5319 ICLR2026 paper notes covering Reinforcement Learning (400), Image Generation (352), Learning Theory (293), LLM Reasoning (241), Model Compression (240), Optimization & Theory (222), Multimodal VLM (211), LLM Agent (209) and other 52 areas. Each note has TL;DR, motivation, method, experiments, highlights, and limitations — 5-minute reads of core ideas. 💡 LLM Reasoning (241) ¶ Vision-R1 is proposed, which constructs 200K high-quality multimodal CoT data via Modality Bridging for cold-start initialization. By combining a Progressive Thought Suppression Training (PTST) strategy with GRPO reinforcement learning, it achieves multimodal mathematical reasoning capabilities close to OpenAI o1 at a 7B parameter scale. DeepMath-103K is a large-scale mathematical reasoning training set specifically designed for Reinforcement Learning from Verifiable Rewards (RLVR). Starting from 2.869 million raw math forum problems, it undergoes rigorous decontamination, difficulty filtering (primarily levels 5–9), and answer verifiability checks. The resulting 103,000 high-difficulty problems have almost no overlap with mainstream evaluation benchmarks, each featuring machine-verifiable answers and three R1-generated solutions. Models trained with RL on this dataset lead in benchmarks like AIME and MATH500, generalizing to non-mathematical reasoning tasks such as biology, physics, and chemistry. DEER enables Large Reasoning Models (LRMs) to trial-answer at "reasoning switch points" within the Chain-of-Thought (CoT). It uses the confidence of these trial answers to judge if the reasoning is sufficient, allowing for training-free dynamic early exit. Across 11 models and 10 benchmarks, it reduces CoT length by an average of 19.1%~80.1% while improving accuracy by 0.3%~5.0%. This paper systematically reveals that the belief "the longer the Chain-of-Thought, the better" is a misconception—task accuracy follows an inverted U-shaped curve relative to CoT length. There exists an optimal length that shortens as model capability increases and task difficulty decreases. The authors explain this phenomenon using a theoretical model of error accumulation, derive a scaling law, and provide two practical recipes: "constructing training data based on optimal length" and "length-filtered voting during inference." The authors decompose the creation of "Reasoning Model SFT Data" into six pipeline stages and conduct over 1,000 controlled ablation experiments. They derive a simple yet counter-intuitive data recipe (high-quality sources + LLM difficulty/length filtering + 16x answer sampling per question + skipping answer verification + using the weaker QwQ-32B as the teacher). Using this, they produced the OpenThoughts3-1.2M dataset and trained OpenThinker3-7B, which outperforms R1-Distill-7B by 15.3/17.2/20.5 percentage points on AIME25, LiveCodeBench, and GPQA respectively, achieving state-of-the-art among open-source models of the same scale. GPG (Group Policy Gradient) returns to the most basic policy gradient, directly optimizing the original RL objective—eliminating the critic, reference model, KL constraints, and surrogate loss, while retaining only group-mean normalization and a gradient debiasing correction, consistently outperforming GRPO in math and multimodal reasoning tasks. DeepConf leverages local confidence signals inherent in LLM generation to dynamically filter low-quality reasoning chains atop parallel thinking (multi-sampling + majority voting). It uses confidence-weighted voting with Top-η% filtering in offline mode and employs the least grouped confidence as a trigger for early stopping and adaptive sampling in online mode. Without training or hyperparameter tuning, it improves GPT-OSS-120B accuracy to 99.9% on AIME 2025 while reducing generation tokens by up to 84.7%. SimpleTIR discovers that the root cause of RL training collapse in multi-turn Tool-Integrated Reasoning (TIR) is the accumulation of low-probability tokens introduced by tool feedback. It proposes a plug-and-play trajectory filtering strategy—discarding entire trajectories containing a "void turn"—to stabilize gradients, raising the AIME24 score of the Qwen2.5-7B base model from a text-only baseline of 22.1 to 50.5. ShinkaEvolve utilizes a "Parent Weighted Sampling + Code Novelty Rejection Sampling + Bandit-style LLM Ensemble Selection" triad to compress LLM-driven program evolution from thousands of evaluations to just 150. It achieves state-of-the-art results across four domains: circle packing, AIME agent scaffolding, ALE-Bench competitive programming, and MoE load-balancing loss. The paper proposes E2H Reasoner , which decomposes training data into four difficulty levels—"trivial, easy, medium, and hard"—and utilizes a probability scheduler (Cosine or Gaussian) to smoothly shift the sampling focus from easy to hard. This approach enables small models to master complex reasoning tasks that are unsolvable zero-shot, while providing theoretical guarantees for CRL convergence and sample complexity. Browse all 241 LLM Reasoning papers → 🦾 LLM Agent (209) ¶ The ACE (Agentic Context Engineering) framework proposes treating context as a continuously evolving "playbook." By utilizing a division of labor among Generator-Reflector-Curator roles and incremental delta updates, it continuously accumulates and refines strategies. This addresses brevity bias and context collapse in existing prompt optimization methods, achieving an average improvement of 10.6% on agent tasks and 8.6% on financial tasks, while reducing adaptation latency by 86.9%. Based on memory and cognitive sciences, the authors decompose the capabilities required by memory agents into four core dimensions: "Accurate Retrieval, Test-Time Learning, Long-Range Understanding, and Selective Forgetting." They construct MemoryAgentBench , the first unified benchmark that simulates multi-turn interaction by chunking long texts and feeding them incrementally to agents. The study finds that no existing long-context models, RAG systems, or commercial memory agents can master all four capabilities simultaneously. MEM1 uses end-to-end reinforcement learning to train LLM agents to embed "memory consolidation" into "reasoning" itself. By maintaining a single, continuously rewritten compact internal state per round and discarding old observations, it maintains near-constant context across arbitrarily long multi-turn tasks, achieving higher performance, lower memory usage, and faster inference. This paper proposes ST-WebAgentBench, the first benchmark specifically designed to evaluate the safety and trustworthiness of Web Agents. Through a hierarchical policy framework and the Completion under Policy (CuP) metric, it reveals that current SOTA agents exhibit severe policy violations in enterprise scenarios. MCP-Bench connects agents to 28 real-world production-grade MCP services (totaling 250 tools across 11 domains such as finance, research, and travel). By utilizing automatically synthesized complex tasks characterized by "fuzzy instructions, multi-objectives, and cross-domain dependencies," combined with a dual-layer evaluation of "rule-based checking + LLM-as-a-Judge," it systematically exposes the genuine deficiencies of 20 mainstream LLMs in long-term planning and dependency reasoning. GTA1 addresses the issue of planning cascade failures through test-time scaling—sampling multiple action proposals per step and selecting the best via a multimodal judge. It further achieves precise localization using a pure RL grounding model (without CoT thinking) that directly predicts coordinates with binary "hit-or-miss" rewards. This two-stage agent reaches SOTA on both grounding and task execution benchmarks. AgenTracer employs "counterfactual replay + programmatic fault injection" to automatically annotate multi-agent failure trajectories, constructing the TracerTraj-2.5K dataset. It then trains a lightweight 8B "failure tracer" using multi-granularity reinforcement learning. On the Who&When benchmark, it localizes decisive errors to specific agents and steps, outperforming giant models like Gemini-1.5-Pro and Claude-3.5-Sonnet by up to 18.18% in agent-level accuracy. Furthermore, providing feedback to off-the-shelf systems like MetaGPT and MaAS leads to performance gains of 4.8~14.2%. MemGen enables LLM agents to determine when to recall in real-time during inference via a "memory trigger," followed by a "memory weaver" that generates machine-native latent token sequences injected into the inference stream. This intertwines memory and reasoning into a dynamic cycle, significantly outperforming parametric and retrieval-based memory while keeping the backbone frozen and without modifying a single parameter. M3-Agent converts real-time visual and audio streams into entity-centric multimodal long-term memory, utilizing a reinforcement learning-trained control model for multi-turn retrieval and reasoning. It outperforms prompt-based closed-source agents and online long video understanding baselines on M3-Bench and VideoMME-long. MedAgent-Pro decomposes the modern clinical "evidence-based diagnosis" process into a two-layer agentic workflow: disease-level standardized planning and patient-level step-by-step evidence reasoning . It utilizes RAG to align with medical guidelines, employs vision/coding tools for quantitative analysis, and utilizes an evidence reflection mechanism to prune unreliable intermediate conclusions. This transforms VLMs from "empirical one-jump responders" into a diagnostic system that is "metric-driven, evidence-based, and traceable." Browse all 209 LLM Agent papers → ⚖️ Alignment & RLHF (102) ¶ This paper extends LLM sycophancy from "agreeing with false facts" to "excessively maintaining user face," proposing a social sycophancy theoretical framework. It constructs the ELEPHANT benchmark to evaluate 11 major LLMs, finding they are on average 47 percentage points more sycophantic than humans in daily advice queries. The study reveals that sycophantic tendencies are rewarded in preference datasets and provides mitigation strategies such as prompt rewriting and DPO. A two-stage Human-AI synergistic preference data curation pipeline is proposed: Phase 1 accumulates approximately 1M preference pairs through 8 iterations of human verification, error-driven adaptive retrieval, and preference-guided LLM annotation; Phase 2 scales the data to 26M pairs using dual-RM consistency filtering. The resulting Skywork-Reward-V2 8B model achieves 97.8% on RewardBench and an average of 88.6% across seven major benchmarks, outperforming all open-source 70B reward models. This paper introduces RewardBench 2, a reward model evaluation benchmark utilizing completely new, unseen human prompts and transitioning from a "1-vs-1" to a "1-vs-3 (1 positive, 3 negatives)" format. Covering six major domains (including new areas like Ties, Precise IF, and Factuality), it is on average 20 points more difficult than the original RewardBench and exhibits significantly stronger correlation with downstream applications such as best-of-N sampling and PPO training. AlphaSteer is proposed to dynamically construct steering vectors by learning a transformation matrix subject to null-space constraints. It generates near-zero vectors for benign inputs (preserving utility) and reconstructs refusal direction vectors for malicious inputs (enhancing safety), providing a theoretical guarantee for the decoupling of safety and utility. The authors theoretically prove that reward over-optimization primarily stems from reward model misspecification in the high-reward tail regions. They propose a rubric-based reward modeling method: utilizing off-policy data (excellent responses generated by strong models) to construct scoring rubrics and refining them through progressive "great vs. greater" differentiation to effectively mitigate over-optimization. It is discovered that many "zero-variance prompts" (where all sampled responses are either all correct or all incorrect) are discarded during GRPO training. The RL-ZVP algorithm is proposed to extract learning signals from these prompts via entropy-guided advantage shaping, achieving improvements of up to 8.61 accuracy points and 7.77 pass rate points across six mathematical reasoning benchmarks compared to GRPO. By revisiting the safety-constrained RLHF objective and proving it possesses a closed-form optimal policy, this work derives an equivalent tractable objective, SafeDPO. It requires only safety-aware data transformation and a safety margin term (one additional hyperparameter) on top of standard DPO. Without needing reward/cost models, it achieves a 96.87% harmless rate on PKU-SafeRLHF-30K while maintaining competitive helpfulness, with training speeds 25x faster than SafeRLHF. Based on representative human surveys of 15,000 individuals across 5 countries, the authors demonstrate that 21 SOTA LLM responses align with only 41% of human preferences ("algorithmic monoculture"). Existing preference datasets fail to learn this diversity because candidate responses are too homogeneous. To address this, "Negative Correlation (NC) sampling" is proposed—using a single prompt to generate four deliberately divergent responses at once. This significantly improves the ability of alignment methods to learn heterogeneous preferences. Consequently, the authors open-source Community Alignment, the largest and most representative multilingual multi-turn preference dataset to date (233,319 comparisons). This paper reveals the "historical context inconsistency" problem in stepwise group-based RL (such as GRPO/GiGPO), where steps within the same group may have different historical contexts, leading to biased advantage estimation. HGPO is proposed to achieve low-bias, balanced-variance advantage estimation through hierarchical grouping and adaptive weighting, achieving significant improvements on ALFWorld and WebShop with minimal extra overhead (<0.001%). WIMHF uses Sparse Autoencoders (SAEs) on the "embedding difference between two candidate responses" to learn a small set of human-readable features. It then quantifies the impact of each feature on preference labels using logistic regression. This process automatically characterizes what a preference dataset "can measure" and what "annotators actually prefer" without pre-defined hypotheses, providing controllable levers for data sanitation and personalization. Browse all 102 Alignment & RLHF papers → 🔒 LLM Safety (185) ¶ This work proposes wd1, a ratio-free weighted log-likelihood policy optimization method for the RL fine-tuning of diffusion language models (dLLMs). By utilizing positive sample weighting and negative sample penalties, it avoids the bias and high variance issues associated with policy ratio estimation in GRPO. It achieves SOTA performance on LLaDA-8B, including +59% on Sudoku and 84.5% on GSM8K. The paper proposes Chain & Hash: an LLM fingerprinting technique that deterministically binds a set of fingerprint questions to their target answers using cryptographic hashing. This allows model owners to provide non-forgeable proof of ownership under pure black-box conditions. Through random padding and meta-prompt diversification during training, the fingerprint survives fine-tuning, quantization, and style-altering prompts. This paper proposes "obfuscation attacks"—adding a loss term to "deceive latent-space monitors" in addition to behavioral goals like jailbreaking or SQL generation. By jointly optimizing adversarial suffixes, the authors reduce the recall of various activation probes (Linear/MLP/OOD) from 100% to 0% while maintaining a 90% jailbreak success rate. This proves that white-box latent-space monitoring is not robust against worst-case attacks, though it identifies an "obfuscation tax": evading probes in complex tasks like SQL generation degrades the model's inherent performance. By rewriting unit tests from LiveCodeBench / SWE-bench into "impossible tasks" that directly conflict with natural language specifications, "passing the test" becomes equivalent to "cheating." This transforms the propensity of LLM coding agents to exploit test loopholes (reward hacking) into a zero-noise, automatically measurable metric. The authors first train a "context-dependent" Llama Nemotron 49B model organism (writing Python type hints during evaluation but omitting them during deployment) and then demonstrate that adding a contrastive activation steering vector to the residual stream can deceive the model into believing it is deployed in an evaluation setting, thereby exposing the "deployment behavior" it intended to hide. This paper advances Constitutional Classifiers from "robust but expensive" safety filters to a production-grade version: by combining an exchange classifier with context awareness, a two-stage cascade, and an activation-based linear probe, it enhances robustness in universal jailbreak red-teaming while reducing computational overhead to approximately \(1/40\) of a single exchange classifier. A safety evaluation benchmark built on real-world MCP servers, utilizing a unified 20-category attack taxonomy (covering server/host/user sides) and multi-turn ReAct tasks across 5 real domains. It systematically reveals that current mainstream LLM agents are generally vulnerable in MCP environments and exhibit a "safety-utility trade-off" where stronger capabilities often correlate with lower safety. Before releasing gpt-oss, OpenAI proactively utilized "Malicious Fine-Tuning (MFT)" to maximize the model's capabilities in the high-risk domains of biology and cybersecurity to estimate the worst-case risks of open-weight release. The conclusion is that even under the strongest adversarial pressure, gpt-oss does not outperform the closed-source o3 and does not substantially advance the existing open-source capability frontier, thereby resulting in limited net additional harm. This paper proposes a comprehensive guardrail solution targeting the "pre-execution" phase, the safest intervention point for LLM agents. It utilizes a controllable synthetic data engine, AuraGen, to generate large-scale annotated risk trajectories to train a lightweight guardian model, Safiron (equipped with cross-planner adapters for unified input formats), to determine, classify, and explain risks. A human-verified Pre-Exec Bench is released, on which Safiron outperforms strong baselines like GPT-5 and Claude-3.7 across four key metrics. This paper proposes Monitor Red Teaming (MRT), a standardized "stress test" workflow utilizing threat models, evasion strategies, and two sabotage benchmarks (including the newly created computer-use benchmark CUA-SHADE-Arena). It systematically tests monitoring systems designed to detect "covert sabotage" by LLM agents and introduces a "hierarchical + sequential" hybrid monitoring scaffold, enabling weaker models to reliably supervise stronger agents. Browse all 185 LLM Safety papers → 👻 Hallucination Detection (40) ¶ By injecting uniform noise into the MLP activations of intermediate LLM layers to approximate the Bayesian posterior, this method captures epistemic uncertainty. It complements aleatoric uncertainty captured via sampling temperature, improving the hallucination detection AUROC on GSM8K from 71.56 to 76.14. This paper identifies "hallucination snowballing" in VLM Multi-Agent Systems (MAS)—where a visual misjudgment by one agent is progressively amplified by subsequent agents via pure text streams. Through turn-wise, layer-wise, and token-level attention analysis, the authors locate "middle-layer unimodal visual tokens" as the critical carriers of visual evidence. They propose ViF: establishing an additional "visual flow" between agents using these visual relay tokens combined with attention reallocation. This model-agnostic approach mitigates snowballing and achieves consistent 2.4–3.8% improvements across 8 benchmarks, 4 MAS structures, and 10 backbones. This paper proposes a unified theoretical framework called the "Hallucination Risk Bound," which decomposes the hallucination risk of LLMs into a data-driven term (representational bias during training) and a reasoning-driven term (instability during decoding) using the triangle inequality. Based on this, the authors design HalluGuard, an NTK-based spectral proxy score that requires no external references or hallucination annotations, achieving consistent SOTA performance across 10 benchmarks, 11 baselines, and 9 backbones. Ours proposes the LVLMs-Saliency gradient-aware diagnostic framework to quantify the visual anchoring strength of each output token. It uncovers a critical pattern: "hallucination occurs when the saliency of previous output tokens regarding the next token prediction decreases." Based on this, a dual-mechanism inference-time framework, SGRS (Saliency-Guided Rejection Sampling) + LocoRE (Local Consistency Restoration), is designed to significantly reduce hallucination rates across multiple LVLMs. Ours proposes the Lumina framework, which detects hallucinations in RAG systems via "context-knowledge signals": MMD is used to measure the extent of external context utilization , while cross-layer token prediction evolution measures internal knowledge utilization . The method generalizes without hyperparameter tuning. Addressing the tendency of Large Reasoning Models (LRMs) to "hallucinate rather than admit ignorance" in factual QA, this paper identifies two pathological reasoning patterns triggered by "factual overthinking." It proposes BARREL, a three-stage training framework (Knowledge Boundary Labeling \(\rightarrow\) Boundary-Aware SFT \(\rightarrow\) GRPO with Reliability Rewards). BARREL improves the reliability of DeepSeek-R1-Distill-Llama-8B from 39.33% to 61.48% while simultaneously increasing accuracy. The AIR (Adaptive vIsual Reinforcement) framework is proposed to reduce MLLM hallucinations during inference without training by utilizing prototype distance-based token pruning and Optimal Transport-guided selective patch reinforcement (LLaVA-1.5-7B CHAIR_S: 22→18.4, POPE Accuracy +5.3%), while maintaining general multimodal performance. This paper proposes RL4HS: using reinforcement learning (GRPO based on span-F1 rewards) to train 7B/14B models to perform "reason-then-locate" for precise hallucination span detection. It introduces Class-Aware Policy Optimization (CAPO) to correct systematic reward biases towards the "no-hallucination" class, outperforming SFT and proprietary large reasoning models (GPT-5, o3) on RAGTruth. CASAL "amortizes" inference-time activation steering into model weights—by training only a sub-module of a single layer using only representation loss (without cross-entropy), the LLM learns to "answer what it knows and abstain from what it doesn't." This reduces hallucination rates by 30%–40% on multiple short-form QA benchmarks while requiring ~30× less compute and ~20× less data than LoRA-style baselines. This work systematically traces LVLM object hallucinations back to the visual encoder for the first time, identifying three major issues: statistical bias (over-emphasis on high-frequency pattern tokens), inherent bias (residual representations of dominant pre-training objects), and vulnerability (feature distortion caused by minor perturbations). It proposes SHIELD, a training-free framework that synergistically defends against these via token re-weighting, token subtraction, and contrastive decoding, outperforming methods like VCD and OPERA on LLaVA-1.5, InstructBLIP, and Qwen-VL. Browse all 40 Hallucination Detection papers → 📊 LLM Evaluation (131) ¶ Through large-scale experiments involving "instruction sharding + simulated dialogue" (200k+ dialogues, 15 LLMs), this paper demonstrates that all top-tier LLMs suffer an average performance drop of 39% in multi-turn underspecified conversations compared to single-turn full instructions. This degradation is primarily caused not by a decline in aptitude, but by a reliability collapse —once a model takes a wrong turn, it becomes "lost" and cannot recover. Proposes DeepResearch Bench , the first systematic benchmark for "Deep Research Agents" (DRA)—comprising 100 PhD-level research tasks across 22 disciplines crafted by experts, supported by two automated and highly human-aligned evaluation frameworks: RACE for report quality and FACT for information retrieval and citation reliability. This paper defines and systematically investigates Preference Leakage (PL) in LLM-as-a-Judge—a phenomenon where judge \(M_J\) systematically favors "related student models" when the synthetic data generator \(M_G\) is associated with \(M_J\) (same model, inheritance, or same family). In same-model scenarios, the PLS reaches 28.7% (Arena-Hard), and this bias is more insidious and harder to detect than egocentric bias. GDPval is a benchmark proposed by OpenAI for "real economic value digital knowledge work": covering 9 industries with the largest contributions to US GDP, 44 occupations, and 1,320 real tasks constructed by professionals with 14 years of experience. Using "model vs. human expert blind win rate" as the core metric, the study finds that the delivery quality of frontier models is linearly approaching industry experts year by year. ARPO is a reinforcement learning algorithm tailored for multi-turn tool-calling agents. It identifies that the token entropy of LLMs spikes after each tool return. Consequently, it adaptively "forks" sampling at these high-entropy steps and employs advantage attribution to propagate the performance differences of branched paths back for learning. This achieves superior performance across 13 reasoning/deep-search benchmarks compared to trajectory-level RL, while using only half the tool-calling budget. ResearchRubrics utilizes 2800+ hours of human effort to pair 101 real-world open-ended research prompts with 2593 expert-written, weighted, fine-grained rubrics. Using LLM-as-Judge to score agents based on these rubrics, the study evaluates mainstream Deep Research (DR) systems and finds that even the strongest agents, such as Gemini DR and OpenAI DR, fail to reach an average rubric adherence rate of 68%. Theoretical bottlenecks are concentrated in implicit requirement inference and multi-source information synthesis. HAL provides a standardized, distributed, and automated infrastructure for AI agent evaluation. Through 21,730 rollouts across the dimensions of "model × scaffold × benchmark," it maps the accuracy-cost Pareto frontier. By using LLMs to analyze 25 billion log tokens, it reveals behaviors hidden by traditional metrics, such as performance degradation with increased reasoning effort, agents searching for answers on HuggingFace, and using incorrect credit cards for bookings. CyberGym constructs a cybersecurity evaluation benchmark that is more than 7 times larger than existing similar benchmarks by using 1,507 historical vulnerabilities from 188 real-world open-source projects on OSS-Fuzz. The core task requires AI agents to generate Proof-of-Concept (PoC) exploits given only a textual vulnerability description and the pre-patch codebase. Results show that even the strongest agent+model combinations achieve only about a 20% success rate. Furthermore, the evaluation process inadvertently discovered 34 0-day vulnerabilities and 18 incomplete patches, proving it to be both a rigorous benchmark for measuring AI progress and a platform capable of generating real-world security impact. SimBench unifies 20 cross-disciplinary datasets from ethics, economics, psychology, and politics into a "population response distribution prediction" task, constructing the first large-scale standardized LLM human behavior simulation benchmark. Systematic evaluation across 45 models reveals that even the strongest current models achieve only a moderate fidelity of 40.80/100. Simulation capability grows log-linearly with scale but does not improve with increased inference compute, and instruction tuning exhibits a distinct "alignment-simulation tradeoff." This paper employs a randomized controlled experiment—involving expert execution and blind review—to verify whether research ideas generated by LLMs truly translate into superior research outcomes. It finds that while LLM ideas receive higher scores when evaluated as standalone "proposals," they suffer significantly larger drops in novelty, excitement, effectiveness, and overall quality after execution. Browse all 131 LLM Evaluation papers → ⚡ LLM Efficiency (171) ¶ Fast-dLLM accelerates bidirectional Diffusion LLMs without retraining by introducing a block-wise approximate KV Cache and replacing fixed top-K parallel decoding with a "confidence threshold" strategy. It achieves up to a 27.6× end-to-end throughput gain on LLaDA and Dream with almost no loss in accuracy. MemAgent segments "unbounded long documents" into fixed-length chunks for streaming processing, replacing the ever-expanding context with a fixed-length, overwritable token memory. By utilizing an extended version of Multi-Conv DAPO for end-to-end training of memory read/write strategies, 8K window models can extrapolate to 3.5M token QA tasks with nearly no performance loss and strictly linear inference complexity. Fast-dLLM v2 