News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 39 of 54 · The Research front page
The authors introduce chunk-level sparse autoencoders (SAEs) that encode mean-pooled activations over contiguous token chunks instead of individual tokens. They show that these models learn features that are more semantically coherent and selective, outperforming standard token-level SAEs in downstream tasks such as retrieval, reasoning detection, and generation steering.
Long and colleagues address a failure mode of on-policy training for vision-language models: a model can reduce object hallucinations simply by making fewer claims, including fewer correct ones. Their framework instead supplies dense evidence about which objects are present or absent and assigns rewards separately to parts of a response, producing descriptions that the authors report are both more informative and more faithful.
Qiu and colleagues modify mixture-of-experts routing so each layer can consult the expert-selection distributions produced for the same token by preceding layers. In an approximately 8 billion parameter model trained from scratch on 100 billion tokens, the method reduced final training loss from 1.6393 to 1.6184 while adding modest computational and memory costs.
Wang and colleagues formulate feed-forward reconstruction and conditional generation as two inference modes of one flow-based model. A single step predicts the conditional mean, while multiple denoising steps produce sharper solutions on the learned data distribution, with accuracy depending on how strongly the input observations constrain the answer.
ProDyGS is a dynamic Gaussian splatting method for novel view synthesis from a single static camera video. By generating proxy images from estimated depth maps, it creates synthetic multi-view supervision. Experiments on the DyNeRF dataset show state-of-the-art quality, outperforming methods that rely on scene flow priors.
The authors introduce CUA-SWE, a benchmark and evaluation pipeline for software engineering with computer use. They find that frontier agents which can view and interact with application interfaces complete tasks more successfully than agents limited to code editing, especially in web and game domains.
The authors present UnStep, a training-free wrapper that accelerates already-distilled causal video diffusion models by running them with fewer denoising steps and a limited temporal KV cache. Two inference-only mechanisms—clean-cache refinement (renoise and refine) and truncated SVD on attention projections—recover the quality lost by these reductions, while an optimized runtime stack further boosts throughput. UnStep achieves 50 frames per second on a single H100 GPU and 77 FPS on a GB200 GPU without quality degradation, all without retraining.
Ma et al. introduce Leverage Score Sampling (Lev), an iterative selection method that uses leverage scores to maximize data diversity efficiently, eliminating costly matrix recomputation. On CommonCrawl data, Lev improves accuracy by up to 1.31% over strong baselines and reduces bits-per-byte on code data by 3.08% relative to DiSF.
The authors introduce a chance-constrained formulation for fine-tuning large language models that bounds the fraction of safety examples whose performance degrades beyond a specified threshold, rather than minimizing average safety loss. They develop a constraint-aware gradient descent method that minimally adjusts the fine-tuning direction to satisfy this constraint, and show it outperforms existing safety-preserving baselines across three tasks and three models.
The authors propose OREO, a framework that enhances visual fidelity of 3D generators by creating a dynamic optimization loop. They use a 2D image editing model to refine rendered views of 3D outputs on the fly, producing high-quality pseudo-targets. Then, a latent contrastive objective distills the improvements back into the 3D generator, resulting in significantly more realistic assets.
The authors present YODAS v3, a weakly labeled speech corpus with over 1.1 million hours of 48kHz stereo audio in 147 languages. They introduce language balanced crawling techniques and demonstrate that the dataset supports competitive ASR and neural codec training. The corpus is released under CC BY 3.0 and available on Hugging Face.
The authors identify a shared early-layer attention routing circuit in VQ-tokenized VLMs that drives object hallucinations, and propose a three-gate diagnostic to detect it. Crucially, swapping a continuous projector for a VQ tokenizer installs the circuit, while a targeted L0 