transforms pre-trained autoregressive Qwen2.5 models into block-diffusion language models via lightweight fine-tuning with approximately 1B tokens. Combined with hierarchical caching and confidence-based parallel decoding, it achieves up to 2.5× speedup over AR decoding without performance degradation. This paper points out that existing Test-Time Training (TTT) approaches fail on long sequences because they adhere to tiny online mini-batches (updating fast weights every 16~64 tokens), causing modern GPU utilization to remain below 5%. The authors take the opposite approach and propose LaCT (Large-Chunk Test-Time Training) , which expands the update granularity to massive chunks of 2K~1M tokens. Combined with window attention to compensate for intra-chunk locality, LaCT achieves 70% GPU utilization with just a few dozen lines of pure PyTorch code. It demonstrates scalability up to 14B parameters and 56K~1M token contexts across three modalities: new view synthesis, language modeling, and autoregressive video diffusion. This paper proposes Discrete Diffusion Forcing (D2F), which transforms pre-trained diffusion language models (dLLMs) into a hybrid AR-diffusion paradigm featuring "block-level autoregression + inter-block parallel decoding." By acquiring this capability via low-cost asymmetric distillation and pairing it with pipeline parallel decoding, the authors demonstrate for the first time that open-source dLLM inference throughput can surpass that of autoregressive LLMs of the same scale (achieving 2.5× speedup over LLaMA3 on GSM8K and 50× over the original dLLM). This paper proposes UniRoute, which encodes each LLM into a "prediction error vector on a small batch of representative prompts." Paired with a bilinear scorer, it allows the trained router to route to new LLMs appearing only at test time without retraining , achieving a better cost-quality trade-off across 30+ unseen models. Two training-free techniques—FreeCache, which reuses stable KV projections, and Guided Diffusion, which uses consistency signals from small AR models to guide parallel demasking—enable a 7B/8B diffusion language model to achieve an average end-to-end speedup of 12x, bringing diffusion LLM latency comparable to or even faster than same-sized autoregressive models for the first time. Diffusion Language Models (DLMs) often internally determine the correct answer mid-way through decoding. Based on this, this paper proposes Prophet, a training-free decoding paradigm that uses the "logit gap between the top-2 candidate tokens" to judge answer convergence. Once converged, it fills all remaining positions in a single step (Early-Submission Decoding), reducing decoding steps by up to 3.4 \(\times\) on LLaDA-8B / Dream-7B with almost no loss in accuracy. Replace "online prefilling of long documents into KV cache" with "offline training of a small learnable KV cache (Cartridge) for each corpus." Use Self-Study (self-generated synthetic dialogues + context distillation) to replicate the general-purpose In-Context Learning (ICL) capabilities in the small cache, achieving a 38.6× reduction in memory and a 26.4× increase in throughput on average. By reshaping the "serial certainty convergence" (predicting token-by-token) of diffusion language models (dLLMs) into "parallel simultaneous convergence" through "Certainty-Forcing Distillation," LLaDA-8B achieves an 8.5× acceleration on GSM8K, reducing decoding steps from 256 to 30 without accuracy loss. Browse all 171 LLM Efficiency papers → 📚 Pretraining (79) ¶ This paper systematically compares 22 data selection strategies for LLM pre-training. It proposes Ask-LLM , which uses instruction-tuned LLMs to directly provide quality scores, and Density , which performs coverage sampling based on kernel density estimation. The study finds that quality filtering (Ask-LLM) can outperform full-scale training while converging 70% faster even when keeping only 10% of the data, whereas coverage sampling typically only "matches" the full-scale performance. This paper proposes H-Net, a hierarchical sequence model that replaces BPE tokenization with a learnable "Dynamic Chunking (DC)" mechanism. The network automatically learns where to split chunks and the granularity of compression on byte-level inputs in an end-to-end differentiable manner. Under compute and data alignment, a single-stage H-Net outperforms BPE-based Transformers, while a two-stage H-Net matches the performance of token-level models twice its size. The authors develop a suite of paired encoder-only and decoder-only models (the ETTIN suite) ranging from 17M to 1B parameters. Using identical data, architectures, and training recipes—varying only in the objective function and attention direction—they achieve SOTA performance on open-data benchmarks for both types. They demonstrate that encoders significantly outperform decoders in classification and retrieval tasks, while the reverse is true for generation, and converting one model type to another via continued training (cross-objective) cannot bridge this performance gap . This paper reformulates "prediction" as "gradient descent optimization on a learned verifier (energy function)" and proposes Energy-Based Transformers (EBTs). This architecture enables cross-modal and cross-task System 2 thinking capabilities (dynamic compute allocation + self-verification) to emerge purely through unsupervised pre-training, outperforming Transformer++ and DiT in both language and vision domains. The successful Joint Embedding Predictive Architecture (JEPA) from computer vision is adapted for LLMs for the first time. By adding a latent space objective—"predicting Code embeddings from Text embeddings"—alongside the standard next-token reconstruction loss, this method significantly outperforms standard fine-tuning and pre-training across four model families and four datasets without sacrificing generative capabilities or suffering from overfitting. A domain-agnostic pipeline utilizing "lynx layout rendering + lightweight LLM cleaning" is proposed to reliably extract and standardize math/code content from Common Crawl. This constructs Nemotron-CC-Math (133B tokens), the highest-quality open-source math pre-training corpus to date, which consistently outperforms FineMath, MegaMath, and OpenWebMath across math, code, and general knowledge tasks. This paper systematically investigates the scaling laws of discrete diffusion language models (DLMs) under various noise types. By employing a unified diffusion framework parameterized by signal-to-noise ratio (SNR) that smoothly interpolates between masked and uniform diffusion, and carefully calibrating batch size and learning rate, the authors find that DLM scaling behavior heavily depends on the noise type. Uniform diffusion is more "data-efficient but parameter-hungry" in data-constrained scenarios. The study scales uniform diffusion models up to 10B parameters / \(10^{22}\) FLOPs, verifying that their scaling laws can compete with autoregressive models (ALMs). Authors find that Diffusion Language Models (dLLMs) often "get it right in the middle but change to wrong at the end" (temporal oscillation) during denoising. They exploit discarded intermediate step predictions as signals using two methods: a training-free Temporal Self-Consistency voting that selects the most stable answer across steps within a single sampling trajectory, and a Temporal Consistency Reinforcement post-training via GRPO using "Negative Temporal Semantic Entropy" as an unlabeled reward. These yield improvements of ~1.5% and up to 25.3% across four math reasoning benchmarks, respectively. This paper proposes CoCoMix, which goes beyond standard next-token prediction by having the model predict high-level concepts extracted via SAE and filtered by attribution. These concepts are compressed into continuous vectors and interleaved into the Transformer hidden state sequence, achieving higher efficiency in language modeling, downstream reasoning, and controllable generation compared to vanilla NTP and knowledge distillation. When compute far exceeds available web data, the authors use "heavy regularization + model ensemble + joint parameter/ensemble scaling + distillation" to compress the pre-training loss of a fixed 200M token budget to an asymptotic value of 3.17. This achieves a 5.17× data efficiency gain over standard recipes, with 83% of the ensemble benefits retained even when distilled into an 8× smaller student model. Browse all 79 Pretraining papers → ✏️ Knowledge Editing (15) ¶ This paper reveals that the root cause of the underestimated performance of fine-tuning in model editing is an incorrect training pipeline (Depth-First sample-by-sample optimization). By correcting this to a standard Breadth-First mini-batch training and combining it with localized parameter tuning to form LocFT-BF, the authors achieve the first support for 100,000 sequential edits and a 72B model scale. Addressing the neglected problem where knowledge editing collateralizes "same-subject but different-relation/object" fine-grained irrelevant knowledge, this paper proposes DiKE: it first uses a reusable disentanglement module to split subject representations into "target-related" and "irrelevant" parts, then performs editing only on the related part while explicitly constraining the irrelevant part to remain unchanged, deriving a closed-form rank-one parameter update formula that maintains mainstream editing performance while preserving fine-grained irrelevant knowledge. This paper proposes the EVOKE benchmark to systematically evaluate the knowledge injection capabilities of Large Multimodal Models (LMMs) for evolving knowledge. It reveals two primary challenges: the poor performance of existing methods and catastrophic forgetting caused by fine-tuning. Furthermore, it explores two response paths: knowledge augmentation and continual learning. ACE identifies a neglected mechanism via neuron-level attribution where "implicit subjects act as query neurons in multi-hop reasoning, activating value neurons layer-by-layer." Accordingly, it refines editing from "layer-level heuristics" to "query-value pathways," outperforming the SOTA PMET by 9.44% on GPT-J and 37.46% on Qwen3-8B in multi-hop factual recall. MoEEdit is the first "routing-stable" parameter-modifying knowledge editing framework for MoE LLMs. It employs "per-expert null space projection" to ensure that edits do not perturb input manifolds for downstream routers, combined with a stochastic Block Coordinate Descent (BCD) solver to decouple computational costs from the total number of experts to the expert hidden dimension. This achieves high editing success rates, strong generalization, and routing stability on sparse architectures simultaneously. This paper proposes KnowledgeSmith, which unifies "knowledge editing" and "machine unlearning" into a single constrained optimization problem. By using knowledge graphs (KG) to automatically generate large-scale evaluation benchmarks across different hierarchies (root/intermediate/leaf) and data scales, the study systematically reveals counter-intuitive phenomena in LLM knowledge updates, such as propagation asymmetry, consistency-capacity trade-offs, and subject dependency. SUIT constrains the "manually calculated" key vector \(k\) and residual vector \(\delta\) in locate-then-edit knowledge editing into low-dimensional subspaces truly relevant to the target edit. This significantly reduces damage to irrelevant knowledge without sacrificing editing success—doubling Specificity on LLaMA3 / GPT-J / Qwen2.5 compared to the strong baseline AlphaEdit. By training Transformers from scratch on synthetic relational knowledge graphs, it is discovered that appropriate regularization leads to the emergence of a bilinear relational structure in hidden layers. This structure not only overcomes the reversal curse but also enables logically consistent propagation of updates to related facts after editing a single fact. Performance degradation in sequential model editing is understood from the perspective of hyperspherical uniformity (Hyperspherical Energy, HE). The SPHERE method is proposed: by projecting editing perturbations onto the orthogonal complement of the primary hypersphere directions of pre-trained weights, stable large-scale sequential editing is achieved, outperforming the strongest baseline by an average of 16.41% on LLaMA3-8B. By using null-space constrained online model editing, this work integrates 3D geometric information provided by VGGT into a 2D generic object tracker. This enhances geometric awareness while maintaining semantic discriminative power, significantly improving tracking performance in scenarios with occlusions and background clutter. Browse all 15 Knowledge Editing papers → 💬 LLM (Other) (55) ¶ This paper proposes CK-PLUG, a plug-and-play, training-free decoding-stage method. It uses "Confidence Gain (CG)" to detect conflicts between parametric knowledge and retrieved context at the token level, then blends "parameter-side" and "context-side" probability distributions using a single hyperparameter \(\alpha\) . This enables continuous, bidirectional, and controllable adjustment between "the model's own memory" and "retrieved context"—allowing the Memory Recall (MR) on LLaMA3-8B to be tuned anywhere between 9.9% and 71.9% while maintaining generation fluency. SPRIG utilizes an "edit-based genetic algorithm + proxy reward model" to automatically assemble a task-agnostic system prompt . A single system prompt's average improvement across 47 task categories matches "task-specific prompts optimized individually for each task," while the combination of both further achieves new SOTA results and transfers across model families and languages. Addressing the inherent flaw in masked diffusion language models where "tokens are fixed once unmasked and cannot be corrected," this paper proposes RemeDi. The model simultaneously predicts token distributions and per-token confidence during each generation step. Based on confidence levels, it decides which positions to unmask and which previously generated tokens to revert to masks for resampling. Coupled with a two-stage "Remask SFT + Remask RL" training protocol, it achieves SOTA results among open-source diffusion language models (GSM8K 89.1%, HumanEval 73.2%). DreamOn resolves the fixed-length generation constraint of Diffusion Language Models (DLMs) by introducing two special states, [expand] and [delete] . This enables variable-length code infilling without architectural modifications, achieving a 26.4% average improvement on HumanEval-Infilling compared to diffusion baselines and reaching performance parity with state-of-the-art (SOTA) autoregressive models. Proposed d²Cache, a training-free approximate KV caching framework for Diffusion-based LLMs (dLLMs). It achieves a \(4.1\times\) inference speedup while improving generation quality through a two-stage strategy: deterministic-prior-guided masked token selection combined with attention-aware non-masked token selection. The Optimas framework is proposed to maintain a Local Reward Function (LRF) aligned with the global reward for each component within a compound AI system. This enables the independent optimization of heterogeneous components (prompts, model parameters, hyperparameters, model selection), achieving an average performance gain of 11.92% across five real-world systems. TableMaster systematically decomposes the "structural features" of tables into four categories of challenges. It then provides a "four-flavor recipe"—constructing a Table-of-Focus, verbalizing for semantic enrichment, adaptive switching between textual/symbolic reasoning, and text-guided symbolic reasoning—integrated into a training-free three-stage framework. This approach improves accuracy from 64.73% to 78.13% on WikiTQ using GPT-4o-mini. This paper proposes the first constrained decoding method to enforce Context-Free Grammar (CFG) constraints on Diffusion Language Models (DLMs) . It abstracts the question of "whether a partial text with holes in any order can be completed into a legal string" as an additive infilling decision problem , reduces it to checking whether the intersection of the target CFG and the regular language of all possible completions is empty, and utilizes highly optimized emptiness checking algorithms to bring the theoretical cubic complexity into a practical range. It increases syntactic correctness to nearly 100% on C++, JSON, and SMILES while slightly improving functional correctness. Proposes BOTS, a unified framework for online task selection in LLM reinforcement finetuning based on Bayesian inference. By fusing explicit evidence (historical pass rates from direct evaluation) and implicit evidence (inferred difficulty of unevaluated tasks via reference model interpolation) with Thompson sampling for exploration-exploitation balance, BOTS achieves up to 50% training acceleration on math, code, and logic tasks with only 0.2% additional overhead. This paper proposes that the "Reversal Curse" is a manifestation of the "binding problem" from cognitive science in Transformers—stemming from inconsistency and entanglement of concept representations. It designs an architecture based on JEPA and memory layers to truly break the reversal curse for the first time without relying on data augmentation. Browse all 55 LLM (Other) papers → 📖 NLP Understanding (2) ¶ This paper proposes the mean activation difference steering method and accompanying quantitative metrics. Through two case studies—rhyme generation and question answering—it systematically proves across 23 open models (1B-32B) that representations of target tokens (rhymes/answers) are formed at early sequence positions (forward planning) and causally influence the generation of intermediate tokens (backward planning). Implicit planning emerges in models as small as 1B, suggesting it is a universal mechanism rather than exclusive to large-scale models. LANE upgrades the classic margin indicator for identifying mislabeled samples to a Label-aware Margin . For negative margins, the penalty is reduced if the mislabeled category is semantically similar to the model's prediction (e.g., "Anger" labeled as "Fear") and increased if they are semantically distant (e.g., "Trust" labeled as "Fear"). Based on this, it applies dynamic weighting to each sample rather than hard deletion, consistently outperforming strong baselines like AUM and HMW across 10 text classification datasets. ✍️ Text Generation (12) ¶ This paper proposes Self-adaptive Schema Scaffolding (S3) —a training-free method that injects a structure template (schema) as a "semi-denoised initial state" directly into the output context of a Diffusion Large Language Model (dLLM). Augmented with null placeholders for adaptive length, S3 allows dLLMs to stably generate valid structured outputs like JSON with fewer denoising steps. It improves structure compliance from the 30%–80% baseline range to over 99%, while maintaining a lower hallucination rate. Starting from the framework of proper scoring rules, this work proves that the negative log-likelihood of the maximum probability output sequence (MSP) is a theoretically sound uncertainty measure. It proposes G-NLL—which approximates this measure using a single greedy decoding pass—matching or exceeding state-of-the-art (SOTA) sampling-based methods across multiple scenarios. Proposes FS-DFM (Few-Step Discrete Flow-Matching), which reduces the sampling steps of discrete flow-matching language models from 1024 to 8 via step-aware training and a cumulative scalar update rule, achieving 128x acceleration while maintaining comparable perplexity and generation quality. This paper points out the mismatch between the default "random unmasking paths" used in training and the actual "planner-guided paths" used during inference in masked diffusion language models. It proposes Planner-Aware Path Learning (PAPL), which reweights the masked diffusion loss using planner confidence to align training more closely with the inference path. This leads to steady quality improvements in protein sequence, text, and code generation. GloSA-sum introduces Topological Data Analysis (TDA) to text summarization for the first time. It utilizes persistent homology to identify the document's semantic skeleton and logical loops, storing them in a "protection pool." A lightweight proxy metric is then used to iteratively delete sentences, achieving fast and accurate compression without losing core logic chains, while effectively shortening contexts for downstream LLM tasks. This paper identifies a persistent " <eos> overflow" early termination issue in instruction-tuned diffusion language models—where longer allocated generation lengths lead to shorter or even collapsed answers (sequences of <eos> ). The root cause is that <eos> serves as both a terminator and a padding token. The authors propose Rainbow Padding: retaining a single <eos> for actual termination while filling remaining positions with a deterministic cycle of \(K\) distinct padding tokens. Using only 7 tokens and single-epoch LoRA, it restores length robustness, improving LLaDA's accuracy on MATH from 0.6% to 32.6%. This paper proposes p-less sampling: a completely hyperparameter-free truncation decoding method. At each step, it uses the "collision probability" \(\sum_v P_\theta(v)^2\) of the entire token distribution as a dynamic truncation threshold. It outperforms methods like top-p and min-p in mathematics, logical reasoning, and creative writing, showing minimal degradation at high temperatures while offering faster inference. Addressing the issues of "poor citation quality and low readability" in long-form reports generated by deep research systems, this paper proposes a tag-based dual-layer intent (paragraph intent + citation intent) writing framework. This framework enhances Large Language Models (LLMs) via direct prompting during inference and distills Small Language Models (SLMs) using intent-aware synthetic data. Across three scientific report generation benchmarks, LLMs improved by an average of +2.9 points and SLMs by +12.3 points, with citation metrics showing particularly significant gains. Antislop treats "AI-typical repetitive phrases (slop) in LLM generation" as quantifiable, locatable, and erasable objects. It first maps model-specific "slop fingerprints" using frequency ratio statistics, then utilizes an inference-time backtracking sampler to precisely suppress these patterns. Finally, it automatically converts the sampler's interception records into preference data for the newly proposed FTPO fine-tuning, permanently welding the suppression capability into the weights—achieving a 90% reduction in slop with almost no performance degradation on GSM8K/MMLU/Creative Writing. This paper proposes Causal-Steer: by treating LoRA as a "causal intervention," it computes the difference in activations with and without LoRA perturbations on the same input . This approach eliminates the need for parallel corpora and extracts a clean style vector. After PCA denoising and robust aggregation via geometric median, it achieves continuous, bidirectional, and linearly interpolatable style control for LLMs using a single scalar \(\alpha\) during inference. Browse all 12 Text Generation papers → 🗣️ Dialogue Systems (10) ¶ "Flip" the dialogue—instead of training LLMs to be better assistants, specifically post-train a User Language Model (User LM) to simulate real human users. This model is used to expose the weaknesses of assistant LMs in realistic multi-turn scenarios (dropping GPT-4o's task success rate from 74.6% to 57.4%). FlyThinker proposes an efficient "think-while-generating" framework that utilizes an independent Reasoner to generate latent reasoning signals at the token level in parallel. These signals are dynamically integrated into the Generator to guide personalized long-form generation while maintaining training and inference efficiency. Based on marketing research, this paper defines four types of non-collaborative user behaviors (unavailable service, tangential chat, impatience, and incomplete utterances) and constructs a simulation framework that maintains goal-alignment. Evaluations on MultiWOZ and \(\tau\) -bench systematically expose behavior-specific failure mechanisms in SOTA tool agents—tangential chat leads to an average SR drop of 29.1%, with different models exhibiting distinct collapse paths (the GPT series falls into repetitive helper API calls, while the Qwen series tends to hallucinate API results). By comparing layer-wise hidden representations (chain-of-embedding) with and without visual input, this study identifies a "Visual Integration Point" (VIP) layer in LVLMs and proposes the Total Visual Integration (TVI) metric to quantify the strength of the language prior. This paper proposes an extremely fast robustness test: on LLM leaderboards based on the Bradley–Terry model (such as Chatbot Arena), removing a tiny worst-case subset (as few as 2 preferences or 0.003%) of human evaluations can change the top-ranked model. The method precisely identifies which specific preferences cause the flip. DRIFT treats the abundant but implicit "user dissatisfaction" (DSAT) from real-world deployments as high-quality negative anchors. Positive samples are dynamically sampled from the current policy, and iterative training is performed using standard DPO. Without requiring human annotations, reward models, or positives generated by stronger models, it enables a 14B model to outperform GPT-4o-mini on WildBench. The authors propose AQuA, the first VQA dataset (7.2K samples) with fine-grained ambiguity levels (4 levels), defining optimal response strategies for each level (Direct Answer/Inference/Enumeration/Request Clarification). The study finds that GPT-5 and Gemini are overconfident, consistently providing direct answers to ambiguous questions. Conversely, a 3B model trained via SFT+GRPO can surpass the strategy adaptation capabilities of closed-source large models. Addressing the issue where LLMs in role-playing only mimic surface-level actions but fail to remember a character's "internal state," this paper proposes automatically compiling character profiles into executable Finite State Machines (CFSM) . It uses code to explicitly record character states and transition rules, further extending this to CPFSM for modeling states via probability distributions. On both synthetic validation and Fandom real-plot benchmarks, it demonstrates superior coherence and interpretability compared to prompt-only state modeling baselines. Proposes Reasoning Inception (ReIn), a test-time intervention method requiring no modification to model parameters or system prompts. By employing an external inception module to detect conversational errors and inject recovery plans into the task agent's reasoning chain, it significantly improves task completion rates across various error scenarios and generalizes to unseen error types. ClarifyVC employs an agent-orchestrated four-stage data augmentation pipeline to "grow" a large volume of ambiguity-rich and protocol-compliant single/multi-turn dialogues from 20,000 real in-vehicle commands. Accompanied by a three-tier evaluation protocol and a Data Quality Score (DQS), fine-tuning on this data improves parsing accuracy by ~15%, ambiguity resolution by ~20%, and achieves 98% protocol compliance for in-vehicle voice commands. 🌐 Multilingual & Translation (8) ¶ This paper systematically analyzes multilingual routing patterns in MoE large language models, discovering that middle layers contain cross-lingually shared experts and that linguistic performance is strongly correlated with alignment to English routing. Based on this, an inference-time routing intervention method is proposed to activate English task-specific experts in middle layers, consistently improving multilingual performance by 1-2% across 3 models, 2 tasks, and 15+ languages. This paper proposes the Adaptive Transfer Scaling Law (ATLAS), which decomposes the effective data volume into three terms: target language, transfer languages, and other languages, while introducing a data repetition saturation function. Validated across 774 multilingual training experiments (10M–8B parameters, 400+ languages), ATLAS significantly outperforms existing scaling laws (multilingual \(R^2\) improved from 0.67 to 0.98). It systematically quantifies the cross-lingual transfer matrix, the capacity constraints of the "curse of multilinguality," and the compute crossover point between pretraining and finetuning. Ours proposes the ASSESS framework, centered on the TransTED Similarity metric. By parsing formal mathematical statements into Operator Trees (OPT) and integrating Lean proof tactic-driven semantic transformations into the standard Tree Edit Distance (TED), the method achieves SOTA performance with 70.16% accuracy and a 0.35 Kappa score on the EPLA benchmark, while remaining reproducible using only CPU resources. By utilizing logit lens and hidden state similarity analysis, this work localizes the final few layers responsible for "language control" in mLLMs. Fine-tuning only these 3-5% of parameters increases language consistency across six languages from <20% to over 98%, achieving performance nearly equivalent to full fine-tuning. Utilizing Sparse Autoencoders (SAEs), it is discovered that unexpected code-switching in LLMs is correlated with abnormally high pre-activation values of target language features. This paper proposes SASFT, a method that constrains target language feature pre-activations during SFT training, reducing unexpected code-switching by more than 50%. This paper proposes the Language Confusion Gate (LCG) : a lightweight two-layer MLP that masks tokens from incorrect language families on-demand during decoding without modifying the base LLM. Trained via "norm-calibrated self-distillation," it reduces language confusion rates by approximately an order of magnitude across multiple models without sacrificing task performance. This paper proposes ALPO (Adaptive Local Preference Optimization) for training expressive subtitle translation LLMs. Empirical findings show that subtitle translation favors free translation and that reasoning-based LLMs outperform chat-based LLMs in paraphrasing capability. After verifying that LLMs as translation evaluators are highly consistent with humans, the authors propose a fine-grained process-supervised preference alignment method (adaptive weighting + dynamic beta + prefix mixing). The 14B model exceeds SOTA models like GPT-4o and DeepSeek-R1 in translation vividness across multiple language directions. DiscoX constructs the first benchmark for discourse-level + expert-level ZH-EN translation (200 articles, average 1712 tokens, 7 domains, 1330 person-hours of manual refinement) and proposes a multi-agent reference-free evaluation system Metric-S, revealing a significant gap where even the strongest LLM (GPT-5-high: 76.66) still lags behind human experts (80.16). 🔍 Information Retrieval & RAG (81) ¶ This paper derives a lower bound theorem for the "dimension required for single-vector embeddings to represent all top-k document combinations" using sphere packing arguments from high-dimensional geometry. It empirically demonstrates through free embedding optimization and a minimalist real-world dataset, LIMIT, that as long as the number of relevant combinations to be represented is sufficient—even for queries as simple as "who likes apples"—dense retrieval models with fixed dimensions are destined to fail. This is a fundamental bottleneck of the single-vector paradigm rather than an issue of data or scale. Addressing the contradiction that GraphRAG often underperforms relative to basic RAG on real-world tasks, this paper proposes GraphRAG-Bench. This benchmark covers the entire pipeline from graph construction to retrieval and generation, featuring tasks across four difficulty levels. It systematically answers "when and why to use graphs": basic RAG suffices for simple fact retrieval, while graph structures provide substantial gains in complex multi-hop reasoning and context summarization tasks requiring the integration of scattered concepts, though at the cost of significantly higher token consumption. LinearRAG identifies that performance bottlenecks in existing GraphRAG methods stem from unstable and expensive relationship extraction. It proposes a "Tri-Graph" that extracts only entities without relations, coupled with a two-stage retrieval process (semantic bridging for entity activation + global importance aggregation for passage retrieval). This approach reduces indexing time by 77% with zero LLM token consumption while outperforming all SOTA models across four benchmarks. This paper transforms a 0.5B Qwen2 decoder into a fully bidirectional encoder, coupled with a "Pre-training → Fine-tuning → Contrastive Distillation" three-stage pipeline, Focal-style reweighting, online hard negative mixing, and high-quality data engineering covering 100+ categories. This allows KaLM-Embedding-V2.5 to achieve SOTA results in the <1B parameter segment on MTEB Chinese and English benchmarks, even competing with models 3–26 times larger. To address the failure of using Multimodal Large Language Models (MLLMs) directly as encoders under complex instructions, this work proposes the Think-Then-Embed (TTE) framework. It first employs a reasoner to generate an "Embedding-Centric Reasoning" (ECR) trajectory, and then utilizes an embedder to produce vectors conditioned on both the original input and this reasoning trajectory. TTE achieves SOTA on MMEB-V2 (TTE \(_t\) -7B 71.5%), leading other open-source models by an absolute margin of approximately 7%. By appending a small set of learnable Meta Tokens to VLM input sequences and organizing information by granularity into these tokens using Matryoshka-nested multi-vector contrastive training, MetaEmbed allows users to freely choose between 1 to 64 vectors at test-time to trade off retrieval accuracy against indexing and latency overhead. This achieves SOTA performance on MMEB and ViDoRe with compact multi-vector representations and scales stably to 32B parameters. Youtu-GraphRAG utilizes a "graph schema" to vertically integrate traditionally isolated graph construction and retrieval. The construction end uses the schema to constrain extraction and perform automatic expansion; the indexing end builds a four-layer knowledge tree via "topology + semantic" dual-perception community detection; the retrieval end uses the same schema to decompose complex questions into atomic sub-queries with iterative reflection. It saves up to 33.60% tokens and improves accuracy by 16.62% across 6 benchmarks compared to SOTA. DeepRAG formalizes the "reasoning while retrieving" process as a Markov Decision Process (MDP). It enables LLMs to autonomously decide whether to "use internal knowledge or perform external retrieval" for each sub-problem during step-by-step problem decomposition. Through a three-step pipeline—binary tree search for data synthesis, imitation learning, and calibration training—DeepRAG achieves a 25.41% relative improvement in answer accuracy across five QA datasets while significantly reducing retrieval frequency. This paper introduces AMemGym, the first on-policy interactive evaluation benchmark for long-horizon conversation memory. By utilizing structured data sampling (User Persona → State Evolution → Personalized QA), it drives LLMs to simulate users in role-play scenarios. The study reveals ranking bias issues in traditional off-policy evaluations and systematically diagnoses the "write/read/utilization" three-stage failure modes in RAG, long-context, and Agent memory systems. MILCO employs a "multilingual connector + English MLM head" to project text from 39+ languages into a shared English vocabulary sparse space. Combined with "Sparse Alignment Pre-training" to prevent semantic collapse and a "LexEcho dual-view" to recover rare entities lost during translation, a single 560M sparse model outperforms dense, sparse, and multi-vector baselines such as BGE-M3 and Qwen3-Embed in both multilingual and cross-lingual retrieval tasks. Browse all 81 Information Retrieval & RAG papers → 💻 Code Intelligence (59) ¶ Proposes Intuitor, an RLIF method that replaces external verifiable rewards with the model's own self-certainty (KL divergence between the output distribution and a uniform distribution). It matches GRPO performance in mathematical reasoning while demonstrating better generalization in out-of-distribution tasks such as code generation. The paper proposes ReasoningBank, a memory framework that distills generalizable reasoning strategies from success and failure experiences judged by the agent itself. It also introduces memory-aware test-time scaling (MaTTS) to establish a synergy between memory and test-time expansion, consistently outperforming baselines on WebArena, Mind2Web, and SWE-Bench (up to 34.2% relative gain) while reducing interaction steps by 16%. This paper trains a 7B masked diffusion code model, DiffuCoder, proposes a local/global AR-ness metric system to characterize the "non-autoregressive" decoding behavior of diffusion LLMs (dLLMs), and designs coupled-GRPO (a diffusion-native RL method using complementary mask coupled sampling), achieving a 4.4% improvement on EvalPlus. PaperCoder is proposed—a multi-agent LLM framework that automatically transforms machine learning papers into runnable code repositories via a three-stage pipeline of Planning, Analysis, and Coding. 88% of the generated repositories were rated as best by the original authors, significantly outperforming baselines on the PaperBench benchmark. The authors construct Ambig-SWE (an underspecified variant based on SWE-Bench Verified) to systematically evaluate the interaction capabilities of LLM programming agents across three dimensions: detecting underspecification, asking clarifying questions, and utilizing interactive information. They find that interaction can improve resolution rates in underspecified scenarios by up to 74%, yet models default to non-interactive behavior and struggle to distinguish between sufficient and underspecified instructions. This paper proposes treating Agentless (workflow-style) training as a "skill prior" for SWE-Agents (multi-turn interactive). By utilizing a recipe of mid-training + cold-start + RL + test-time self-play, the open-source model Kimi-Dev achieves 60.4% on SWE-bench Verified (a SoTA for workflow solutions). It is further upgraded into an agent with 48.6% pass@1, comparable to Claude 3.5 Sonnet, using a lightweight SFT of 5k trajectories. VERINA uses 189 manually refined Lean programming tasks to decompose "verifiable code generation" into three independent yet combinable base tasks: CodeGen, SpecGen, and ProofGen. It provides a multi-stage specification evaluator combining "theorem proving + full coverage testing." Results show that even the strongest o3 achieves only 72.6% code correctness and 52.3% specification success, while the proof success rate is as low as 4.9%. Instead of using the LLM as a direct "player," it is tasked with translating game rules and a few match trajectories into executable Python Code World Model (CWM) (including state transitions, legal actions, terminal state detection, plus value functions and hidden state inference functions). This code is then processed by classical planners like MCTS/ISMCTS for deep search. Across 10 games (including 4 entirely new OOD games), the approach tied with or outperformed Gemini 2.5 Pro in 9 games. RECODE-H transforms "research code generation" from a one-shot task into multi-turn human-computer collaboration: it features 102 repository-level tasks from real top-tier conference papers and official repositories, equipped with unit tests and a five-level feedback hierarchy. Using ReCodeAgent (multi-turn ReAct + memory compression) as a strong baseline, the study systematically quantifies how "finer feedback leads to more accurate LLM corrections"—the Recall of GPT-5 increases from 29.4% without feedback to 71.6% with the strongest feedback. QLCoder embeds an LLM-Agent into an iterative loop of "candidate query generation \(\to\) CodeQL execution scoring \(\to\) feedback-based patching." It constrains reasoning using a custom MCP toolkit (CodeQL Language Server for syntax consistency + RAG vector database for semantic grounding) to automatically synthesize CodeQL queries from CVE metadata that "alert on vulnerable versions and stay silent on fixed versions." It achieved a 53.4% success rate and an F1 score of 0.70 across 176 real-world Java CVEs, significantly outperforming vanilla Claude Code (10%) and existing IRIS/CodeQL query suites (F1 0.048 / 0.073). Browse all 59 Code Intelligence papers → 🎨 Image Generation (352) ¶ The long-used VAE in latent diffusion is replaced with a "frozen pretrained representation encoder (DINOv2 / SigLIP2 / MAE) + a trained lightweight ViT decoder." By implementing three specific modifications for high-dimensional latents, the Diffusion Transformer is successfully adapted, achieving an unconditional FID of 1.51 and a guided FID of 1.13 on ImageNet 256×256. The convergence speed is 47× faster than SiT and 16× faster than REPA. DiffusionNFT proposes a novel online RL paradigm for diffusion models: instead of performing policy optimization on the reverse sampling process (e.g., GRPO), it utilizes a contrastive training approach on the forward process via a flow matching objective for positive and negative samples. This defines an implicit policy improvement direction, achieving 3-25× speedups over FlowGRPO while eliminating the need for CFG. TempFlow-GRPO identifies that "treating all denoising steps equally" in existing Flow-GRPO training is a core bottleneck. By using a triad of "process rewards via trajectory bifurcation + noise-level reweighting + seed grouping," it matches optimization intensity to the real exploration potential of each step, achieving SOTA on GenEval and PickScore with significantly fewer steps (GenEval 0.63→0.97, ~10× training efficiency). SVG is proposed to replace the VAE latent space with frozen DINOv3 self-supervised features for building diffusion models. By supplementing fine-grained details via a lightweight residual encoder, it achieves faster training, more efficient inference, and cross-task universal visual representations. PixNerd replaces the final linear projection of the Diffusion Transformer with a "per-patch implicit neural field head" that dynamically generates weights from Transformer features. This head decodes fine-grained pixels within large patches, enabling single-stage, end-to-end diffusion in the original pixel space without relying on VAEs or cascaded multi-scale architectures. It achieves a 1.93 FID on ImageNet \(256 \times 256\) with nearly 8x lower latency than previous pixel-space diffusion models. This paper systematically proves that Representation Alignment (REPA) accelerates diffusion model training not by relying on the global semantic information of the target representation (ImageNet linear probe accuracy), but rather on the spatial self-similarity structure between its patch tokens . Based on this, the authors propose iREPA with only 4 lines of code (convolutional projection + spatial normalization), which consistently accelerates REPA convergence across 27 encoders, various model scales, and training recipes. The authors propose rCM (score-regularized continuous-time consistency model), which extends continuous-time consistency distillation to 14B parameter text-to-image/video models for the first time. By combining forward divergence (consistency) and backward divergence (score distillation), the method matches the quality of DMD2 while preserving diversity, achieving 15-50× acceleration. The authors propose the first systematic "benchmark evaluation → reward model → reinforcement learning training" pipeline for image editing: constructing the EditReward-Bench benchmark, training the EditScore series of reward models (7B-72B, outperforming GPT-5), and successfully applying it to Online RL training to significantly enhance the performance of image editing models. A 14B causal Transformer is utilized to perform next-token prediction directly on continuous image tokens , paired with a lightweight 157M flow matching head as a sampler. Without relying on heavy diffusion backbones or vector quantization, this approach achieves pure autoregressive text-to-image quality comparable to top-tier diffusion models. Constructs a high-quality dataset, EditReward-Data, containing 200K expert-annotated preference pairs, and trains the EditReward model. This model achieves SOTA human alignment across multiple image editing benchmarks and significantly improves downstream editing model performance when used as a data filter. Browse all 352 Image Generation papers → 🎬 Video Generation (97) ¶ LongLive utilizes a frame-level causal autoregressive framework combined with a trio of features—KV-recache, streaming long training (train-long-test-long), and short-window attention with frame-level attention sinks. It fine-tunes a 1.3B short-video model within 32 GPU-days into an interactive long-video generator capable of real-time generation (20.7 FPS) on a single H100, supporting real-time prompt switching and video lengths up to 240 seconds. Rolling Forcing transforms frame-by-frame autoregressive video diffusion into a rolling multi-frame joint denoising process, utilizing initial frame attention sinks to anchor global appearance. This achieves near 16 FPS real-time generation of multi-minute long videos on a single GPU while significantly suppressing long-term error accumulation. Self-Forcing++ utilizes a short-video bidirectional diffusion teacher as a "short-window error corrector." It performs Extended DMD training by randomly sampling degraded segments from long-video trajectories generated by the student, combined with a rolling KV cache and optical flow rewards. This extends a 1.3B autoregressive video model from 5-second to 100-second and even 4-minute generation while significantly mitigating overexposure, darkening, stagnation, and error accumulation. VideoPhy-2 utilizes 3,940 multi-event prompts derived from 197 real-world actions. Generated videos from modern text-to-video models are scored by humans across three axes: Semantic Adherence, Physical Commonsense, and Physical Rules. The results reveal that even the strongest model, Wan2.2-27B-A14B, achieves only \(47.7\%\) joint performance on the hard subset. Furthermore, a 7B VideoPhy-2-AutoEval evaluator was trained to reduce human evaluation costs. SANA-Video replaces the full attention in video DiTs with linear attention to reduce complexity from \(O(N^2)\) to \(O(N)\) . By leveraging the additive property of linear attention, a "constant memory" block autoregressive KV cache is designed. This allows a 2B model to be trained on 64 H100s in 12 days (only 1% of MovieGen's cost), producing 720×1280 minute-long videos that match Wan2.1-14B on VBench while being 16× faster during inference. The authors reframe long video generation as an "internal information retrieval" problem and propose Mixture of Contexts (MoC) — a parameter-free yet trainable sparse attention routing module. It allows each query to dynamically select a few relevant chunks plus mandatory anchors (text + local window) while using causal masking to avoid feedback loops. This maintains or even improves identity/motion/scene consistency in minute-long videos while pruning 85% of token pairs and reducing attention FLOPs by 7×. UniVideo utilizes a frozen MLLM for multimodal understanding and instruction parsing, and an MMDiT for high-fidelity image/video generation. It unifies video understanding, text-to-video, image-to-video, in-context video generation, and mask-free video editing into a single natural language instruction framework, achieving performance comparable to or better than specialized models across multiple video generation and editing tasks. By aligning the intermediate features of a video diffusion model to the geometric representations of the 3D foundation model VGGT (using decoupled angular and scale alignment objectives), the diffusion model trained on pure video data "internalizes" 3D structures. This significantly improves geometric and temporal consistency in long-term video generation and enables the extraction of explicit 3D geometry during inference. Ours proposes JavisDiT, a Joint Audio-Video Diffusion Transformer model that achieves fine-grained audio-visual spatio-temporal alignment through a Hierarchical Spatio-Temporal Prior Synchronizer (HiST-Sypo). Additionally, a new benchmark, JavisBench (comprising 10K complex scene samples), and a new evaluation metric, JavisScore, are introduced. The authors propose MotionStream, the first real-time streaming video generation system with motion control. The method trains a bidirectional motion-controlled teacher with a lightweight track head, then distills it into a causal student via Self Forcing and DMD. By introducing attention sinks and a rolling KV cache, the system achieves a total match between training and inference distributions. It reaches 17 FPS (29 FPS with Tiny VAE) at 480P on a single H100 GPU, supporting infinite-length generation at constant speed. Browse all 97 Video Generation papers → 🧩 Multimodal VLM (211) ¶ This paper proposes HiCo, a hierarchical video token compression method that reduces long video context from the clip level to the video level by approximately \(1/50\) (averaging only 16 tokens per frame). Combined with a multi-stage short-to-long training strategy, the LongVid dataset containing 114K long videos, and a more challenging multi-hop NIAH evaluation, VideoChat-Flash achieves better performance than GPT-4o / Gemini-1.5-Pro on both short and long video benchmarks at the 7B scale, while achieving 99.1% accuracy on the 10,000-frame NIAH test. Six 3D vision researchers spent 300+ hours manually crafting 1,000 multi-image spatial reasoning multiple-choice questions from 120,000 real images to form MMSI-Bench. Among 37 mainstream MLLMs, the strongest open-source model scores only 30%, GPT-5 reaches 41.9%, while humans achieve 97%. It also provides an automated error diagnosis pipeline leveraging manual reasoning annotations. WorldSense is the first real-world omnimodal video understanding benchmark that mandates audio-visual synergy . It comprises 1,662 synchronized audio-visual segments and 3,172 multiple-choice questions, each designed such that "it cannot be answered correctly if either audio or video is removed." Results show that even the strongest Gemini 2.5 Pro achieves only 65.1% accuracy, while most open-source omnimodal models perform close to random guessing. Without modifying the VLM architecture or introducing point cloud modalities, this work augments 2D vision-language models with 3D indoor scene understanding capabilities using only a visual prompting paradigm consisting of "videos + reconstructed Bird’s Eye Views (BEV) + cross-frame consistent object ID markers." It achieves SOTA performance in both zero-shot and fine-tuning settings. DeepEyesV2 aims to truly weave "external tool calling" into the inference process of multimodal models. It allows models to autonomously decide when to write Python code or initiate a web search within a single inference trajectory, backfilling tool outputs for further reasoning. The authors found that pure RL fails to learn stable tool calling; thus, they proposed a "cold-start SFT + reinforcement learning" two-stage training approach. This resulted in consistent improvements across perception, mathematical reasoning, and search benchmarks (e.g., MathVerse +7.1; MMSearch 63.7%, significantly surpassing the 53.8% of specialized search models). StreamingVLM utilizes a unified framework of "overlapping short segments during training and reusing compact KV cache during inference," enabling 7B-class VLMs to maintain low latency, long-range memory, and second-level real-time commentary capabilities over hours of video streams. This paper proposes the VLMBias counterfactual visual evaluation framework, which systematically modifies iconic visual elements in animals, logos, flags, chessboards, game boards, optical illusions, and patterned grids. It finds that mainstream VLMs achieve an average accuracy of only 17.05% on objective counting tasks, with 75.70% of responses reverting to commonsense priors rather than visual evidence. VL-JEPA replaces the traditional autoregressive token generation of VLMs with non-autoregressive prediction of semantic embeddings for target text. Under identical training settings, it consumes fewer parameters and converges faster than token-space VLMs, while naturally supporting selective decoding for classification, retrieval, VQA, and online video scenarios. OmniVinci employs three architectural improvements for "visual-audio alignment" (OmniAlignNet semantic alignment, temporal grouping, and Constrained Rotary Time Embedding) alongside a data pipeline capable of synthesizing 24 million dialogues. Trained on only 0.2T tokens, this open-source omni-modal LLM understands video, audio, speech, and text simultaneously, outperforming Qwen2.5-Omni (while using only 1/6 of its training tokens) across multiple cross-modal, audio, and visual benchmarks. RECA treats the visual understanding embeddings of a Unified Multimodal Model (UMM) as dense "visual prompts." By re-aligning the understanding and generation branches through post-training with unlabeled image reconstruction, it significantly improves 1.5B UMMs on GenEval, DPGBench, and image editing benchmarks without requiring extra captions, GPT-4o distillation, or reinforcement learning. Browse all 211 Multimodal VLM papers → 🧠 VLM Reasoning (112) ¶ Thyme enables Multimodal Large Language Models (MLLMs) to autonomously generate and execute code during reasoning for image operations (e.g., cropping, rotation, contrast enhancement) and mathematical calculations. This capability is activated through a two-stage training process: a "500k SFT cold start" followed by "GRPO-ATS reinforcement learning," consistently outperforming Qwen2.5-VL baselines across nearly 20 benchmarks, particularly in high-resolution perception and complex reasoning. Mini-o3 employs a triplet of "hard-sample dataset + diverse cold-start trajectories + over-turn masking" to enable a VLM trained on only 6 interaction turns to naturally extend to dozens of trial-and-error exploration turns during inference, reproducing OpenAI o3-style deep visual search capabilities and achieving new SOTA results. OmniSpatial is the first comprehensive spatial reasoning benchmark built on cognitive psychology. It systematically covers 4 dimensions and 50 subcategories (Dynamic Reasoning, Complex Spatial Logic, Spatial Interaction, and Perspective Transformation) with 8.4K human-annotated QA pairs. Results show that the o3 model achieves only 56.33% compared to a human score of 92.63%→ revealing that complex spatial reasoning remains a core bottleneck for VLMs. Identifying that 67% of errors in multimodal RLVR stem from the neglected bottleneck of "inaccurate visual perception," this paper proposes PAPO. It introduces an implicit perception KL loss between "original vs. masked images" (plus double entropy regularization) into the GRPO/DAPO optimization objective. Without additional annotations, reward models, or teacher models, PAPO achieves an overall improvement of 4.4%–17.5% across 8 multimodal reasoning benchmarks and a 30.5% reduction in perception errors. Constructed ZEBRA-COT, the first large-scale diverse interleaved text-image reasoning dataset (182K reasoning trajectories across 18 domains). Scaffolding experiments demonstrate that visual CoT has a potential improvement of up to +43% for frontier models, and fine-tuning enables Anole-7B and Bagel-7B to acquire endogenous visual reasoning capabilities. The MindCube benchmark (21,154 questions / 3,268 images) is proposed to systematically expose the deficiency of VLMs in "reconstructing unseen spaces from limited views," where they perform near random guessing. A "map-then-reason" (SFT + RL) scheme is introduced, where the model first draws a cognitive map and then reasons upon it, improving the accuracy of Qwen2.5-VL-3B from 37.8% to 61.3%. The study proposes VTool-R1, the first framework that trains VLMs via Reinforcement Learning Fine-tuning (RFT) to generate interleaved textual and visual intermediate reasoning steps, enabling models to "think with images." LVR enables Multimodal Large Language Models (MLLMs) to move beyond "thinking" solely in text space. Instead, it uses the LLM's last hidden states to autoregressively reconstruct question-related visual semantics directly in the visual embedding space ("think before speaking"). Combined with a modified GRPO reinforcement learning, this approach significantly outperforms the "Think about/with Images" paradigms on perception-intensive VQA tasks. Addressing ego-centric multi-view scenarios (e.g., autonomous driving/robotics) where cameras simultaneously cover front, rear, left, and right views, this paper establishes the first outdoor 3D spatial reasoning benchmark, Ego3D-Bench (8.6K QA). It proposes Ego3D-VLM , a training-free, plug-and-play framework that localizes queried objects in 3D global coordinates to generate a compact "textual cognitive map." Feeding this map into any VLM improves MCQA accuracy by an average of 12% and reduces absolute distance RMSE by an average of 56%. VGR enables Multimodal Large Language Models (MLLMs) to "replay visual memory" during the thinking process—autonomously framing key regions during reasoning and retrieving high-resolution visual tokens to continue thinking. Coupled with a set of VGR-SFT data containing grounding signals, it significantly outperforms baselines on fine-grained image understanding tasks such as ChartQA, AI2D, and MMStar, while using only 30% of LLaVA-NeXT's visual tokens. Browse all 112 VLM Reasoning papers → ⚡ VLM Efficiency (18) ¶ SP-VLA categorizes VLA action sequences into "deliberative" and "intuitive" types. Deliberative types invoke the large model, while intuitive types are approximated by lightweight Ridge Regression. Simultaneously, spatial-semantic dual-perceptive token pruning is applied, achieving 1.5× lossless acceleration on LIBERO and 2.4× acceleration on SimplerEnv. PRUNESID is a training-free vision token compression framework that balances token semantic importance and information diversity through a two-stage pipeline consisting of "Principal Semantic Component Analysis (PSCA) clustering + Intra-group Non-Maximum Suppression (NMS)". By dynamically allocating token budgets according to image complexity, it maintains 96.3% relative accuracy on LLaVA-1.5 with only 11.1% of tokens, and achieves 92.8% relative performance under extreme compression (5.6%) on LLaVA-NeXT. The authors propose the HiDrop framework, which performs deep functional analysis of MLLM layers (Shallow = Propagators, Middle = Fusion Centers, Deep = Language Reasoning). It designs a three-stage strategy: Late Injection (skipping shallow layers), Concave Pyramid Pruning (pruning in middle layers), and Early Exit (exiting in deep layers). This approach compresses approximately 90% of vision tokens with negligible performance loss and achieves a 1.72× training speedup. It is observed that KV Cache in LVLMs exhibits modal-specific and head-specific semantic redundancy. Since selection based solely on importance leads to a loss of semantic coverage, MixKV is proposed to adaptively mix importance and diversity scores per head for KV Cache compression, achieving an average improvement of 5.1% under extreme compression. VisionTrim is a training-free acceleration framework for Multi-modal Large Language Models (MLLMs). It employs two plug-and-play modules: DVTS (selecting dominant vision tokens by balancing global semantics and local spatial continuity) and TGVC (re-clustering discarded tokens into supplementary tokens via text guidance). By compressing vision tokens during both vision encoding and LLM decoding stages, it maintains 98.8% average performance on LLaVA-1.5 while removing 88.9% of vision tokens. iLLaVA breaks the inertia of "compressing tokens only in the LLM stage" by inserting token merging into both the image encoder and the LLM . Using an "information tokens + recovery tokens" merging strategy to retrieve useful information from discarded tokens, it achieves 2× throughput and 4× prefilling acceleration training-free while maintaining >95% performance. This paper discovers that existing visual token pruning methods collapse on visual grounding (VG) tasks because they destroy the "global spatial reference system" constructed by positional encodings. Consequently, it proposes Nüwa—a two-stage pruning framework inspired by swarm intelligence (Boids) that employs a "Partition-Align-Aggregate" strategy on the vision encoder side to preserve spatial anchors, followed by text-guided refinement in the middle of the LLM. This approach improves performance retention on VG tasks from ~7% to 47% while maintaining VQA performance at 95%. Photon is a multimodal large model that directly processes full 3D medical volume data (CT/MRI). It utilizes "Instruction-conditioned Token Scheduling (ITS)" to adaptively determine the number of visual tokens to retain for each question, and "Surrogate Gradient Propagation (SGP)" to ensure discrete token dropping remains differentiable during training. This approach achieves SOTA accuracy on medical VQA while providing approximately 5x training speedup and two-thirds memory savings. This work reveals the implicit visual coordinate system (IVC tokens) established by RoPE positional encoding in LVLMs and proposes a training-free, prompt-aware vision token pruning strategy. By preserving both IVC tokens and semantic foreground tokens, it reduces visual tokens by approximately 50% while maintaining \(\ge 99\%\) of the original performance. LearnPruner empirically debunks the prevalent assumption that "attention score = token importance." It points out that [CLS] attention in vision encoders is contaminated by attention sinks, while in LLMs, only "text-to-vision" mid-layer attention is reliable. Consequently, it replaces [CLS] attention with a learnable pruning module and superimposes text-guided pruning at the LLM mid-layers. By retaining only ~5.5% of vision tokens, it maintains 95% performance and achieves a 3.2× speedup. Browse all 18 VLM Efficiency papers → 🎵 Audio & Speech (80) ¶ The authors propose MMSU (5,000 audio QAs, 47 tasks), the first speech understanding and reasoning benchmark to systematically integrate linguistic theories. Evaluating 22 SpeechLLMs reveals that existing models still exhibit significant performance gaps in phonological perception and complex reasoning. YuE scales the LLaMA2 architecture to trillions of tokens to train the first open-source "lyrics-to-song" foundation model. By employing dual-token track decoupling (separate prediction of vocals/accompaniment), structural progressive conditioning (interleaving lyrics and audio by segments), and re-designed music in-context learning , it generates songs up to 5 minutes long with aligned lyrics and vivid vocals, matching or exceeding certain commercial systems (e.g., Udio, Tiangong) in musicality. The study constructs RedTeamCUA, the first red teaming framework for CUA in hybrid Web-OS environments, along with RTC-Bench containing 864 test cases. It systematically evaluates the vulnerability of 9+ frontier CUAs to indirect prompt injection, finding that all CUAs are attackable (highest ASR 83%). Furthermore, more capable models prove more dangerous—the fact that the Attempt Rate (AR) is significantly higher than the Attack Success Rate (ASR) implies that improvements in model capabilities will directly translate into higher attack success rates. Stitch is proposed to enable "thinking while talking" in Spoken Language Models (SLMs) by interleaving silent reasoning tokens with speech tokens in chunks. It leverages the idle computation time during audio playback to perform reasoning. Stitch-S achieves first-frame latency identical to non-reasoning baselines while improving mathematical reasoning accuracy by approximately 15 percentage points. AudioX utilizes a unified framework based on a Diffusion Transformer (DiT), integrated with a lightweight "Multimodal Adaptive Fusion (MAF)" module and a self-constructed 7-million-sample multimodal dataset, IF-caps. This allows a single set of model weights to generate high-fidelity sound effects and music from arbitrary combinations of text, video, and audio, significantly outperforming specialized models in fine-grained instruction following. By constructing a 5-million-scale multi-cultural, full-length, hierarchically-annotated music dataset (MF-Skills + MF-Think) and applying a "SFT → CoT Cold Start → GRPO Reinforcement Learning" training recipe onto an enhanced Audio Flamingo 3 backbone, Music Flamingo elevates audio language models from "identifying surface attributes" to "performing hierarchical, theory-aware music reasoning like a trained musician," achieving new SOTA results on 12+ music understanding and reasoning benchmarks. This paper is the first to systematically reveal inherent safety vulnerabilities in diffusion large language models (dLLMs) caused by bidirectional modeling and parallel decoding mechanisms. It proposes the DiJA jailbreak attack framework, which achieves nearly 100% attack success rates on multiple aligned dLLMs through interleaved mask-text prompts. To fill the missing puzzle piece of "large-scale human preference corpora for naturalness" in speech synthesis, this paper introduces a comprehensive toolkit: a dataset (99K preference pairs), a benchmark (1000 high-consistency samples), and a reward model. By utilizing a "SFT cold-start + GRPO reinforcement" two-stage strategy, Qwen2.5-Omni-7B is trained into a generative reward model, SpeechJudge-GRM. It achieves a 77.2% accuracy rate (79.4% with inference-time voting) in judging speech naturalness, significantly outperforming the classic Bradley-Terry reward model (72.7%). The ParaS2S framework is proposed, comprising ParaS2SBench—a benchmark for evaluating paralinguistic awareness (emotion/sarcasm/age/gender) in Speech-to-Speech (S2S) interaction—and ParaS2SAlign, a GRPO-based RL alignment framework. It enables S2S models to learn the ability to adjust responses according to speaking styles with minimal annotated data. Ours proposes TASTE (Text-Aligned Speech Tokenization and Embedding), which aligns speech tokens with text transcriptions through a cross-attention mechanism. This achieves high-quality speech reconstruction at an extremely low bitrate (~150 bps) and makes joint text-speech modeling direct and efficient. The 1.3B parameter TASLM outperforms 7B pre-trained SLMs. Browse all 80 Audio & Speech papers → 🔎 AIGC Detection (30) ¶ This paper proposes the detection principles "All Patches Matter, More Patches Better," identifying that existing AI-generated image (AIGI) detectors suffer from a "Few-Patch Bias"—focusing only on a minimal set of patches. A Panoptic Patch Learning (PPL) framework is designed, using Randomized Patch Reconstruction (RPR) and Patch-wise Contrastive Learning (PCL) to spread discriminative power across all patches. This significantly improves cross-generator generalizability and robustness on GenImage, DRCT-2M, AIGCDetectBenchmark, and real-world Chameleon datasets (e.g., CLIP backbone achieves 97.2% mAcc on GenImage with a std of only 1.7). EditLens moves beyond binary "Human vs. AI" classification by using lightweight similarity metrics (cosine distance, soft n-grams) as intermediate supervision to fine-tune a regression model. It continuously predicts "how much the text was edited by AI," achieving SOTA performance in both binary (F1=95.6%) and ternary (macro-F1=90.4%) classification tasks. Constructed the largest AI-generated peer review dataset to date (788,984 reviews), systematically evaluated the performance of 18 AI text detection methods in peer review scenarios, and proposed the Anchor detection method leveraging original papers as context, significantly outperforming all baselines at low False Positive Rates. The DiNCo method is proposed to expose "suggestibility bias" by having LLMs independently evaluate automatically generated distractors (plausible but incorrect alternative answers). By normalizing with the total confidence across distractors and fusing two complementary dimensions—generation consistency and verification consistency—it significantly improves confidence calibration in short-form QA and long-form generation tasks. This paper explains the effectiveness of "rewrite-based" LLM text detection methods from a geometric projection perspective and proposes L2D. Instead of using fixed distances to measure the difference between original and rewritten text, L2D adaptively learns a distance function , achieving an average improvement of 41.5%~75.4% over the strongest baselines across 100+ settings. This paper introduces ArEnAV , the first large-scale audio-visual deepfake dataset targeting "Arabic-English intra-sentential code-switching (CSW)" (387k videos, 765+ hours). Utilizing an integrated generation pipeline with 4 TTS paths and 2 lip-sync models, the authors perform "content-driven" semantic manipulation of real YouTube videos. They systematically demonstrate that existing SOTA detection/localization models and human evaluators almost entirely fail in these multilingual, code-switching scenarios. Addressing the issue where existing AI-generated video detectors destroy critical forgery artifacts by scaling or cropping input frames to a fixed low resolution (e.g., 224×224), this paper proposes a "native-scale" detection framework. Based on Qwen2.5-VL, the visual Transformer directly processes videos at arbitrary original resolutions and durations. The work also constructs a 140k training set covering 15 generators and a high-fidelity benchmark, Magic Videos, achieving new SOTA results across multiple benchmarks. The authors propose HiDA-Net: a dual-path input architecture using "global thumbnail + full-resolution patches covering the whole image," combined with feature aggregation, token-level forgery localization, and JPEG quality factor estimation. It achieves "no pixel left behind" and significantly advances the SOTA in high-resolution AIGI detection. Focusing on "looks real but defies logic" semantic anomalies in AI-generated images (violating physics, common sense, or anatomy), this paper formalizes the task of "detection + explanation + scoring." By utilizing the multi-agent pipeline AnomAgent and lightweight human verification, the authors curate the AnomReason benchmark with 21.5K images and hundreds of thousands of structured quadruplet annotations. They propose semantic matching metrics SemAP/SemF1; the fine-tuned AnomReasonor-7B outperforms all open-source VLMs in semantic detection, approaching the performance of GPT-4o. Through close reading annotations of 8,618 expressions by 26 professional writers, this study reveals that n-gram novelty is insufficient to measure textual creativity—approximately 91% of expressions with high n-gram novelty are not considered creative, and a negative correlation exists between high n-gram novelty and low pragmaticality in open-source LLMs. Browse all 30 AIGC Detection papers → 🧊 3D Vision (195) ¶ The state update of the recurrent 3D reconstruction model CUT3R is reformulated as a test-time online learning problem. By deriving a closed-form, per-token adaptive learning rate based on the alignment confidence between memory states and new observations to gate state updates, this method significantly mitigates long-sequence forgetting without retraining or adding parameters. It improves global pose accuracy by 2× over the baseline while remaining efficient (20 FPS, 6 GB VRAM) for sequences of thousands of images. STREAM3R reformulates dense 3D reconstruction as a "frame-by-frame causal attention problem in a decoder-only Transformer." Whenever a new image arrives, it performs causal cross-attention with the cached historical frame features to regress pointmaps. This enables incremental online reconstruction using KVCache and sliding window attention similar to LLMs, achieving performance superior to or comparable with existing streaming methods in depth estimation and 3D reconstruction for both static and dynamic scenes, while offering faster inference. Scenethesis is a training-free agentic framework that utilizes LLMs to draft coarse layouts, vision foundation models for visual grounding and scene graph extraction, and a physics-aware optimizer (semantic correspondence + SDF contact/support constraints) for object-wise pose correction. A GPT-5 judge verifies spatial consistency and triggers re-planning, enabling the generation of collision-free, stable, and interactive 3D scenes for both indoor and outdoor environments. HoloPart introduces the concept of "amodal segmentation" from 2D to 3D, proposing the new task of "3D Part Amodal Segmentation"—decomposing a global mesh into geometrically complete semantic parts (rather than fragmented surface patches) using a diffusion model that incorporates local attention and global shape context for part completion. The authors construct OmniWorld, a 4D world modeling dataset spanning four domains (simulator, robot, human, and internet) with over 300 million frames, featuring five modalities: depth, camera pose, text, optical flow, and foreground masks. By combining self-collected game engine data with 12 public datasets and an automated annotation pipeline, they demonstrate that fine-tuning existing SOTA models on OmniWorld leads to significant gains in 3D geometric reconstruction and camera-controllable video generation. Constructs OpenFly, a comprehensive platform for Aerial Vision-Language Navigation (VLN): integrates 4 rendering engines (UE/GTA V/Google Earth/3DGS); develops a fully automatic data generation toolchain (point cloud acquisition → semantic segmentation → trajectory generation → GPT-4o instructions); builds a large-scale dataset of 100,000 trajectories across 18 scenes; proposes a keyframe-aware VLN model, OpenFly-Agent (Keyframe Selection + Visual Token Merging), which outperforms existing methods by 14.0% and 7.9% in Success Rate (SR) for seen and unseen scenes, respectively. Prior Depth Anything employs a two-stage "coarse-to-fine" pipeline to fuse precise but sparse metric depth priors measured by sensors with complete but relative geometric structures predicted by monocular depth models. A single model unifies three tasks—depth completion, super-resolution, and inpainting—in a zero-shot manner, matching or even exceeding the performance of specialized SOTAs across 7 real-world datasets. Human3R freezes the online 4D reconstruction foundation model CUT3R and uses Visual Prompt Tuning (VPT) to insert "human prompts." This allows the model to simultaneously output multi-person SMPL-X meshes (everyone), dense scene point clouds (everywhere), and camera trajectories (all-at-once) in a single feed-forward pass at 15 FPS with 8 GB VRAM, reaching SOTA after training on a single GPU for just one day. YoNoSplat uses a feedforward model to directly predict per-view local 3D Gaussians, camera poses, and intrinsics from an arbitrary number of unposed and uncalibrated multi-view images, which are then aggregated into a global scene. By employing a "mix-forcing" training strategy, pairwise distance normalization, and Intrinsic Condition Embedding (ICE), it resolves pose-geometry entanglement and scale ambiguity. It achieves SOTA performance in both posed and unposed settings, reconstructing a scene from 100 images in 2.69 seconds. Trace Anything represents every pixel in a video as a continuous 3D trajectory and directly predicts the trajectory field of the entire video through a single feed-forward inference. This achieves efficient 4D dynamic scene representation without requiring depth, optical flow, 2D trackers, or per-scene optimization. Browse all 195 3D Vision papers → 🎯 Object Detection (30) ¶ RF-DETR utilizes DINOv2 internet-scale pre-training combined with end-to-end weight-sharing NAS to train a "supernet." This approach allows a single training session to export an entire accuracy-latency Pareto curve via grid search without retraining. It is the first real-time detector to exceed 60 AP on COCO and outperforms GroundingDINO on the real-world dataset RF100-VL with a 20x speedup. This paper proposes TreeBench (the first traceable visual reasoning benchmark, consisting of 405 highly challenging VQA tasks where OpenAI-o3 achieves only 54.87%) and TreeVGR (a training paradigm that jointly supervises grounding and reasoning via reinforcement learning with Dual IoU rewards). The TreeVGR-7B model achieves gains of +16.8 on V*Bench, +12.6 on MME-RealWorld, and +13.4 on TreeBench, demonstrating that traceability is crucial for advancing visual reasoning. It is demonstrated through linear probing experiments that CLIP's Bag-of-Words (BoW) behavior originates from cross-modal alignment failure rather than a lack of binding information in the encoders. LABCLIP is proposed, which significantly restores attribute-object binding capabilities by training only a lightweight linear transformation. ASEntmax (Adaptive-Scalable Entmax) is proposed, replacing softmax attention with \(\alpha\) -entmax using learnable temperature. The work theoretically and experimentally proves that sparse attention achieves \(1000\times\) length extrapolation, resolving the attention dispersion problem of softmax under long contexts. To address category, size, and location biases in object detection training data, this paper proposes a "Diagnosis-Synthesis" debiasing framework. It identifies truly under-represented data groups using a Representation Score (RS) that goes beyond frequency. It recalibrates layouts based on RS and synthesizes high-fidelity samples using a Visual Blueprint (color rectangle pixel conditions) combined with Dual Generative Alignment . This approach improves rare classes by 3.6 mAP and large objects by 4.4 mAP, achieving a layout accuracy 15.9 mAP higher than the previous L2I SOTA. DeCo-DETR decouples "online text encoder invocation" and the "competition between localization and alignment" in open-vocabulary detection. It employs an LVLM to offline distill a reusable hierarchical semantic prototype pool as a substitute for the text encoder during inference and utilizes dual-stream gradient isolation to separate localization and semantic alignment training. This approach achieves a gain of 3.1--5.8 points on OV-COCO novel classes while compressing single-image inference latency to 135ms. Ours proposes PaAno, a lightweight time-series anomaly detection method based on patch-level representation learning. It utilizes a 1D-CNN encoder with triplet loss and pretext loss to learn a patch embedding space. Anomaly scores are calculated via the distance to normal patches stored in a memory bank. It achieves comprehensive SOTA results on the TSB-AD benchmark with only 0.3M parameters and inference times in seconds. Aiming at the weakly supervised task of "training rotated object detectors with only a single point annotation," this paper proposes Point2RBox-v3. It utilizes Progressive Label Assignment (PLA) to feed scale information from pseudo-labels into multi-level FPN label assignment and adopts Prior-Guided Dynamic Mask Loss (PGDM-Loss) to use SAM for sparse scenes and Watershed for dense scenes. It achieves a new SOTA on six remote sensing benchmarks including DOTA-v1.0 (66.09% for the two-stage version). WS-COC is the first framework to utilize Multimodal Large Language Models (MLLM) for weakly-supervised class-agnostic object counting. Using only image-level total counts for supervision, it activates the counting capabilities of MLLMs through three simple strategies: "Binary Dialogue Tuning + Comparative Ranking Optimization + Global-Local Fusion." It approaches or even surpasses some fully-supervised methods using point-level supervision on four datasets, including FSC-147. Proposed D24FAD, a dual-distillation framework combining Teacher-Student Distillation (TSD) on query images and Student Self-Distillation (SSD) on support images, supplemented by a Learn-to-Weight (L2W) mechanism for adaptive support evaluation. It achieves 100% AUROC using 2-shot on the APTOS fundus dataset. Browse all 30 Object Detection papers → ✂️ Segmentation (31) ¶ SAM 3 unifies "finding and segmenting all instances of a concept in images/videos" (Promptable Concept Segmentation, PCS) into a single model. By using noun phrases or visual exemplars as prompts, it outputs masks and cross-frame identities for all matching instances via a shared backbone + detector + memory tracker. Supported by a human-AI collaborative data engine producing a training set with 4M concept labels, SAM 3 doubles the accuracy of existing systems in both image and video PCS. This paper introduces Segment Concept (SeC), which injects object-level "concept representations" extracted by Large Vision-Language Models (LVLMs) into a SAM 2.1-style Video Object Segmentation (VOS) pipeline on demand. This approach significantly reduces appearance-based interference and object reappearance failures in complex multi-shot scenarios while establishing the SeCVOS benchmark specifically for evaluating semantic-level VOS capabilities. The VINCIE framework is proposed, demonstrating for the first time that in-context image editing models can be learned entirely from native video data. By annotating videos as interleaved multimodal sequences and designing three proxy tasks (NIP/CSP/NSP), it reaches SOTA on multi-turn editing benchmarks, increasing the 5-turn editing success rate from <2% to 25% compared to baselines. This work formalizes Referring Video Object Segmentation (RVOS) as an ODE flow problem that continuously deforms video latent representations into masks under language guidance. By fine-tuning the pre-trained text-to-video (T2V) model Wan2.1 and employing three strategies focused on the trajectory starting point, the method achieves SOTA performance on MeViS, Ref-YouTube-VOS, and Ref-DAVIS17. To address the issues of expensive labeling, data scarcity, and fragmented sub-tasks (DIS / HR-SOD) in Salient Object Detection (SOD), this paper proposes a multimodal diffusion pipeline to simultaneously generate images and pixel-level masks. By incorporating iterative generation with hard-example feedback, the authors create S3OD, a high-resolution synthetic dataset of 139,000 images. Coupled with an ambiguity-aware multi-mask decoder, models trained solely on synthetic data reduce cross-dataset errors by 20–50% and achieve SOTA results on DIS and HR-SOD after fine-tuning. This paper analyzes the interaction friction between the [CLS] global token and local patch tokens in Vision Transformers. It observes that normalization layers implicitly differentiate these two token categories. By introducing specialized processing paths in normalization layers and early QKV projections, the authors achieve a segmentation performance gain of over 2 mIoU with only an 8% increase in parameters, while maintaining classification accuracy. The study identifies that SAM2 exhibits sparse perception patterns similar to biological vision (the decoder focuses on the foreground while the encoder computes globally; only a few tokens in memory frames are effective and remain temporally consistent in saliency). Based on this, Efficient-SAM2 is proposed, eliminating redundant computation through Object-Aware Sparse Window Routing (SWR) and Sparse Memory Retrieval (SMR). This achieves a 1.68× end-to-end acceleration on SAM2.1-L with only a 1% accuracy loss. It is discovered that the self-attention of text-to-image diffusion models exhibits an "Instance Emergence Point" (IEP) during the denoising process, where the self-attention shows intense divergence at object boundaries. TRACE generates high-quality instance edges through IEP localization + ABDiv edge extraction + single-step distillation, achieving 81× inference acceleration. Without any instance annotations, it improves unsupervised instance segmentation by +5.1 AP, and its tag-supervised panoptic segmentation outperforms point-supervised methods by +1.7 PQ. RegionReasoner is proposed as a multi-round visual reasoning framework based on reinforcement learning. By utilizing reference grounding rewards and global-local consistency rewards, the model is compelled to explicitly reference coordinates of specified regions and maintain semantic coherence throughout reasoning trajectories. This results in significant improvements in multi-round localization and segmentation accuracy on the newly constructed RegionDial-Bench. This paper defines a new task, "Urban Socio-Semantic Segmentation" (segmenting entities like schools and parks defined by social attributes rather than visual appearance from satellite imagery), constructs the SocioSeg dataset (unifying heterogeneous geospatial data into a single rendered digital map layer), and proposes the SocioReasoner framework. SocioReasoner mimics the human annotator's two-stage reasoning process of "localization, rendering feedback, and refinement" using a VLM, and optimizes this non-differentiable prompt generation pipeline end-to-end via GRPO reinforcement learning, outperforming SOTA models across three-level hierarchical tasks while demonstrating strong zero-shot generalization. Browse all 31 Segmentation papers → 🖼️ Image Restoration (61) ¶ This paper proposes ContinuousSR, which reconstructs a low-resolution image into a continuous 2D Gaussian field at once via the "pixel-to-Gaussian" paradigm. Any subsequent magnification is achieved through a single fast rendering (approx. 1ms), surpassing SOTA in quality across seven benchmarks (+0.18 dB on Manga109) while achieving a 19.5× speedup when continuously scaling across 40 levels.

SeedVR2 compresses a multi-step diffusion-based video restoration model into a one-step generator via diffusion adversarial post-training. It utilizes adaptive window attention, progressive distillation, and discriminator feature matching loss to support high-resolution video restoration, achieving perceptual quality comparable to or better than multi-step models in a single inference step. Flower transforms a pre-trained flow-matching generative model into a linear inverse problem solver. At each time step, it predicts the clean destination, applies a proximal projection for data consistency using the observation operator, and then advances along the flow trajectory. It achieves superior results compared to existing flow-based solvers on image restoration tasks such as denoising, deblurring, super-resolution, and inpainting. InterActHuman is proposed to achieve audio-driven video generation in multi-person and human-interaction scenes through an automated spatio-temporal layout-inferring mask predictor and an iterative mask-guiding strategy, supporting independent lip-sync and body movements for each character. RestoreVAR adapts the visual autoregressive model VAR from pure image generation into an all-in-one image restoration model. It utilizes continuous latents of degraded images as cross-attention conditions, further supplemented by a latent refiner and a continuous latent decoder to recover details. It achieves superior restoration quality among generative AiOR methods while reducing the multi-second inference time of LDM-based methods to approximately 0.201 seconds. The paper proposes Solution Flow Models (SoFlow), which directly learn the solution function \(f(x_t, t, s)\) of a velocity ODE (mapping \(x_t\) at time \(t\) to the solution at time \(s\) ). Trained from scratch using a combination of Flow Matching loss and a JVP-free solution consistency loss, it achieves superior 1-NFE FID compared to MeanFlow on ImageNet 256 (XL/2: 2.96 vs 3.43). UniRestorer hierarchically organizes the image degradation space into multi-granularity degradation groups and trains corresponding MoE restoration experts. By jointly routing via degradation estimation and granularity estimation, the universal restoration model leverages fine-grained degradation priors while remaining robust against incorrect degradation estimation. DiffusionBlocks is proposed to interpret the layer-wise updates of residual networks as discretization steps of a continuous-time diffusion process. This allows partitioning the network into blocks that can be trained completely independently, reducing training memory by a factor of \(B\) (the number of blocks) while maintaining performance competitive with end-to-end training. Vivid-VR performs generative video restoration by attaching a ControlNet to a pre-trained T2V diffusion Transformer (CogVideoX1.5-5B). It utilizes a "concept distillation" training strategy, where the T2V model synthesizes its own text-aligned training data to suppress distribution drift during fine-tuning. Combined with a lightweight control feature projector and a dual-branch connector, it achieves more realistic textures and robust temporal consistency across real, synthetic, and AIGC videos. This paper establishes a "standard ruler" for Diffusion Posterior Sampling (DPS) algorithms: by utilizing Lévy process signals—which allow for exact Gibbs sampling—as the test distribution, it obtains "gold standard" posterior samples at the distribution level. The authors systematically evaluate mainstream DPS algorithms (C-DPS / DiffPIR / DPnP) across four types of inverse problems (denoising, deconvolution, inpainting, and partial Fourier reconstruction) using MMSE optimality gap and posterior coverage metrics. The conclusion reveals that these algorithms are generally not calibrated . Browse all 61 Image Restoration papers → 🛰️ Remote Sensing (11) ¶ Earth-Agent is the first Earth Observation (EO) Agent framework based on the Model Context Protocol (MCP) tool ecosystem. it unifies RGB and spectral remote sensing data, achieving cross-modal, multi-step, and quantitative spatio-temporal reasoning by dynamically invoking 104 expert tools. The proposed Earth-Bench benchmark includes 248 expert tasks and 13,729 images. Experiments demonstrate that Earth-Agent significantly outperforms general-purpose Agents and remote sensing MLLMs. TerraFM is designed for multisensor Earth observation data, treating Sentinel-1 SAR and Sentinel-2 optical imagery as natural augmented views of the same location. Through modality-specific patch embedding, per-position cross-attention fusion, and dual-centering DINO training for long-tail land cover, it achieves strong generalization on classification and segmentation tasks in GEO-Bench and Copernicus-Bench. This paper proposes "Perceptually-Grounded Geospatial Chain-of-Thought" (Geo-CoT), which decomposes the analysis process of remote sensing VLMs into three steps: "Planning → Grounding Evidence → Synthesis." Each step anchors assertions to specific pixel regions using bounding boxes. By constructing the Geo-CoT380k dataset with 380,000 structured reasoning entries and employing two-stage alignment (SFT for cognitive structure and GRPO for faithfulness refinement), the resulting RSThinker significantly outperforms existing SOTA models across over ten remote sensing tasks, including visual grounding, counting, detection, captioning, and VQA. This paper presents the first systematic measurement of the Intrinsic Dimension (ID) of Geographic Implicit Neural Representations (Geographic INR). It finds that the true ID of 256-512D embeddings is only 2-10. A high ID in the frozen embedding space correlates positively with downstream performance, while a low ID in the supervised task-head activation space correlates with high performance, revealing a dual mechanism of "Representativeness vs. Task-Alignment." SelvaBox constructs the largest open-access high-resolution UAV RGB tree crown detection dataset for tropical forests. Using a unified multi-resolution detection benchmark, it demonstrates that high-resolution inputs, DINO-Swin detectors, and cross-dataset training significantly improve in-distribution and zero-shot generalization for tropical tree crown detection. MoRA treats human mobility graphs as the "structural backbone" for multimodal fusion. Using CLIP-style asymmetric contrastive learning, it aligns POIs, satellite imagery, and demographics with a billion-edge mobility graph. It outperforms SOTA by an average of 12.9% across 9 socioeconomic downstream tasks using 128-dimensional representations and provides the first empirical evidence of scaling laws in geospatial representation learning. OF-Diff utilizes category labels to directly extract "shape mask priors" of remote sensing objects to constrain diffusion generation. An "online distillation" framework is employed to distill mixed features containing real image information into a shape-dependent decoder. This enables the model to generate high-fidelity, layout-consistent remote sensing images without requiring real image references during inference. Finally, DDPO reinforcement fine-tuning is used to further align with the real distribution, resulting in a 4–8% mAP improvement for categories such as airplanes, ships, and vehicles in downstream detection tasks. This paper proposes ABOM, a task-free adaptive meta black-box optimizer that parameterizes evolutionary operators (selection, crossover, and mutation) as differentiable attention modules. By utilizing self-generated data to update parameters online during the optimization process, it achieves competitive zero-shot performance on synthetic benchmarks and UAV path planning. Starting from a single satellite image and predefined ground camera trajectories, SatDreamer360 utilizes triplane scene representations, ray-guided pixel attention, and panoramic epipolar-constrained temporal attention to generate geometrically aligned and cross-frame consistent 360° ground panorama sequences within a diffusion model, outperforming Sat2Density, ControlS2S, and EscherNet on the newly constructed VIGOR++ benchmark. The authors propose TAMMs—the first unified framework to jointly execute Temporal Change Description (TCD) and Future Satellite Image Forecasting (FSIF) in a single MLLM-Diffusion architecture. It awakens the temporal reasoning capabilities of a frozen MLLM through a Temporal-Aware Module (TAM) and translates change understanding into generative control signals via a Semantic Fusion Control Injection (SFCI) mechanism. Browse all 11 Remote Sensing papers → 🔍 Anomaly Detection (10) ¶ The authors discovered that frozen foundation visual encoders "secretly" possess the ability to distinguish anomalies—the area of an anomalous region in an image is positively correlated with the distance of its features to the natural image manifold. By training a lightweight non-linear projection operator (FOUNDAD) atop the encoder to pull anomalous features back to the normal manifold and scoring based on the difference before and after projection, SOTA performance is achieved in few-shot, category-agnostic industrial anomaly detection. MRAD replaces the parametric fitting of \(p(y|x)\) in mainstream ZSAD with "similarity retrieval from a feature-label memory bank." The training-free version suppresses WinCLIP, and combined with two linear fine-tuning layers and dynamic prompts injected with regional priors, it achieves SOTA across 16 industrial and medical datasets. The authors propose W1-ACAS: a post-hoc, tuning-free adaptive conformal anomaly detection framework. It maps prediction errors from pre-trained Time Series Foundation Models (TSFMs) into anomaly scores directly interpretable as false positive rates (p-values) and learns weights online by minimizing the Wasserstein distance to maintain stable false positive control under non-stationary data. This work repositions the LLM from a "data processor" to an "algorithmic strategist"—it analyzes the algorithmic description of a detector without touching real data, reasons about its logical blind spots, and generates a reusable Python synthesis code. This code creates "hard anomalies" specifically designed to deceive that detector, upgrading the original one-class problem into a more separable two-class problem. It consistently enhances five mainstream detectors across 36 tabular anomaly detection benchmarks. ReTabAD is the first "context-aware" tabular anomaly detection benchmark. It restores discarded textual semantics (feature descriptions, domain knowledge, original categorical values) into 20 curated datasets, provides implementations for 20 classic/deep/LLM-based algorithms, and proposes a training-free zero-shot LLM framework. Experiments demonstrate that semantic context improves detection AUROC by an average of 7.6 percentage points, allowing zero-shot LLMs to approach Prev. SOTAs. UniOD trains one universal outlier detection model using a batch of historical labeled datasets. It first unifies tabular datasets of any dimension or semantics into "multi-scale similarity graphs + SVD features," then transforms outlier detection into node binary classification using a GIN+GT dual-path graph network. Once trained, the model performs training-free and parameter-tuning-free inference for any unseen new dataset, achieving average AUROC/AUPRC scores that outperform 17 baselines across 30 benchmarks with lower latency. This paper theoretically maps the learning process of Transformer encoders on multivariate time series to the classical STAR statistical model. It proposes ALoRa-T, which applies low-rank regularization to self-attention, using the "rank" of the attention matrix as an anomaly signal for detection and tracing anomalies back to specific variables for localization using interpretable contribution weights. ConFVG utilizes the second smallest eigenvector of the graph Laplacian (Fiedler vector) to guide the masking strategy of a Graph Autoencoder (GAE) for structural-aware representation learning under label scarcity. It then employs subgraph attention fusion and a Mean Teacher framework to continuously adapt to evolving fraud patterns in unlabeled online streams, achieving real-time healthcare fraud detection. This paper proposes VPC, which utilizes a set of fixed equiangular prototypes to map ID classes and OOD samples into two distinct predefined subspaces. By using the difference in \(L_2\) activation intensity between these two subspaces as an OOD score, it consistently reduces FPR95 in Outlier Exposure (OE) training scenarios on CIFAR and ImageNet-1k. PIRN targets few-shot multimodal industrial anomaly detection for RGB images and 3D surface normals. It reconstructs normal features of each modality using adaptive prototype codebooks and enhances texture and geometric cues through cross-modal normality communication, achieving superior detection and localization performance on MVTec 3D-AD, Eyecandies, and Real-IAD D3. 🧑 Human Understanding (45) ¶ Addressing the lack of public data for "active interactive digital humans," this work constructs SpeakerVid-5M—the first large-scale, high-quality dataset for audio-visual dyadic interactive digital human generation (8,743 hours, 5.2 million single-person clips, 770,000 dialogue pairs). It also introduces an auto-regressive video dialogue baseline and the VidChatBench evaluation benchmark. This paper proposes BAH, the first multimodal dataset for Ambivalence/Hesitancy (A/H) recognition in videos. It contains 1,118 videos (8.26 hours) from 224 participants across 9 Canadian provinces, annotated by behavioral science experts, and provides baseline experimental results at both frame and video levels. Ours proposes the new task of "Free-Form Hand-Object Interaction (HOI) Generation" along with WildO2, an in-the-wild 3D dataset automatically reconstructed from web videos. A three-stage framework, TOUCH (Contact Map Prediction → Multi-level Conditional Diffusion → Physical Constraint Refinement), is designed to move beyond "stable grasping" priors, enabling the generation of diverse and physically plausible hand poses—such as pushing, poking, and rotating—based on fine-grained textual instructions. A VLM translates high-level instructions into a part-level bipartite graph of "Relative Motion Dynamics (RMD)," automatically constructing target states and reward functions for reinforcement learning. This enables physically simulated characters to complete long-horizon interactions with static, dynamic, and articulated objects without motion capture data or manual reward tuning. To address high evaluation costs in Descriptive Multimodal Emotion Recognition (DMER), EmoPrefer is proposed as the first emotion preference dataset and benchmark. It systematically explores whether MLLMs can replace human annotators for emotion preference judgment. The best approach (Qwen2.5-Omni) achieves a 67.21% two-class WAF, leaving room for further improvement. Text2Interact addresses text-driven two-person 3D interaction generation. It first utilizes InterCompose to synthesize high-quality interaction data from LLMs and single-person motion priors, then employs InterActor with word-level text conditioning, dual-person motion interaction attention, and adaptive interaction loss to enhance motion realism, text alignment, and cross-distribution generalization. Ours proposes TEMU-VTOFF—a Dual-DiT architecture for the virtual try-off (VTOFF) task. It employs a collaborative division between a feature extractor and a garment generator, utilizing Multi-modal Hybrid Attention (MHA) to fuse image, text, and mask information to resolve visual ambiguity. Additionally, a DINOv2-driven garment aligner is designed to preserve high-frequency details. Ours achieves SOTA performance in multi-category scenarios on both VITON-HD and Dress Code datasets. EasyTune transforms the fine-tuning paradigm for diffusion models from "calculating reward gradients after the full denoising trajectory" to independently optimizing at each denoising step . This breaks the recursive gradient dependency between steps, reducing VRAM usage from \(O(T)\) to \(O(1)\) and enabling denser optimization. Combined with a Self-refined Preference Learning (SPL) module that converts retrieval models into motion reward models without human annotation, it outperforms DRaFT-50 by 7.7% in alignment (MM-Dist) on HumanML3D, while using only 31.16% of its additional VRAM and speeding up training by 7.3×. Motion-R1 combines a "Decomposed Chain-of-Thought (CoT) Data Engine" with "RL Binding": the former uses an LLM to decompose high-level instructions into temporal/causal sub-action chains for cold-start SFT; the latter uses GRPO to directly incorporate "motion similarity + semantic similarity + format" as rewards, eliminating the need for expensive human preference annotations while generating semantically aligned and realistic 3D human motions. Ours proposes the Snippet paradigm: organizing gait silhouette sequences into multiple "snippets," where each snippet consists of frames randomly sampled within a continuous interval. This approach balances short-range temporal context with long-range temporal dependencies. Using a 2D convolutional backbone, it achieves 77.5% Rank-1 on Gait3D, surpassing all 3D convolutional methods. Browse all 45 Human Understanding papers → 📹 Video Understanding (48) ¶ This paper proposes "spatial supersensing," a paradigm shift from passive task-driven sensing to active world modeling. It first proves via the VSI-SUPER benchmark that brute-force context expansion (including Gemini-2.5 and the self-trained Cambrian-S) fails completely on spatial recall and counting tasks in arbitrarily long videos. It then introduces a self-supervised "Latent Frame Prediction" head that uses prediction error ("surprise") as a control signal to drive memory management and event segmentation, significantly outperforming strong commercial baselines on long-video spatial tasks. Video-STAR reformulates Open-Vocabulary Action Recognition (OVAR) as a sequential decision process of "selecting tools first, then decomposing sub-actions": during inference, a Multimodal Large Language Model (MLLM) calls domain-specific tools (e.g., pose estimation, human detection, online retrieval) to supplement visual evidence and decomposes holistic actions into discriminative sub-action primitives for scoring and matching; this is coupled with a hierarchical reward system (rewarding accuracy, tool efficiency, and sub-action relevance) to train the model via GRPO, shifting it from "relying on text priors" to "vision-grounded reasoning," significantly advancing the SOTA across five benchmarks: HMDB-51, UCF-101, K-400/600, and SSv2. FOCUS reformulates the task of "selecting the most query-relevant video frames under a strict token budget" as a Combinatorial Pure Exploration (CPE) problem in multi-armed bandits. By treating short video segments as arms and adaptively allocating the scoring budget using empirical means and Bernstein confidence radii, it significantly improves long video QA accuracy while viewing less than 2% of total frames. To address the inefficiency caused by excessive visual tokens in Video LLMs, PPLLaVA utilizes a "prompt-video relevance map" calculated by CLIP as a dynamic 3D convolution kernel to compress tokens. This approach reduces visual sequences by up to 1/18 while preserving key information relevant to user instructions, achieving both speedup and performance gains across seven video understanding benchmarks. CaReBench utilizes 1,000 manually annotated videos—each with captions exceeding 200 words and explicitly split into spatial and temporal versions —to establish a benchmark capable of simultaneously evaluating fine-grained video captioning and retrieval. It introduces two new metrics, ReBias and CapST, to quantify the spatiotemporal bias of VLMs, and provides a two-stage SFT baseline, CARE, which unifies captioning and retrieval into a single MLLM. FlashVID is proposed as a training-free inference acceleration framework for Video Large Language Models (VLLMs). By jointly modeling spatial and temporal redundancy through Tree-based Spatiotemporal Token Merging (TSTM), it maintains 99.1% of LLaVA-OneVision's performance while retaining only 10% of visual tokens. Furthermore, it enables a 10x increase in input frame capacity for Qwen2.5-VL. Ours proposes LaGoVAD, a language-guided open-world video anomaly detection paradigm. By modeling the anomaly definition as a random variable input in natural language form, and combining it with dynamic video synthesis and contrastive learning regularization strategies, it achieves zero-shot SOTA performance across seven datasets. Ours proposes VideoNSA, which introduces Native Sparse Attention (NSA) into video-language models. Through a hybrid sparse attention mechanism with dynamic gating across three branches—compression, selection, and sliding window—it achieves 128K token video understanding using only 3.6% of the attention budget. It comprehensively outperforms token compression and training-free sparse attention baselines in long video understanding, temporal reasoning, and spatial understanding tasks. This paper recasts the inherently ambiguous task of "predicting future motion from a single image" as a conditional generation task on a dense grid of point trajectories . By using a trajectory VAE to compress full-image point trajectories into a latent space and sampling diverse possible futures via rectified flow matching, the proposed method is both more accurate than regressive trajectory predictors and more physically plausible than large video models that generate RGB pixels before tracking. TAPTRv3 targets any point tracking in long videos. Built upon the DETR-like point query framework of TAPTRv2, it introduces spatial context cross-attention, visibility-aware long-term attention, and scene cut-triggered global matching. These enhancements effectively reduce feature drift under long sequences, occlusions, and camera switches, setting new state-of-the-art results on multiple TAP benchmarks. Browse all 48 Video Understanding papers → 🚗 Autonomous Driving (50) ¶ VADv2 reformulates end-to-end driving planning from "regressing a single trajectory" to "learning a probability distribution over the action space." It first discretizes the continuous action space into a 4096-word planning vocabulary via furthest trajectory sampling, then uses a NeRF-inspired probabilistic field and cascaded Transformer to predict probabilities for each candidate action, and finally samples a trajectory for vehicle control. Using only camera inputs, it achieved a Driving Score of 85.1 on CARLA Town05 and led several benchmarks including Bench2Drive, NAVSIM, and 3DGS. Apple utilized Vision Pro to collect 829 hours of egocentric video paired with 3D hand joint tracking (EgoDex), covering 194 tabletop manipulation tasks. They systematically evaluated imitation learning strategies (BC/DDPM/FM + Transformer) on this dataset, providing the largest data foundation to date for the scaling and training of dexterous manipulation. ReCogDrive replaces the "trajectory as text generation" paradigm with a "Cognitive VLM + Diffusion Planner" framework. It first injects human driving cognition into a VLM through a hierarchical data pipeline, then treats VLM hidden states as conditions for a diffusion planner to output continuous trajectories. Finally, a DiffGRPO reinforcement learning stage, tailored for diffusion policies, optimizes safety and comfort within the NAVSIM simulator. This achieves SOTA performance on both NAVSIM (PDMS 90.8) and Bench2Drive, while being 3.5× faster than pure text output. DriveVLA-W0 adds a "predict future images" world model task to autonomous driving VLA, using dense visual self-supervision signals to fill the "supervision deficit" left by sparse action supervision. This effectively "amplifies" the data scaling law across 70M frames, allowing the model to consistently improve rather than reaching early saturation. RAP utilizes lightweight 3D rasterization to generate controllable counterfactual views and recovery scenes from real driving logs. It then stabilizes the transfer of these synthetic samples to real-image planners through feature-space Raster-to-Real alignment, significantly enhancing end-to-end planning robustness on closed-loop/long-tail benchmarks such as NAVSIM, WOD-E2E, and Bench2Drive. Ours proposes Astra, a general interactive world model that enables action-conditioned long-range video prediction on pre-trained video diffusion models through an autoregressive denoising framework. It introduces ACT-Adapter (action injection), noise-enhanced historical memory (alleviating visual inertia), and Mixture of Action Experts (unifying heterogeneous action modalities), achieving SOTA fidelity and action-following capabilities across autonomous driving, robotic manipulation, and scene exploration. The paper treats autonomous driving trajectory planning as "language modeling"—first pre-training a motion token predictor autoregressively on expert data to learn "driving like a human," and then applying reinforcement learning fine-tuning with rule-based rewards and an improved GRPO (VD-GRPO). This explicitly aligns the model with driving principles such as safety, comfort, and compliance, achieving SOTA on nuPlan, especially under interactive reactive settings. ReflectDrive discretizes 2D driving space into an action codebook, uses a pre-trained Diffusion Language Model (DLM) for VLA trajectory planning, and layers a gradient-free "reflection mechanism"—performing local searches on unsafe tokens to find safety anchors, followed by diffusion inpainting to regenerate surrounding trajectories. It achieves a PDMS of 91.1 (approaching the human score of 94.8) on the NAVSIM closed-loop benchmark. SMART-R1 introduces R1-style Reinforcement Fine-Tuning (RFT) to multi-agent traffic simulation for the first time, proposing the Metric-oriented Policy Optimization (MPO) algorithm and an "SFT-RFT-SFT" iterative training strategy. It achieved first place on the WOSAC 2025 leaderboard with a Realism Meta score of 0.7858. BridgeDrive proposes replacing truncated diffusion with a diffusion bridge to achieve anchor-guided trajectory planning in autonomous driving. This ensures theoretical symmetry between forward and backward processes, achieving success rates of 74.99% (PDM-Lite) and 89.25% (LEAD) in Bench2Drive closed-loop evaluations, surpassing previous SOTA by 7.72% and 2.45%, respectively. Browse all 50 Autonomous Driving papers → 🤖 Robotics & Embodied AI (162) ¶ X-VLA encodes hardware and collection variances from each robot data source into a set of learnable soft prompts. Combined with a concise Transformer + flow matching action generation framework, it achieves robust cross-embodiment adaptation after pre-training on large-scale heterogeneous robot data. This work utilizes the pre-trained video generation foundation model Cosmos-Predict2-2B as a base, without modifying any network architecture and using only one-stage fine-tuning . It "encodes" robot actions, future states, and state values as "latent video frames" for joint denoising and generation. This allows the model to simultaneously serve as a policy, world model, and value function. It achieves SOTA on LIBERO (98.5%), RoboCasa (67.1%), and real-world dual-arm ALOHA tasks, with an additional 12.5-point improvement using best-of-N planning. HybridVLA enables a single LLM backbone to simultaneously perform diffusion denoising and autoregressive action prediction within a unified token sequence. By adaptively fusing both paradigms through a confidence-based collaborative ensemble, it achieves performance gains of 17% in simulation and 19% on real robots over SOTA models. Inspired by the dual memory system in cognitive science, this work proposes the MemoryVLA framework. It introduces a Perceptual-Cognitive Memory Bank (PCMB) into the VLA model to capture long-term dependencies through memory retrieval, gated fusion, and consolidation mechanisms, significantly outperforming CogACT and π₀ across 150+ tasks in SimplerEnv, LIBERO, and real-world environments. SimpleVLA-RL adapts outcome-driven online RL from the LLM domain into a closed-loop robotic training framework suitable for Vision-Language-Action (VLA) models. By utilizing interactive trajectory sampling, binary success rewards, and exploration-enhanced GRPO, it significantly improves data efficiency, generalization, and success rates for long-horizon manipulation across LIBERO, RoboTwin, and real-world robotic tasks. Ours transforms pre-trained passive video diffusion models into a controllable, multi-view, and long-term consistent robotic world model. This allows general-purpose VLA policies to perform closed-loop rollouts in "imaginary space," enabling policy evaluation without real robots and improving success rates by 44.7% through fine-tuning on synthesized success trajectories. GE integrates a "multi-view video world model (GE-Base)" and a "lightweight parallel action decoder (GE-Act)" into a unified video generation framework. The action branch directly reads multi-scale, full-resolution latent representations from the video DiT via block-wise alignment. Combined with slow-fast asynchronous inference, it generates 54-step action trajectories within 200ms on a single RTX 4090 and enables transfer to new robotic embodiments using only 1 hour of teleoperation data. OneTwoVLA unifies fast action execution and slow language reasoning within a single VLA. The model adaptively triggers reasoning using [BOR] during critical moments and outputs actions directly via [BOA] otherwise, significantly outperforming non-reasoning VLAs and dual-system approaches in long-horizon manipulation, error recovery, human-robot interaction, and open-vocabulary visual grounding. Spatial Forcing utilizes geometric latents from a pre-trained 3D foundation model (VGGT) to supervise the intermediate visual tokens of a VLA. This enables robotic policies to acquire stronger spatial understanding without requiring additional depth maps or point clouds during inference, leading to improved success rates, convergence speed, and data efficiency on LIBERO, RoboTwin, and real-robot tasks. Inspired by the human brain's left-hemisphere semantic understanding and right-hemisphere spatial cognition, this paper proposes JanusVLN—the first dual implicit neural memory framework designed for VLN. It models spatial-geometric and visual-semantic memories as fixed-size KV Caches, achieving efficient spatial reasoning using only RGB video and reaching SOTA performance on the VLN-CE benchmark. Browse all 162 Robotics & Embodied AI papers → 🎮 Reinforcement Learning (400) ¶ ReTool employs a training framework of "cold-start SFT + tool-augmented RL" to enable LLMs to autonomously learn "when and how to call a code interpreter" during long-chain reasoning. By using only outcome-based rewards, a 32B model achieved 67.0% on AIME2024, significantly surpassing the text-only RL baseline (40.0%) while using only one-third of the training steps. This paper introduces Rubrics as Rewards (RaR) , treating itemized rubric checklists as reward functions for on-policy reinforcement learning. This extends RL with Verifiable Rewards (RLVR)—previously limited to "verifiable" tasks like math or code—to real-world reasoning domains like medicine and science where no single standard answer exists. RaR achieves up to a 31% improvement over popular LLM-as-judge Likert baselines on HealthBench and a 7% improvement on GPQA-Diamond. Addressing the debate over whether RLVR truly improves reasoning capabilities or merely enhances sampling efficiency, this paper proposes a new metric, CoT-Pass@K (requiring both correct answers and correct reasoning). Using a theoretical framework for GRPO, it proves that as long as the base model possesses a "logic prior" where correct CoT more likely leads to correct answers, binary rewards based solely on answer correctness will implicitly drive up the probability of generating correct reasoning, thereby authentically extending the reasoning boundary of base models. R-Zero initializes two roles, a "Challenger" and a "Solver," from a single base model. The Challenger is rewarded for generating difficult problems at the edge of the Solver's capability (accuracy \(\approx 50\%\) ), while the Solver is rewarded for solving them. The two are trained alternately using GRPO in a co-evolutionary process. Without any human-authored questions or labels, this method improves the mathematical reasoning average of Qwen3-4B-Base by \(+6.49\) and general reasoning by \(+7.54\) . This paper discovers that the KL divergence between the base model and the fine-tuned model on the new task distribution can predict catastrophic forgetting. It explains why on-policy RL, compared to SFT, tends to find high-reward solutions closer to the original policy, thereby forgetting less when learning new tasks. By integrating a trio of "Scaffolded Data Synthesis + Compiler Feedback-driven Self-Correction + Model Averaging," an open-source Lean theorem prover achieves a new SOTA. The 8B model outperforms the 671B DeepSeek-Prover-V2, and the 32B model reaches a 90.4% pass@32 on MiniF2F with 20x fewer parameters and a significantly lower computational budget. This work redefines reward modeling as a reasoning task and proposes the RM-R1 series of Reasoning Reward Models (ReasRM). Through reasoning distillation, RL training, and the Chain-of-Rubrics (CoR) mechanism, it outperforms 70B and GPT-4o models on three major reward model benchmarks by an average of 4.9%. This work mathematically proves from an RL policy gradient perspective that the SFT gradient implicitly encodes a pathological reward structure of inverse probability weighting ( \(1/\pi_\theta\) ). This causes excessively large gradients for low-probability tokens, which limits generalization. The authors propose DFT (Dynamic Fine-Tuning), which eliminates this weighting via a one-line code modification (multiplying CE loss by the token probability: \(-p\log p\) ). DFT significantly outperforms SFT in mathematical reasoning, code generation, and multimodal tasks, and even surpasses GRPO/PPO in offline RL settings. This paper proposes RLCR (Reinforcement Learning with Calibration Rewards), which overlays a Brier score term on top of standard "binary correctness rewards." This allows reasoning models to output a calibrated confidence while generating answers. Without significant loss in accuracy, it reduces Expected Calibration Error (ECE) from 0.37 to 0.03 on HotpotQA and reverses the degradation trend of standard RL—where models typically become "more confident and more chaotic" as training progresses—on out-of-distribution (OOD) tasks. This work replaces the "arithmetic mean" used in GRPO for optimizing token-level rewards with a "geometric mean." By leveraging the inherent robustness of the geometric mean to outliers, the method suppresses extreme importance sampling ratios, thereby stabilizing policy updates without sacrificing exploration capability. Mathematically, it achieves a Pass@1 improvement of up to 4.1% over GRPO in reasoning tasks. Browse all 400 Reinforcement Learning papers → 🎁 Recommender Systems (24) ¶ Addressing the pain points where explicit Chain-of-Thought (CoT) in LLM recommendations is both difficult to obtain as supervisory data and slow during inference, this paper proposes LatentR3. By adding a LatentRATT attention layer at the top of the LLM to compress reasoning into a continuous latent space (requiring only 1 latent token) and employing a modified GRPO (utilizing PPL continuous rewards + batch-level advantage), the model learns latent reasoning end-to-end without any CoT supervision. This approach yields relative improvements of 17.0% and 8.4% when applied to BIGRec and D3, respectively. The authors construct Search Arena—the first large-scale search-augmented LLM human preference dataset (24,069 conversations + 12,652 preference votes across 71 languages). The study discovers that user preferences are heavily influenced by citation count (even when citations do not support statements), community-driven platforms are preferred over Wikipedia, and search augmentation does not degrade general chat performance, whereas general LLMs significantly deteriorate in search scenarios. VISTA decouples the target attention of candidates over ultra-long user histories into a two-stage process: first, compressing million-length histories into hundreds of summary tokens to be cached; second, performing lightweight attention only on these cached tokens downstream. This keeps training and inference costs constant and has been deployed on Meta’s recommendation platform serving billions of users. This paper proposes the ConvRec-R1 two-stage framework to train LLM-based conversational recommender systems: first, a Remap–Reflect–Adjust distillation pipeline is used to generate high-quality demonstrations from a black-box teacher that are "grounded within the target catalog" for SFT warmup; then, Rank-GRPO (recrafting GRPO to treat "each rank" in the recommendation list as an action unit) is applied for RL alignment. This allows small models (0.5B–3B) to converge faster on REDDIT-V2 in terms of Recall/NDCG and match or even exceed GPT-4o. This work proposes ProPerSim, a simulation framework that constructs 32 user personas based on the Big Five personality traits within the Smallville household environment. AI assistants perform proactive recommendation decisions every 2.5 minutes. Through DPO preference learning over a 14-day simulation, user satisfaction improved from 2.2/4 to 3.3/4, validating for the first time the feasibility of unifying proactivity and personalization. PESO transforms continual learning for LLM-based generative recommendations from "stacking multiple frozen adapters" into "a single evolving LoRA + a proximal regularization term." By gently anchoring each update to the previous stage's state, the model automatically balances retaining long-term preferences and absorbing new ones, consistently outperforming cumulative LoRA and simple evolving LoRA across three real-world datasets. Addressing the issue where shoving item-IDs into an LLM causes collaborative signals and linguistic semantics to conflict, this paper proposes IDIOMoE: splitting the FFN of each pre-trained LLM block into a text expert and an item expert . Using static token-type gating to route tokens based on their type (item-id tokens go to the item expert, others to the text expert), the model decouples "collaborative filtering" and "semantic understanding" into different subnetworks. This achieves state-of-the-art recommendation performance on both public and industrial-scale datasets while maintaining the original LLM's linguistic capabilities. RPM upgrades black-box LLM personalization from "aligning final responses" to "aligning underlying reasoning processes." It automatically extracts a structured user model of "features → factors → statistics" from raw user history, constructs personalized reasoning paths for each history entry, and feeds these reasoning examples to the model via feature-based retrieval. This enables the LLM to reason following the user's private logic, consistently outperforming existing response-level personalization methods across four task categories with enhanced interpretability. This paper theoretically explains the root cause of "one-epoch overfitting" in CTR/CVR models—where performance collapses after the first epoch—using Rademacher complexity. It identifies that the unconstrained growth of embedding norms expands the generalization bound. Consequently, the authors propose AdamAR, an adaptive regularization method that allocates norm budgets based on feature frequency: applying light regularization for high-frequency features and heavy regularization for low-frequency ones. This approach eliminates multi-epoch overfitting while improving single-epoch performance and has been deployed in Alibaba's search advertising system. It is theoretically proven that any Multi-Generator-Evaluator ranking system can be approximated with smaller error by a larger generator-only model that satisfies the scaling law. Accordingly, GoalRank is proposed—training a large generator-only ranking model by constructing group-relative reference policies with a reward model. It significantly outperforms SOTA in online A/B tests. Browse all 24 Recommender Systems papers → 🔄 Self-Supervised Learning (81) ¶ PonderLM is proposed, which introduces a "pondering" mechanism during the pre-training phase. It transforms predicted probability distributions into continuous embeddings via weighted sums and performs repeated forward passes. Without requiring annotated data or reinforcement learning, a 2.8B model outperforms a 6.9B model across 9 downstream tasks. HiMAE integrates masked autoencoders into a U-Net-style hierarchical 1D CNN, allowing intermediate layers to naturally correspond to embeddings at different temporal resolutions. This transforms "resolution" from a hyperparameter into a probe-based diagnostic tool, while the model is small enough to perform sub-millisecond inference on smartwatch CPUs. This paper theoretically demonstrates that the InfoNCE loss induces representations to converge toward a Gaussian distribution through two complementary mechanisms: an empirical idealization path (alignment + spherical uniformity → Gaussian) and a regularized path (vanishing regularization → isotropic Gaussian). These findings are validated using synthetic data and CIFAR-10. Inspired by the fruit fly olfactory circuit, this work proposes the Fly-CL framework. It achieves SOTA performance in pre-trained model-based continual learning while significantly reducing training time through a three-stage progressive decorrelation process involving sparse random projection, top-k operation, and streaming ridge classification. Addressing the combinatorial generalization deficit in Goal-Conditioned Behavioral Cloning (GCBC), which fails to "stitch" novel state-goal pairs, this paper proposes BYOL- \(\gamma\) : a self-predictive representation learning objective that samples future states using a geometric distribution to approximate the successor measure. As an auxiliary loss for BC, it requires neither TD learning nor negative samples, achieving average success rates across OGBench stitching tasks that surpass all baseline methods. CSRv2 utilizes "progressive k-annealing + sparse supervised contrastive learning + full-backbone fine-tuning" to advance Contrastive Sparse Representation (CSR) into the ultra-sparse range of \(k\le 4\) . This approach reduces dead neurons from 80% to 20% and achieves a 14% accuracy gain at \(k=2\) . It enables embeddings with only 2 active dimensions to match the performance of CSR at \(k=8\) or MRL at 32 dimensions, providing up to a 300× improvement in computational and memory efficiency compared to dense embeddings. This paper proposes bidirectional Predictive Coding (bPC), which employs a single energy function to accommodate both "top-down generative" and "bottom-up discriminative" inference. This allows the same biologically plausible local circuit to perform accurate classification like discPC and generation/reconstruction like genPC, outperforming existing unidirectional or hybrid PC models in brain-inspired tasks such as cross-modal association and occlusion completion. NEO discovers that input distribution shifts cause a global translation in penultimate embeddings shared across samples and classes. By re-centering test features to the origin using a single global centroid vector, it outperforms seven mainstream TTA methods with zero optimization, zero hyperparameters, and near-zero additional overhead . The paper proposes Regularized Latent Dynamics Prediction (RLDP), which maintains feature diversity by adding simple orthogonal regularization to a self-supervised latent next-state prediction objective. RLDP matches or exceeds complex SOTA representation learning methods in zero-shot RL, demonstrating significant advantages particularly in low-coverage scenarios. SplitLoRA transforms the long-standing challenge of "determining the minor subspace dimension" in continual learning from a heuristic threshold into a solvable optimization problem. It derives a theoretical upper bound for "Stability Loss + Plasticity Loss" as a function of the subspace dimension \(k\) , solves for the module-specific optimal \(k^*\) for each LoRA module, and fixes the LoRA down-projection matrix \(A\) to this subspace while training only \(B\) . It outperforms existing methods by 2%–5% on ImageNet-R, CIFAR-100, and DomainNet. Browse all 81 Self-Supervised Learning papers → 📐 Optimization & Theory (222) ¶ Polar Express transforms the polar decomposition approximation in Muon from heuristic Newton-Schulz coefficient searches into solving for the worst-case error-optimal combination of odd polynomials each round. While remaining purely matrix-multiplication-based and bfloat16-friendly, it allows Muon updates in GPT-2 training to converge faster and more stably toward valid polar factors. Based on a unified and fair hyperparameter tuning and end-to-end evaluation protocol, this study systematically compares 11 deep learning optimizers. It reveals that the 1.4–2× speedups claimed by new optimizers largely originate from "weak baselines." True speedups do not exceed 1.4× and decay to 1.1× as model scale increases; furthermore, it confirms that matrix-based optimizers (Muon/Soap/Kron) indeed outperform scalar-based ones. This paper proposes MIRAS, a framework that unifies sequence modules like Transformers, linear RNNs, TTT, and Titans as "associative memory via online optimization at test-time." By extending the design axes of attentional bias and retention, it introduces three attention-free models—MONETA, YAAD, and MEMORA—which outperform various modern recurrent baselines in language modeling, commonsense reasoning, and long-context needle recall. Add a single line of code to any momentum optimizer: only update coordinates where the "update direction" and "current gradient" share the same sign; otherwise, zero out the update for those coordinates and scale up others proportionally. This yields "cautious" versions like C-AdamW / C-Lion, which consistently accelerate LLM pre-training and image classification without modifying original hyperparameters. CALM enables the simultaneous evolution of "prompts for heuristic generation" and the "underlying LLM" itself. Within an LLM-driven evolutionary heuristic design loop, each "prompt-response-performance" triplet is treated as reinforcement learning data. A local 7B INT4 model is fine-tuned online using GRPO, allowing heuristics generated on a single 24GB GPU to outperform SOTA methods relying on the GPT-4o-mini API across multiple combinatorial optimization tasks. This paper reveals the mechanism behind Muon's speed advantage over Adam through the lens of "associative memory": Muon's update rule normalizes gradient singular values, which naturally matches the outer-product superposition structure of associative memory, thereby enabling more balanced learning of low-frequency "tail" knowledge in heavy-tailed data. MuonBP enables each device to perform orthogonalization only on local shards under tensor parallelism, executing global orthogonalization every \(P\) steps. By using two distinct learning rates—"block step" and "full step"—it eliminates the throughput loss caused by inter-device communication in Muon. On an 8B model, it achieves approximately 8% speedup over Muon with even better convergence results. This paper provides the first non-convex convergence guarantee for the practical Muon optimizer (which uses Newton-Schulz approximation instead of exact SVD polar decomposition). It proves that the convergence rate matches the idealized SVD version up to a constant factor that decays doubly exponentially with the number of Newton-Schulz steps \(q\) , and that Muon suffers \(\sqrt{r}\) times less rank-dependent loss than its vector counterpart SGD-M. HeuriGym places LLMs into an agentic closed loop of "problem reading—code writing—execution feedback—iterative refinement," requiring them to write complete heuristics from scratch for 9 real-world combinatorial optimization (CO) problems (EDA, biology, logistics, etc.). It uses a new metric, QYI, to measure solution quality and feasibility—results show that even the strongest GPT-o4-mini-high and Gemini-2.5-Pro score only 0.6 (where experts score 1.0), exposing weaknesses in tool usage, long-range planning, and adaptive reasoning. The authors complete the μP/CompleteP scaling rules ("tune on small models, transfer to large models") across four critical training axes—width, depth, batch size, and training duration. They further demonstrate that under proper parameterization, even fine-grained "per-module" hyperparameters (individual learning rates, weight decays, and Adam parameters for each tensor/layer type) can be directly transferred from a 50M parameter proxy model to a 7.2B parameter model, achieving a ~1.3× training speedup. Browse all 222 Optimization & Theory papers → 📐 Learning Theory (293) ¶ This paper theoretically answers "what pre-training actually leaves for post-training (RL / test-time scaling)"—the answer is not cross-entropy, but a quantity called coverage profile . The authors prove that next-token prediction implicitly optimizes coverage, and that coverage generalizes faster than cross-entropy without being hindered by sequence length, thereby explaining the anomaly of "why models with lower cross-entropy can have worse Best-of-N performance." This paper demonstrates, using an analytically simplified attention model, that attention behavior undergoes a phase transition as the context length \(n\) increases, driven by the scaling factor \(\beta_n = \gamma \log n\) . The critical point occurs precisely at \(\beta_n \asymp \log n\) (specifically \(\gamma_c = \tfrac{1}{1-\rho}\) ), providing the first rigorous theoretical justification for the logarithmic scaling used in methods like YaRN and Qwen. This paper precisely maps the Empirical Risk Minimization (ERM) problem of two-layer neural networks (diagonal and quadratic) under weight decay training to LASSO and low-rank matrix compressed sensing. This mapping allows for the first analytical characterization of a complete excess risk phase diagram (8 phases, including benign/harmful overfitting and interpolation peaks) in a regime where "true feature learning occurs." Furthermore, it establishes a first-principles correspondence between each scaling law phase and the post-training weight spectrum (bulk / spike / heavy-tail), explaining the empirical observation that "heavy-tailed weight spectra \(\leftrightarrow\) better generalization." This paper decomposes the total loss of language modeling into two terms: "Bayes risk, which decreases as context length increases" and "approximation loss, which increases as context length increases." By introducing Intrinsic Entropy , the Bayes risk is strictly linked to context length, explaining the counter-intuitive phenomenon that "longer context is not necessarily better" and deriving an optimal context length determined by the training data volume. This paper proves from two complementary perspectives, "statistical estimation" and "network approximation," that memorization in diffusion models (reproducing training samples rather than generalizing) fundamentally stems from two provable "separations" between the true score function and the empirical score function: the true score does not minimize the denoising score matching loss, and the empirical score requires a network that scales with the number of samples to approximate. Based on this, a pruning mitigation method for DiT is proposed. This paper leverages the Random Energy Model (REM) from statistical physics to provide an asymptotically exact theory of signal propagation in deep Transformers at initialization. It unifies "rank collapse" and "entropy collapse" as a single phase transition controlled by the query/key initialization variance \(\beta\) . Based on this, it derives an algorithm to generate "trainability maps," guiding practitioners on selecting residual strength and initial weights to ensure trainability in deep models. This paper analyzes the commonly used Best-of-N (BoN) inference-time alignment through a "softened" framework, Soft Best-of-N (SBoN). It provides KL divergence upper bounds relative to the reference policy and upper/lower regret bounds relative to the optimal policy. The authors prove that when the proxy reward model quality is poor and overoptimization occurs, a finite inverse temperature \(\beta\) makes the regret bound for SBoN tighter than that for BoN, thereby mitigating reward hacking. This is a position + theoretical paper: the authors use computational complexity (the TC circuit class) to categorize machine learning problems into "efficiently parallelizable" and "inherently serial." They argue that key tasks such as reasoning, decision-making, and physical simulation belong to the latter. They prove for the first time that even if a diffusion model samples for thousands of steps, its computational depth remains constant, rendering it unable to solve inherently serial problems. Therefore, merely stacking parallel compute (wider networks, more GPUs) is destined to plateau; progress must rely on scaling serial computation. This paper discovers that the loss landscape of LLMs presents "basins" as model scale increases—any perturbation of parameters within the basin preserves performance, while moving outside leads to a sudden collapse. Based on this, randomized smoothing is used to prove that performance degradation from arbitrary fine-tuning or jailbreaking is bounded by the basin radius. A GO optimizer is proposed to actively enlarge these basins to mitigate catastrophic forgetting. This paper provides the first non-asymptotic oracle complexity bounds for estimating the normalizing constant \(Z\) via Jarzynski Equality (JE) and Annealed Importance Sampling (AIS). It replaces isoperimetric inequalities with the "action of the curve" to characterize difficulty, demonstrates that geometric interpolation suffers from exponential action on multimodal distributions, and proposes a new algorithm based on reverse diffusion sampling. Browse all 293 Learning Theory papers → 🔗 Causal Inference (64) ¶ NextQuill reformulates LLM personalization as a causal problem—viewing both model predictions and ground-truth user responses as the joint result of "User History/Features \(\times\) Context." By using causal effects (do-calculus), it isolates the core preference-driven components and employs two alignment losses to learn only these parts, achieving deeper personalization than undifferentiated alignment. CausalFM adapts the "tabular foundation model" PFN to causal inference: it uses Structural Causal Models (SCMs) to generate synthetic priors and pre-trains a Transformer on synthetic data. This enables the model to directly provide Bayesian-style CATE estimates for backdoor, front-door, and instrumental variable settings via in-context learning without retraining. This paper systematically investigates the over-reliance of preference models on five surface features (verbosity, structure, jargon, sycophancy, vagueness). By using causal counterfactual pairs, it quantifies that bias originates from distributional imbalances in training data and proposes Counterfactual Data Augmentation (CDA) as a post-training method, reducing the average miscalibration rate between model and human judgment from 39.4% to 32.5%. This paper proposes Knowledgeable-R1, a reinforcement learning framework that enables LLMs to resist interference from misleading retrieval contexts in RAG scenarios while preserving the ability to utilize reliable context. This is achieved through joint sampling of parametric knowledge (PK) and contextual knowledge (CK) trajectories, combined with local/global advantage calculation and adaptive asymmetric advantage transformation. This paper points out that large-scale multimodal data do not follow the generation assumption of a single Directed Acyclic Graph (DAG). It proposes a Latent Partial Causal Model utilizing "undirected edges to connect two sets of latent coupled variables." On both spherical and convex latent spaces, it is proven that representations learned by Multimodal Contrastive Learning (MMCL), such as CLIP, differ from ground-truth latent variables by a linear orthogonal transformation and a permutation transformation, respectively. This provides the first theoretical guarantee for "component-wise decoupling" in MMCL and implements it via a plug-and-play decoupling pipeline (FastICA / PCA+FastICA), achieving improvements in few-shot learning and domain generalization. This paper proposes IGC-Net : the first neural network to estimate temporal Conditional Average Potential Outcomes (CAPO) through pure regression-based iterative G-computation end-to-end. It correctly adjusts for time-varying confounding while bypassing the zero-division instability of Inverse Probability Weighting (IPW) and the high-dimensional full distribution estimation of G-Net. This paper proposes a decompositional evaluation framework based on Structural Causal Models (SCM), splitting LLM counterfactual reasoning into four stages (causal variable identification → causal graph construction → intervention identification → outcome reasoning). It systematically diagnoses ability bottlenecks across 11 multimodal datasets and suggests tool augmentation and advanced elicitation strategies to improve performance. This paper proposes the WO-learner (overlap-weighted orthogonal meta-learner), which focuses estimation on samples that are truly likely to receive the target intervention sequences by applying an "overlap weight" to training samples. Combined with a Neyman-orthogonal weighted population risk function, it maintains stability in low-overlap scenarios where "overlap probability decays exponentially with the prediction horizon." It outperforms existing meta-learners across synthetic, semi-synthetic, and real-world datasets. This work derives a closed-form expression + semiparametrically efficient estimator for the sharp bounds of the value function in personalized off-policy learning under unobserved confounding. It simplifies the originally unstable minimax optimization into a standard minimization problem and proves that minimizing this estimator yields the optimal confounding-robust policy. The E-CIT (Ensemble Conditional Independence Test) framework is proposed. By splitting data into subsets, executing tests independently, and merging results based on a p-value aggregation method for stable distributions , it reduces the computational complexity of any conditional independence test to linear with respect to sample size. Simultaneously, it maintains or even enhances test power in complex scenarios such as heavy-tailed noise and real-world data. Browse all 64 Causal Inference papers → 🔬 Interpretability (195) ¶ The study proposes GEPA (Genetic-Pareto), a prompt optimizer that diagnoses issues and iteratively optimizes prompts through natural language reflection based on a small number of execution trajectories. It outperforms GRPO by an average of 6% (up to 20%) across six tasks while using only 1/35 of the sampling volume. This is an analytical paper: the authors use the Hungarian algorithm to align multiple Sparse Autoencoders (SAEs) that differ only in their initialization random seeds while seeing identical data. They find that learned features only partially overlap (only 30% on Llama 3 8B). Furthermore, larger models/SAEs exhibit lower overlap, and TopK is more unstable than ReLU. This demonstrates that SAEs identify a "practical decomposition of the activation space" rather than a "unique and objective feature list" actually used by the model. The authors perform "model diffing" on GPT-4o before and after fine-tuning using Sparse Autoencoders (SAEs). They find that a set of "misaligned persona" features (notably a "toxic persona" feature #10) is the primary internal cause of the "emergent misalignment" phenomenon—where fine-tuning on narrow-domain erroneous data leads to broad-domain misalignment. Based on this, they achieve misalignment prediction, steering-based suppression, and "re-alignment" using small amounts of benign data. This paper demonstrates that two seemingly independent puzzles in LLMs—attention sinks and compression valleys—are actually two facets of the same mechanism: massive activations in the residual stream . Based on this, it proposes the Mix-Compress-Refine three-phase information flow theory, unifying the explanation of why embedding tasks are strongest in the middle layers while generation tasks require the full depth. This paper reframes "understanding model activations" as an open-ended question-answering task, LatentQA. Given an activation and a natural language question, a fine-tuned decoder LLM answers directly in natural language. This enables both "reading" activations (monitoring) and "writing" to activations (steering) using gradients backpropagated from natural language-described losses. This paper systematically analyzes the phenomenon of VLMs "seeing evidence but answering incorrectly" in VQA. It finds that deep-layer attention often successfully locates the correct visual evidence, but this information is not fully utilized during the generation stage. Accordingly, the authors propose Visual Evidence Augmentation (VEA), a training-free test-time visual evidence highlighting method, which consistently improves accuracy across various models including LLaVA, Qwen, Gemma, and InternVL on multiple evidence-based VQA tasks. Using an Information Bottleneck/Rate-Distortion framework to measure embeddings from 40+ LLMs and classic human taxonomic cognitive data on a unified "compression ↔ semantics" scale, this work reveals that LLMs achieve more "information-theoretically optimal" aggressive compression than humans, but at the cost of fine-grained semantics (typicality structures). The seemingly "inefficient" human conceptual organization instead serves as a source of adaptive flexibility. This paper mathematically proves that decoder-only Transformer language models are almost surely injective (different prompts produce different last-token representations). Based on this, the authors propose the SIPIT algorithm, which can precisely reconstruct input text from hidden states in linear time. Using causal mediation and causal abstraction, this paper reverse-engineers a universal "lookback mechanism" that large language models (LLMs) rely on to track character beliefs (Theory of Mind). The model copies reference information into "pointers" and "addresses" at different tokens, retrieving "payloads" via QK-attention lookbacks to achieve character-object-state binding, belief retrieval, and visibility updates. This paper conducts a "sanity check" on the currently popular Sparse Autoencoders (SAEs). By applying SAEs to both trained Transformers and randomly initialized Transformers, the authors find that commonly used automated interpretability scores (auto-interp AUROC) and reconstruction metrics are almost indistinguishable between the two. This suggests that high interpretability scores alone cannot prove that an SAE has captured computational features actually learned by the model. Browse all 195 Interpretability papers → 📦 Model Compression (240) ¶ LightMem is proposed as a three-stage lightweight memory system inspired by the human Atkinson-Shiffrin memory model. Through cognitive sensory memory pre-compression, topic-aware short-term memory integration, and offline sleep-time updates, it achieves up to a 7.7% accuracy improvement on LongMemEval while reducing token consumption by up to 38x. TurboQuant is a data-oblivious (requires no calibration, ready for online use) vector quantization algorithm: it first applies a random rotation to "transform" the coordinates of any input vector into nearly independent Beta distributions, then applies pre-solved Max-Lloyd optimal scalar quantizers to each coordinate. This keeps MSE distortion within a constant factor ( \(\approx 2.7\) ) of the information-theoretic lower bound across all bitrates and dimensions. To address bias in inner product estimation, it adds a 1-bit QJL layer to process residuals for unbiased estimation, outperforming existing product quantization in KV cache compression and ANN retrieval. The authors discover that attention weights in Diffusion Transformers can be decomposed into a "small number of high-rank + large number of extremely low-rank" components. They propose SLA—applying precise sparse attention to critical blocks, linear attention to marginal blocks, and skipping negligible blocks. Integrated into a single GPU kernel and requiring only a few thousand fine-tuning steps, SLA reduces attention computation by approximately 95% and achieves a 2.2× end-to-end acceleration in video generation with nearly zero quality loss. This paper theoretically proves that "expert merging" introduces irreducible errors due to the loss of the router's independent, input-dependent modulation capability over experts. Consequently, it proposes REAP—a one-shot pruning criterion that simultaneously considers router gating values and expert activation norms. REAP significantly outperforms merging and other pruning methods across various SMoEs ranging from 20B to 1T, particularly in generative tasks and at a 50% compression rate, achieving near-lossless performance on Qwen3-Coder-480B and Kimi-K2 for code generation. This paper systematically deconstructs the promise of "free speedup and preserved accuracy" offered by hardware-native 4-bit floating-point formats (MXFP4/NVFP4). Through theoretical proof of quantization errors, it identifies why existing quantization techniques fail on these formats. The authors propose MR-GPTQ, a tailored algorithm for FP4 characteristics, and the QuTLASS GPU kernel, achieving 2.2x~4x end-to-end acceleration on B200/RTX5090 while recovering MXFP4 accuracy from a 10% drop to near-NVFP4 levels. DiCache proposes a training-free adaptive caching strategy for diffusion models. It allows DiT to use shallow online probes during inference to determine when to reuse cache and how to combine historical caches. It improves speed while maintaining higher fidelity relative to the original model across WAN 2.1, HunyuanVideo, and Flux. This paper invests thousands of GPU hours to systematically reassess "layer selection metrics" and "post-pruning fine-tuning methods" for LLM layer pruning. It reaches two counter-intuitive conclusions: the simplest "reverse-order pruning of the last few layers" outperforms various complex metrics, and "partial layer fine-tuning of only the lm head + last 3 layers" exceeds the standard LoRA. The authors provide a theoretical explanation using Pre-LN gradient flow, ultimately outperforming existing SOTA pruning methods by 2.36%–19.45% on Llama-3.1-8B, Llama-3-8B, and Llama-3-70B. The authors propose Landscape of Thoughts (LoT), the first tool to visualize LLM reasoning trajectories as 2D topographic maps. By using perplexity-based features and t-SNE projections, LoT reveals behavioral patterns in reasoning and can be adapted into a lightweight verifier to improve reasoning accuracy and test-time scaling effects. OBS-Diff resurrects and adapts the classic Optimal Brain Surgeon (OBS) pruning for large-scale text-to-image diffusion models. Through a "Timestep-aware Hessian," it makes pruning criteria more sensitive to early denoising steps. By using "Module Packages," it amortizes expensive layer-wise calibration. Without requiring any training or fine-tuning, it supports unstructured, N:M semi-structured, and structured granularities, significantly outperforming baselines like Wanda and DSnoT at high sparsities (50%–70%). SteerMoE is proposed to detect behavior-associated experts via contrastive paired inputs and steer the behavior of MoE LLMs by activating or deactivating specific experts during inference (+20% safety, +27% faithfulness). The study also reveals a unique safety alignment vulnerability in MoE models (safety drop -100%). Browse all 240 Model Compression papers → 🕸️ Graph Learning (118) ¶ RelGT is proposed as the first Graph Transformer specifically designed for relational databases. By utilizing multi-element tokenization (a 5-tuple of features/type/hop/time/local structure) and a hybrid local-global attention mechanism, it consistently outperforms GNN baselines across 21 tasks in the RelBench benchmark, achieving gains of up to 18%. Ours proposes the RAS framework, which dynamically constructs query-specific knowledge graphs at inference time. Through a three-stage process of iterative retrieval planning, text-to-triple transformation, and graph-augmented answering, RAS achieves structured reasoning. It delivers improvements of up to 7.0% and 8.7% for open-source and closed-source LLMs, respectively, across 7 knowledge-intensive benchmarks. This paper constructs City-Networks, a million-node scale, large-diameter transductive dataset based on the road networks of four real-world cities. It uses "local eccentricity" to label node classification tasks that naturally require long-range information. Building on this, it proposes a Jacobian-based "hop-wise influence" metric to directly quantify the distance of the neighbor information utilized by any GNN/GT. This replaces the traditional indirect argument of "global attention vs. local aggregation" performance gaps with a measurable, comparable, and theoretically grounded solution. This paper proposes CtrlHGen, upgrading abductive reasoning on knowledge graphs (inferring rational logical hypotheses from observed entities) into a "controllable" task. This allows users to specify the semantic direction and structural complexity of hypotheses. Through data augmentation via sub-logic decomposition, it mitigates the "scarcity of long hypothesis samples." By using a smooth semantic reward involving Dice/Overlap alongside a condition-following reward, it addresses "reward oversensitivity." On three KG datasets, CtrlHGen demonstrates better adherence to control signals and superior semantic similarity compared to baselines. This paper proposes defining in/out-degree Laplacians for cellular sheaves on directed graphs to construct the Cooperative Sheaf Neural Network (CSNN). This allows nodes to independently choose information propagation or reception strategies, simultaneously mitigating oversquashing and addressing heterophilic tasks. By migrating the Prior-Fitted Network (PFN) paradigm from tabular data to graphs, the authors pre-train NodePFN on thousands of synthetic graphs generated from controllable priors. This allows for training-free, single-forward-pass general node classification on arbitrary real-world graphs, achieving a 71.27% average accuracy across 23 benchmarks. DHG-Bench is the first comprehensive benchmark for Hypergraph Neural Networks (HNNs). Under a unified experimental protocol, it systematically evaluates 17 SOTA HNN algorithms against 22 datasets (covering node, hyperedge, and hypergraph task granularities) across four dimensions: effectiveness, efficiency, robustness, and fairness . Through extensive controlled experiments, it reveals systemic shortcomings in existing HNNs, such as "performance collapse when switching data/tasks," "inability to handle large graphs," "vulnerability to feature/label noise," and "lower fairness compared to MLPs." HYPER is the first foundation model for link prediction on knowledge hypergraphs. By encoding "positional interactions between relations" into transferable base relations, the model achieves zero-shot generalization to hypergraphs containing entirely new entities, new relations, and arbitrary arities. Addressing the issue where existing graph diffusion/flow models "decouple nodes and edges for independent linear interpolation," leading to non-smooth probability paths and convergence difficulties, this paper models graphs as coupled colored Gaussian systems using Markov Random Fields (MRF). By constructing smooth, closed-form, simulation-free probability paths based on Bures-Wasserstein (BW) displacement between graph distributions, the proposed BWFlow framework achieves superior performance, faster convergence, and highly efficient sampling in planar graph and molecular generation. DAMR models KGQA as a Monte Carlo Tree Search (MCTS) guided by an LLM planner. The LLM selects top- \(k\) relevant relations during the expansion step to prune the search space, while path scoring is delegated to a lightweight Transformer scorer (jointly encoding questions and relation sequences via cross-attention). This scorer is online fine-tuned using pseudo-paths generated during the search process. DAMR outperforms all SOTAs on WebQSP (Hits@1 94.0) and CWQ (78.0), while reducing LLM calls by over 50% and token consumption by over 75%. Browse all 118 Graph Learning papers → 📈 Time Series (121) ¶ This work discovers that text paired with time series exhibits a periodicity similar to the time series itself (Chronological Textual Resonance). It proposes the TaTS framework, which transforms textual representations into auxiliary variables to enhance the forecasting and imputation performance of any existing time series model in a plug-and-play manner. The paper proposes the Relational Transformer (RT) architecture. Through task table prompting, cell tokenization, and Relational Attention mechanisms, the model can be pre-trained on multiple relational databases and transferred zero-shot to unseen datasets and tasks. A 22M parameter model achieves a zero-shot AUROC of 93% compared to fully supervised methods, significantly outperforming a 27B LLM (84%). TimeOmni-1 proposes the first unified time series reasoning model. Through TSR-Suite (the first reasoning-oriented time series dataset suite) and a two-stage training process (SFT for injecting time series priors + RL for refining reasoning), it significantly outperforms GPT-4.1 across multiple time series reasoning tasks. Aurora is the first multimodal time series foundation model : it is pre-trained on a cross-domain corpus of "time series + textual description + endogenous images." It utilizes modality-guided attention to inject domain knowledge from text/images into time series modeling and employs "prototype-guided flow matching" for generative probabilistic forecasting. This allows it to achieve SOTA performance in both deterministic and probabilistic forecasting under zero-shot and few-shot cross-domain scenarios. Addressing the fundamental issue where the MSE loss generates "autocorrelation bias" when label sequences exhibit autocorrelation, DistDF shifts from estimating conditional likelihood to directly aligning the conditional distributions of predicted and label sequences. It employs the "Joint-distribution Wasserstein distance" (with a provable upper bound) as a proxy objective, leveraging the Bures–Wasserstein closed-form solution under Gaussian assumptions. This serves as a plug-and-play regularization term added to MSE, consistently achieving state-of-the-art results across multiple datasets and backbone models. To address the distribution mismatch in time series forecasting caused by "forcing historical statistical patterns onto future distributions," this paper proposes TimeAlign—a plug-and-play dual-branch framework. It utilizes a "future reconstruction" branch (present only during training) to provide a target distribution for alignment. By employing global and local alignment, the prediction branch's representation is pulled toward the true future distribution, reducing MSE/MAE by 3.27%/5.20% relative to the state-of-the-art across 8 benchmarks. Starting from neuroscience first principles, Uni-NTFM designs Heterogeneous Feature Projection (HFPM) to decouple time-frequency encoding, Hierarchical Topological Embedding (TE) to unify heterogeneous electrode configurations, and MoE Transformer to achieve functional modularity and sparse coding. Pretrained on 28,000 hours of EEG data with 1.9B parameters, it achieves SOTA in linear probing and fine-tuning across 9 downstream tasks. The authors propose a minimalist baseline called "context parroting"—which simply identifies the most similar segment in historical trajectories and copies the subsequent evolution as the prediction. On zero-shot forecasting of low-dimensional chaos, turbulence, coupled oscillators, and ECG signals, this method outperforms leading foundation models like Chronos, TimesFM, Time-MoE, Moirai, and DynaMix in both accuracy and long-term attractor reconstruction, while being six orders of magnitude cheaper in inference , thereby exposing that current foundation models have not truly "learned physics." Addressing the flaw in multi-step time-series forecasting where MSE treats each future step as an independent, equal-weighted task, this paper derives a "quadratic learning objective" weighted by the inverse conditional covariance matrix from a maximum likelihood perspective. Using a bilevel optimization framework (QDF), this weighting matrix is treated as a learnable parameter and learned on a hold-out set to maximize generalization. As a plug-and-play loss replacement for MSE, it consistently achieves SOTA results across 8 datasets and various forecasting models. MixLinear employs a dual-pathway linear architecture of "temporal segmentation for local trends + frequency adaptive low-rank filtering for global trends," reducing long-term time series forecasting (LTSF) model parameters to only 0.1K (45–176) while achieving accuracy comparable to or better than mainstream lightweight models on 8 benchmarks. Browse all 121 Time Series papers → 🏥 Medical Imaging (87) ¶ CodeBrain develops an EEG foundation model using a "time-frequency dual-codebook decoupled tokenizer + multi-scale architecture with parallel global structural convolutional SSM and sliding window attention." After pre-training on the largest public EEG corpus, it consistently outperforms existing EEG foundation models across 10 datasets in 8 task categories and provides codebook-level interpretability. This paper identifies for the first time a failure mode called "Attention Dynamic Instability" in Attention-based MIL for Whole-Slide Imaging (WSI). It proposes ASMIL: a unified framework that stabilizes attention using an EMA-updated anchor model distillation, suppresses attention over-concentration with a normalized sigmoid, and mitigates overfitting via token random dropout, achieving up to 6.49% F1 improvement across multiple pathological datasets. The study proposes the Δ-LFM framework: utilizing ArcRank loss to construct patient-specific temporally aligned trajectories in latent space (consistent angle + monotonically increasing magnitude), extending the flow matching time range from \([0, 1]\) to \([0, T]\) actual time intervals for arbitrary time-point prediction. It outperforms 8 baseline methods across three Alzheimer's longitudinal MRI benchmarks and introduces a progression-specific metric, Δ-RMAE. The LaVCa method is proposed to generate natural language descriptions (captions) for each voxel in the human visual cortex using LLMs. Through a four-step pipeline consisting of "encoding model → optimal image selection → MLLM caption generation → LLM keyword refinement + sentence composition," it reveals voxel-level visual selectivity more accurately and diversely than existing methods like BrainSCUBA. OmniCT utilizes a Spatial Consistency Enhancement (SCE) module comprising "slice composition + tri-axial positional encoding + MoE hybrid projection" to unify 2D slices and 3D volumes into a single LVLM token space. It further incorporates Organ-level Semantic Enhancement (OSE) to explicitly inject anatomical region priors into representations. Combined with the MedEval-CT dataset of 1.7 million samples and a hybrid benchmark, OmniCT significantly outperforms existing medical and general LVLMs in both slice-driven and volume-driven CT tasks (7B model averages 81.45 for slices and 66.15 for volumes). U2-BENCH is the first benchmark to systematically evaluate the ultrasound understanding capabilities of Large Vision-Language Models (LVLMs). By sampling 7,241 cases across 15 anatomy sites from 40 authorized datasets and defining 8 clinical tasks across four categories (classification, detection, regression, and text generation), the evaluation of 20 open-source and closed-source models reveals that while models perform acceptably on image-level classification, they generally fail in spatial reasoning and clinical language generation. This paper proposes Multi-Granular Language Learning (MGLL), a plug-and-play contrastive learning framework. By jointly optimizing soft CLIP loss, point-wise loss, and smooth KL divergence, MGLL aligns medical images with multi-label, multi-granular textual descriptions. It consistently outperforms SOTA methods on fundus and X-ray datasets and can be embedded into multimodal large language models (MLLMs) as a vision encoder, improving diagnostic accuracy by up to 34.1%. Estimates the predictive variance of an "infinite randomly initialized ensemble" using a single model and a single forward pass. By reformulating the ensemble variance as a supervised regression problem with kernel similarity labels, the method avoids training actual ensembles or inverting the Gram matrix. It provides uncertainty estimates comparable to or better than Deep Ensembles, validated on OOD detection and sparse reward RL exploration. BrainGFM models fMRI brain networks as graphs and employs "Graph Contrastive Learning + Graph Masked Autoencoding" for large-scale pre-training on 400,000 brain graphs across 27 datasets and 8 brain atlases. By using meta-learning optimized graph prompts for few-shot adaptation and BioClinicalBERT-encoded language prompts for zero-shot transfer, a frozen foundation model can perform direct diagnosis across diverse atlases, brain disorders, and task settings. The Brain-IT framework is proposed, which utilizes a brain-inspired Brain Interaction Transformer (BIT) to cluster functionally similar brain voxels into cross-subject shared Brain Tokens. It predicts localized semantic and structural image features to achieve high-fidelity reconstruction from fMRI, matching the performance of prior methods using 40 hours of data with only 1 hour of data. Browse all 87 Medical Imaging papers → 🩺 Medical LLM (20) ¶ SimpleToM reveals a critical deficiency in LLM Theory of Mind: while frontier models accurately infer others' mental states (Explicit ToM), their performance drops sharply when applying this knowledge to predict or judge behaviors (Applied ToM), exposing a significant gap between "knowing what" and "how to use what is known." The authors collaborated with 100 licensed mental health professionals to construct CounselBench, a dual-component benchmark for open-ended mental health QA . It includes 2,000 expert evaluations with dimension-level scoring and span annotations (CounselBench-Eval), and 120 clinician-authored adversarial prompts designed to induce specific failure modes (CounselBench-Adv). The study reveals that LLMs currently exhibit "high scores alongside persistent safety hazards" in counseling scenarios and demonstrates that LLM-as-Judge is unreliable in this high-risk domain. In collaboration with 100 licensed mental health experts, CounselBench was constructed as a dual-component benchmark—CounselBench-EVAL (2,000 expert evaluations across six dimensions) and CounselBench-Adv (120 adversarial questions with 1,080 response annotations). The study systematically reveals that while LLMs achieve high superficial scores in open-ended mental health QA, they harbor safety hazards such as overgeneralization and unauthorized medical advice, while also proving that LLM-as-Judge is severely unreliable in safety-critical domains. This paper constructs Cancer-Myth—an adversarial dataset verified by hemato-oncologists containing 585 oncology patient questions with false presuppositions. The study finds that leading LLMs, including GPT-5, Gemini-2.5-Pro, and Claude-4-Sonnet, achieve a success rate of no more than 43% in correcting these false presuppositions. Furthermore, mitigation techniques such as defensive prompting trigger significant over-corrections on "no-false-presupposition" questions and degrade performance on other medical benchmarks, highlighting a critical safety gap in medical LLM patient communication. The authors constructed MedAgentGym, the first unified Agent training environment for biomedical data science. It comprises 72,413 task instances covering 12 real-world scenarios across 129 categories, equipped with executable sandboxes and verifiable ground truth. A systematic benchmark evaluation of 29 LLMs revealed a gap between commercial and open-source models. By employing efficient multi-threaded trajectory sampling and offline/online RL, they trained Med-Copilot, achieving +43.02%/+45.28% improvements and reaching performance competitive with GPT-4o. The authors constructed the MedEvidence benchmark—rewriting conclusions from 100 Cochrane Systematic Reviews (SRs) into 284 closed-ended questions paired with their source studies. This allows LLMs to replicate expert conclusions under "same material" controlled conditions. Evaluating 25 LLMs revealed: reasoning models are not necessarily better, marginal gains diminish with model size, and medical fine-tuning often decreases performance. Models generally lack "scientific skepticism" regarding low-quality evidence, failing to match expert conclusions in at least 37% of cases. Doctor-R1 models outpatient inquiry as a partially observable multi-turn decision-making process. By utilizing a "multi-agent interaction environment + two-level reward architecture + experience memory" for experiential agentic reinforcement learning, an 8B doctor agent learns to ask questions strategically and empathetically while maintaining diagnostic accuracy. It outperforms 32B open-source models and closed-source models like GPT-4.1 on HealthBench and MAQuE. This paper proposes the ATPO (Adaptive Tree Policy Optimization) algorithm, which models multi-turn medical dialogues as a Hierarchical Markov Decision Process (H-MDP). It dynamically allocates rollout budgets through an uncertainty-aware adaptive tree expansion mechanism, guiding exploration via a composite uncertainty measure of Bellman error and action-value variance. Using Qwen3-8B, it outperforms GPT-4o on three medical dialogue benchmarks. Addressing the overconfidence issue where LLMs provide diagnoses despite incomplete information in multi-round clinical consultations, KnowGuard proposes an "investigate-before-abstain" paradigm. This approach shifts abstention decisions from model self-assessment to systematic cross-round evidence exploration over a medical knowledge graph. By using a rolling-updated contextual evidence pool to identify "missing evidence," the model decides whether to continue questioning or provide a diagnosis. On a self-constructed open-ended multi-round benchmark, KnowGuard improved average diagnostic accuracy by 3.93% and converged in an average of only 5.74 rounds. The paper proposes mCLM (Modular Chemical Language Model), which represents molecules as sequences of synthesizable building blocks. This allows LLMs to generate molecules that satisfy both pharmacological functions and automated synthesis feasibility, showing significant improvements in pharmacokinetic and toxicological properties across 430 FDA-approved drugs. Browse all 20 Medical LLM papers → 🧬 Computational Biology (155) ¶ La-Proteina utilizes a "partially latent" representation—explicitly modeling \(\alpha\) -carbon coordinates while compressing other atomistic details and sequences into a fixed-dimension per-residue latent variable. This approach transforms the mixed discrete-continuous and variable-dimension challenge of all-atom protein modeling into a pure continuous, fixed-dimension problem. By applying flow matching to jointly generate sequences and all-atom structures, it achieves SOTA performance in all-atom co-designability, diversity, and structural plausibility, scaling effectively to proteins up to 800 residues. SimpleFold treats protein folding as a conditional generation task from "amino acid sequence to all-atom 3D structure." By utilizing only standard Transformer blocks with a flow-matching objective, it completely discards AlphaFold2’s MSA, pair representations, triangle updates, and equivariant modules. Scaled to 3B parameters on 9M distilled structures, it approaches SOTA on standard folding benchmarks and excels particularly in conformational ensemble generation. TRIBE feeds intermediate layer representations from three pre-trained foundation models (text, audio, and video) into a temporal Transformer to predict fMRI responses of 1000 brain parcels end-to-end. By integrating "nonlinear + cross-subject + multimodal" designs, it won the Algonauts 2025 Brain Encoding Competition with a significant lead among 267 teams. DriftLite proposes leveraging the degree of freedom between the drift and potential functions in the Fokker-Planck equation. By solving for an optimal control drift via a lightweight linear system to proactively stabilize particle weights, it addresses weight degradation in Sequential Monte Carlo with minimal cost, significantly outperforming Guidance-SMC baselines in Gaussian mixture, molecular system, and protein-ligand co-folding tasks. HEIST models tissue as a two-layer hierarchical graph consisting of a "spatial cell graph + intracellular gene co-expression networks." Through cross-level directed attention, gene representations are modulated by the spatial microenvironment while cell representations are shaped by internal transcriptional states. This approach overcomes fixed gene vocabularies, enables zero-shot migration to proteomics, and sets new SOTA performance in clinical prediction, cell annotation, and gene imputation tasks. This paper integrates the concept of "well-tempered metadynamics" from molecular dynamics—applying online repulsive biases along collective variables (CVs)—into the state-of-the-art diffusion sampler ASBS to create WT-ASBS. During training, it continuously accumulates biases along low-dimensional CVs to force the discovery of rare conformations; during inference, the bias is reweighted to restore the Boltzmann distribution. This marks the first time a diffusion sampler has characterized reaction surfaces involving bond breaking/formation with wall-clock times significantly lower than metadynamics. VCWorld is proposed as a cell-level white-box simulator that integrates structured biological knowledge graphs with the iterative reasoning capabilities of Large Language Models (LLMs). It simulates signaling cascades triggered by drug perturbations in a data-efficient manner, generating interpretable step-by-step predictions and explicit mechanistic hypotheses, achieving SOTA on drug perturbation benchmarks. The authors propose MF-(TS)NPE: pre-training a neural density estimator using cheap low-fidelity simulations followed by fine-tuning with a small number of expensive high-fidelity simulations, reducing the required high-fidelity simulation budget for Bayesian inference by up to two orders of magnitude. CryoLVM introduces the Joint-Embedding Predictive Architecture (JEPA) and a SCUNet backbone to the domain of 3D cryo-EM density maps. It performs self-supervised pre-training in representation space using 7,302 real experimental maps from EMDB, combined with a novel histogram distribution alignment loss for fine-tuning. It consistently outperforms specialized methods like DeepEMhancer, EMReady, EM-GAN, and IsoNet across three downstream tasks: sharpening, super-resolution, and missing wedge completion. This paper proposes scDFM, a generative framework based on Conditional Flow Matching (CFM). It ensures distribution-level fidelity through MMD regularization and utilizes a PAD-Transformer backbone to process noisy and sparse single-cell data. On combinatorial perturbation prediction tasks, it achieves a 19.6% reduction in MSE compared to the strongest baseline, CellFlow. Browse all 155 Computational Biology papers → ⚛️ Physics & Scientific Computing (69) ¶ PBFM incorporates PDE residual constraints as a secondary objective during training. It replaces manual loss weighting with Conflict-free Gradients (ConFIG) and eliminates the Jensen gap by reconstructing clean samples through unrolling. This allows Flow Matching to simultaneously approach physics consistency and distributional accuracy without increasing inference overhead, pushing the Pareto front of "Physics vs. Distribution" forward across three PDE benchmarks. This paper proposes PDNS (Proximal Diffusion Neural Sampler), which models "sampling from unnormalized target distributions" as a stochastic optimal control problem in path measure space. It utilizes the Proximal Point Method to decompose a one-shot global optimization into a sequence of sub-problems with KL proximity constraints. This allows the sampler to gradually approach the target along a geometric interpolation path between \(\pi\) and a reference distribution, alleviating mode collapse in strongly multi-modal tasks (Molecular Dynamics, Ising/Potts, etc.) and achieving SOTA on multiple continuous and discrete benchmarks. ARROW redesigns both the "next-step prediction model" and the "long-term autoregressive rollout strategy" in global weather forecasting: it unifies 6/12/24-hour scales using a multi-interval prediction model and employs a DQN scheduler to adaptively select the next jump based on current weather states, simultaneously reducing error accumulation and preserving fine-grained atmospheric variations in mid-to-long-term forecasts. This paper proposes Frozen-PINN, which freezes randomly sampled spatial basis functions and utilizes least squares and adaptive ODE solvers to advance time-varying output layer coefficients. By fundamentally bypassing gradient descent training, it achieves faster training, higher accuracy, and explicit temporal causality across various time-dependent PDEs. This study presents the first systematic investigation into initialization schemes for spline-based KANs. It proposes LeCun/Glorot-inspired variance preservation schemes and a tunable power-law initialization family. Large-scale experiments involving over 126K model instances demonstrate that power-law initialization consistently outperforms baselines in function fitting and PDE solving. Furthermore, the Glorot scheme shows significant gains in large-parameter models, with NTK spectral analysis revealing the underlying optimization dynamics. LD-EnSF replaces expensive full-space numerical forward simulations with a learnable latent dynamics network (LDNet), migrates the Ensemble Score Filter (EnSF) entirely into an extremely low-dimensional latent space, and aligns sparse irregular observations using a history-aware LSTM encoder. This synergistically accelerates data assimilation by several orders of magnitude while maintaining high precision. This paper systematically refutes the promise of "zero-shot super-resolution" for Machine-Learned Operators (MLOs) such as the Fourier Neural Operator (FNO). By decomposing multi-resolution inference into two sub-capabilities—"resolution interpolation" and "frequency information extrapolation"—the authors find that FNO fails at both and exhibits severe aliasing. Neither physical constraints nor band-limited learning effectively addresses this issue. Finally, a simple yet effective multi-resolution training protocol is proposed, achieving robust cross-resolution generalization using a minimal amount of high-resolution data. CFO "borrows" flow matching from generative modeling to learn the right-hand side (RHS) dynamics of time-varying PDEs. By fitting splines to trajectories and using finite differences to estimate temporal derivatives at nodes as velocity field labels, it trains a neural operator to regress this analytical velocity. This bypasses backpropagating through ODE solvers (unlike Neural ODEs), enables training on irregular time grids, and allows inference at arbitrary resolutions. Using only 25% irregularly sampled data, it reduces relative error by up to 87% compared to full-data autoregressive baselines. This paper combines the classical Domain Decomposition Method (DDM) with neural operators to propose a "local-to-global" framework. A local neural operator is trained only on randomly generated basic shapes (simple polygons). During inference, an arbitrary geometric domain is partitioned into small subdomains, solved domain-by-domain, and stitched using additive Schwarz iterations (termed Schwarz Neural Inference, SNI). This reduces the relative error on completely unseen geometries from 20%~167% (direct inference) to single digits. RealPDEBench is the first scientific machine learning benchmark that packages real-world experimental measurement data alongside paired numerical simulation data . Covering 5 complex physical systems, 3 task categories, 9 metrics, and 10 baselines, it systematically reveals the significant gap between simulation and real-world data and demonstrates that "pre-training on simulation followed by fine-tuning on real data" consistently improves both accuracy and convergence speed. Browse all 69 Physics & Scientific Computing papers → 🌍 Earth Science (7) ¶ This paper adapts the "Common Task Framework (CTF)" approach—which catalyzed benchmarks like ImageNet and AlphaZero in NLP/CV—to seismology. It provides three multi-scale seismic wavefield datasets alongside a 12-point scoring protocol using hidden test sets. Evaluating 18 mainstream scientific machine learning models reveals that most complex architectures fail to outperform a naive "all-zero" baseline. RainPro-8 utilizes a MaxViT-U-Net with only 36.7M parameters to fuse multi-source data from radar, satellite, and Numerical Weather Prediction (NWP). Through "ordered consistent loss + single-forward prediction," it outputs high-resolution probabilistic precipitation forecasts for Europe over 8 hours in one go. It outperforms existing NWP, extrapolation, and deep learning nowcasting methods while being 48x faster in inference than MetNet-like models. For Weather Foundation Model (WFM) fine-tuning, this paper proposes WeatherPEFT: the forward pass uses Task-Adaptive Dynamic Prompting (TADP) to extract "variable × resolution × spatiotemporal" task features from encoder embedding weights to generate soft prompts, while the backward pass employs Stochastic Fisher-Guided Adaptive Selection (SFAS) to update only a small subset of parameters with the highest Fisher information. It matches or exceeds Full-Tuning on three downstream tasks using only ~0.3%–4% trainable parameters. OmniField models "scientific observation data" (climate, air pollution) as a continuous neural field conditioned on available modalities . Utilizing Multimodal Crosstalk (MCT) blocks and Iterative Cross-Modal Refinement (ICMR), it aligns heterogeneous signals before decoding. This unified framework supports reconstruction, interpolation, and prediction without gridding or interpolation preprocessing, reducing average error by 22.4% relative to 8 strong baselines while maintaining performance under severe sensor noise. GeoFAR decomposes the low-frequency bias in climate super-resolution into two problems: "under-represented frequency components" and "missing geographical conditions." It utilizes DCT frequency convolutional kernels to extract fine-grained frequency band representations and modulates these representations pixel-wise using a geographic implicit representation (Geo-INR) composed of longitude, latitude, and elevation. This approach significantly reduces high-frequency errors and prediction biases in complex terrain across multi-scale climate downscaling tasks such as ERA5, PRISM, and CERRA. TianQuan-S2S integrates "long-term climatological means" into patch embeddings via attention fusion and injects learnable Gaussian noise into each layer of a ViT. This specifically addresses the "model collapse" (increasingly blurry predictions) of data-driven models in 15–45 day subseasonal forecasting, outperforming both the numerical model ECMWF-S2S and the data-driven FuXi-S2S on the ERA5 dataset. This paper extends Neural Tangent Kernel (NTK) theory to Full Waveform Inversion (FWI), proposing a "Wave-based NTK" to unify the characterization of traditional FWI and Continuous Representation FWI (CR-FWI). It explains the phenomenon "why INR representations are more robust but converge slowly at high frequencies" through eigenvalue decay rates. Based on this, it designs IG-FWI, a hybrid of INR and multi-resolution grids, achieving a superior trade-off between robustness and convergence speed. 📡 Signal & Communications (7) ¶ Three core improvements are proposed from an SSM perspective: exponential-trapezoidal discretization, complex-valued state spaces, and Multi-Input Multi-Output (MIMO) formulation. These enhance model quality and state-tracking capabilities significantly without increasing decoding latency, pushing the performance-efficiency Pareto frontier forward. Proposes Directer (Dynamic Rejection Steering), which significantly enhances the instruction-following capabilities of LLMs by dynamically adjusting KV cache steering intensity and introducing plausibility constraints at each decoding step, while avoiding text quality degradation caused by oversteering. To address the fatal issue in LLM-driven lossless compression where prediction probabilities must be bit-level identical across encoder and decoder to avoid "cascading collapse," this paper proposes PMATIC—an alternative to arithmetic coding that quantizes bit probabilities into bins and uses low-entropy helper bits to synchronize both ends on the same quantized probability. PMATIC tolerates bounded prediction mismatches, theoretically guarantees correct decoding, and achieves perfect restoration under real-world cross-machine non-determinism while maintaining compression rates significantly superior to traditional tools like gzip and cmix. The authors implement the classic information-theoretic Gray-Wyner Network as a learnable three-channel codec, utilizing a \(\beta\) -parameterized objective to decouple "common" and "private" information between two vision tasks while enabling an adjustable tradeoff between "transmit rate" and "receive rate." Addressing the limited representation of binary spikes and surrogate gradient mismatch in Spiking Neural Networks (SNNs), this paper proposes Parametric Invertible Transformation (PIT). PIT applies conjugate invertible linear transformations before and after neuron firing: "rearranging" the membrane potential distribution into a quantization-friendly form before firing and "augmenting" integer spikes into spatiotemporal real-valued outputs after firing. This is coupled with a modified surrogate gradient that pushes inputs away from quantization decision boundaries. The method also characterizes SNN spatiotemporal representation capacity through linear algebra. Across CIFAR, ImageNet, and DVS datasets, various architectures achieved new SOTA results (e.g., SEW ResNet34 improved by 5.62%). EfficientMPT replaces the \(O(n^2)\) standard attention in Transformer-based error-correcting code (ECC) decoders with a linear-complexity EEC attention based on "global query vectors + element-wise multiplication." While maintaining error correction performance comparable to state-of-the-art (CrossMPT), it reduces GPU memory and FLOPs by dozens of times for long LDPC codes. Its parameter count is independent of code length, allowing it to serve as a fine-tuneable "foundation model" for error correction. To address Wireless Signal Recognition (WSR) pre-training, this work introduces the "noising-denoising" paradigm of diffusion models into signal self-supervision and proposes TS-DDAE. Gaussian noise is injected into IQ signals in both temporal and spectral domains simultaneously, followed by a joint restoration using a specialized dual-encoder TS-Net (temporal self-attention + spectral channel attention). The learned representations outperform the best baseline by an average of 1.32% across 4 datasets and multiple tasks like AMC/WTC, exceeding the AMC SOTA model IQFormer by approximately 8.75%. 👥 Social Computing (17) ¶ INTIMA distills three psychological theories—parasocial interaction, attachment, and anthropomorphism—along with qualitative coding of real Reddit user posts into a benchmark containing 31 behaviors and 368 emotional probes. By using LLMs to automatically label model responses as "Reinforcing Companionship," "Maintaining Boundaries," or "Neutral," the study finds that Gemma-3, Phi-4, o4-mini, GPT5-mini, and Claude-4 all significantly lean toward reinforcing companionship. Notably, models tend to set fewer boundaries as user vulnerability increases. Addressing the dual-use risks of knowledge editing (KE), this paper proposes EditScope to infer edited target entities from weights (up to 99% accuracy) and a training-free edit reversal method based on SVD bottom-rank approximation (up to 94% reversal rate), relying solely on edited weights without requiring edit prompts or original weight information. Provided with only 20 Big Five personality item responses of an individual, LLMs are tasked to role-play and predict that individual's responses to 9 other psychological scales. The results show that the "inter-scale correlation structure" reconstructed by LLMs aligns highly with real human data ( \(R^2>0.88\) ). Analysis of reasoning chains reveals a two-stage abstraction process where LLMs compress raw scores into natural language personality summaries before reasoning—indicating genuine psychological reasoning rather than mere semantic pattern matching. This work proposes SocialHarmBench, the first safety evaluation benchmark specifically targeting socio-political harms. It consists of 585 prompts covering 7 domains and 34 countries, revealing systematic safety vulnerabilities of current LLMs in politically sensitive scenarios such as historical revisionism and propaganda manipulation. Ours proposes SAGE, a unified VPR training framework: it introduces a lightweight Soft Probing module to enhance local feature discriminativity, reconstructs an online affinity graph merging geographical distance and visual similarity every epoch, and focuses on the hardest samples through greedy weighted clique expansion. By freezing the DINOv2 backbone and training only 1.96M parameters, it achieves comprehensive SOTA across 8 benchmarks. This paper constructs the BiasFreeBench benchmark, which systematically compares eight mainstream debiasing methods (four prompting + four training) within a unified framework for the first time. Focusing on bias evaluation at the LLM response level, it proposes the Bias-Free Score metric and finds that prompting methods (especially CoT) generally outperform training methods, while DPO shows outstanding generalization across bias types. Reveals the "premature satisfaction" issue in DPO—where the gradient is unnecessarily decayed by pessimistic signals from the reference policy when it assigns a lower probability to the chosen response than the rejected one (~45% of pairs), even if the policy remains incorrect ( \(\Delta_\theta < 0\) ). Proposes HyPO (a one-line change: \(\max(0, \Delta_{ref})\) to clip the reference margin), achieving a 41.2% relative improvement over DPO on AlpacaEval 2.0. Ours proposes the RAVEN audit framework to detect concept-conditioned semantic divergence in LLMs—a propaganda-like behavior pattern where high-level conceptual cues (ideologies, public figures) trigger abnormally consistent stance responses—by combining intra-model semantic entropy and cross-model divergence. This paper proposes introducing Tsallis entropy (a generalized form of Shannon entropy) into Test-Time Adaptation (TTA) for VLMs, further developing it into Adaptive Debiasing Tsallis Entropy (ADTE). By customizing the debiasing parameter \(q^l\) for each category, ADTE selects more reliable high-confidence views than Shannon entropy without distribution-specific hyperparameters. It outperforms SOTA on ImageNet, its five variants, and ten cross-domain benchmarks. The authors propose GRADIEND—a gradient-based encoder-decoder architecture that learns interpretable monosemantic features (exemplified by gender) from model gradients through a single bottleneck neuron. It can identify which weights encode specific features and directly modify model weights via the decoder to eliminate bias, achieving SOTA debiasing performance on all baseline models when combined with INLP. Browse all 17 Social Computing papers → 🛡️ AI Safety (141) ¶ The authors construct VPI-Bench (306 samples), the first comprehensive visual prompt injection attack benchmark, systematically evaluating the security of Computer-Use and Browser-Use Agents across 5 platforms. Findings reveal that Browser-Use Agents are extremely fragile (100% AR on Amazon/Booking), and even Anthropic's CUA exhibits serious vulnerabilities (up to 59% AR), with system prompt defenses proving ineffective. MoReBench proposes evaluating the structural quality of the reasoning process (rather than the correctness of the final conclusion) of reasoning models across 1,000 moral dilemmas using 23,018 expert-written rubric criteria. The study finds that neither scaling laws nor performance on math/code benchmarks can predict a model's moral reasoning capabilities. This study presents the first controlled evaluation comparing AI agents and human cybersecurity experts within the same real-world production network (a university environment with approximately 8,000 hosts). The researchers simultaneously deployed 10 professional penetration testers, 6 existing agent scaffolds, and a self-developed multi-agent framework, ARTEMIS. ARTEMIS ranked second overall, identifying 9 valid vulnerabilities with an 82% valid submission rate, outperforming 9 out of 10 human professionals. In contrast, off-the-shelf scaffolds like Codex and CyAgent performed poorly. The results highlight AI's advantages in systematic enumeration, parallel exploitation, and cost-efficiency, while revealing critical weaknesses in GUI operations and high false-positive rates. This paper points out that traditional class unlearning assumes "class label = target concept to be removed," whereas real-world deletion requests often involve a mismatch between the two. To address this, the authors decouple forget data, model output, and target concepts into three label domains, defining target/model/data mismatch tasks. They propose the TARF framework, which utilizes "representational gravity" to identify data sharing the same concept hidden in the remaining set, and employs a three-phase dynamic objective (annealed gradient ascent + target-aware gradient descent) to precisely extract the target concept and approximate the retrained model. This paper characterizes "sample contribution to training" as the input sensitivity of the pre-trained model toward that sample. It proposes MU-Mis, which utilizes only the pre-trained model and forget data without accessing any remaining data. By minimizing the "sensitivity difference between target and irrelevant classes," it directly erases the contribution of forget samples. It is the first remaining-data-free method to achieve utility parity with state-of-the-art (SOTA) methods that rely on remaining data. This paper presents the first systematic study of watermark-based user-level detection and attribution for AI-generated content. It providing theoretical analysis (TDR/FDR/TAR bounds), an efficient watermark selection algorithm (A-BSTA), and cross-modal (image and text) experimental validation. The results demonstrate that detection and attribution inherit the accuracy and (lack of) robustness of the underlying watermarking methods. LAMDA constructs a long-term malware benchmark covering over 1 million Android APKs from 2013 to 2025. Using Drebin static features, family labels, and multiple temporal splitting systems, it reveals that existing malware detectors degrade rapidly under real-world concept drift. LingoLoop injects subtle adversarial perturbations into input images to trap Multimodal Large Language Models (MLLMs) into infinite repetitive generation. By utilizing "POS-prior-based EOS suppression" and "hidden state contraction-induced loops," it produces up to 367× more tokens than normal inputs when generation limits are relaxed, causing compute/energy-exhaustion Denial-of-Service (DoS). This paper proposes the Banded Inverse Square Root (BISR) matrix factorization method. By imposing a banded structure on the inverse correlation matrix (rather than the correlation matrix itself), it achieves the first asymptotically optimal factorization error bound for multi-epoch Differentially Private SGD, accompanied by a low-storage optimized variant, BandInvMF. This paper utilizes the concepts of oracle machines and "reductions" from computability theory to rigorously formalize diverse Human-in-the-Loop (HITL) human oversight schemes into three categories—Trivial Monitoring, Endpoint Action, and Involved Interaction. Based on this, it establishes a failure mode classification system and analyzes blind spots in UK/EU laws, ultimately revealing an unavoidable "Accountability ↔ Technical Interpretability" trade-off. Browse all 141 AI Safety papers → 📂 Others (115) ¶ AnyUp proposes the first encoder-agnostic learnable feature upsampling method. By employing feature-agnostic convolutional layers and a window attention mechanism, it can perform high-quality upsampling for arbitrary visual features across any resolution with only a single training session. It achieves SOTA performance on tasks such as semantic segmentation and depth estimation. This paper systematically examines nine common myths in the field of graph machine learning regarding oversmoothing, oversquashing, homophily/heterophily, and long-range dependencies. By refuting each myth with concise counter-examples, the authors decouple "oversquashing" into two independent concepts— computational bottleneck and topological bottleneck —resolving widespread conceptual confusion in the literature. SmellNet establishes a machine olfaction benchmark using low-cost portable gas sensors to collect real-world temporal signals from 50 natural ingredients and 43 categories of odor mixtures, proposes SCENTFORMER which integrates temporal differencing, sliding windows, and GC-MS chemical priors. The rigid "direct state swap" in Parallel Tempering (PT) is replaced with an "accelerated swap": neural transports (Normalizing Flows / Controlled Diffusion / Diffusion Models) are used to push the two states towards each other before performing a Metropolis acceptance check. This enables high-probability exchanges even when adjacent annealed distributions have minimal overlap, significantly increasing the round-trip count between the reference and target distributions while maintaining the asymptotic unbiasedness of MCMC and providing low-variance free energy estimates. Propose the Local Bayesian Influence Function (BIF), which replaces the infeasible Hessian inverse in classical influence functions with covariance estimated via SGLD sampling. This achieves architecture-agnostic data attribution for models with billions of parameters and reaches SOTA in retraining experiments. PlanetAlign is proposed as a PyTorch-based network alignment benchmark library that integrates 18 datasets across 6 domains, 14 methods covering three major categories (consistency, embedding, and Optimal Transport), and standardized evaluation workflows. Through large-scale systematic experiments, it reveals the comprehensive lead of OT-based methods (PARROT/JOENA) in effectiveness and the differentiated performance of various methods in scalability and robustness. This paper utilizes algebraic geometry tools to systematically analyze MLPs and CNNs with polynomial activations: it proves the finite identifiability of MLPs and the unique identifiability of CNNs, reveals that sparse subnetworks correspond to singularities on the neuromanifold, and provides a geometric explanation for the sparse bias in MLPs through the lens of "critical exposure"—a property that CNNs do not possess. LPWM is the first self-supervised object-centric world model capable of scaling to real-world multi-object datasets. The core innovation is learning independent latent action distributions for each particle (per-particle latent actions). By utilizing a causal spatio-temporal Transformer to encode all frames in parallel, it supports diverse conditional generation (actions, language, image goals, multi-view). It achieves SOTA in video prediction and demonstrates imitation learning capabilities (89% success rate on OGBench task3). This paper proposes the TabStruct evaluation framework and the global utility metric to measure the structural fidelity of tabular data generators without requiring ground-truth causal graphs. By systematically comparing 13 generators across 29 datasets, the study reveals that diffusion models significantly outperform other methods in maintaining global structural integrity. By decomposing AI model errors into bias (systematic misalignment) and variance (incoherent behavior), this study finds that: longer reasoning leads to higher incoherence; larger models become more incoherent on difficult tasks. This suggests that future super-intelligent AI is more likely to manifest "industrial accident" style unpredictable failures rather than consistently pursuing incorrect goals. Browse all 115 Others papers →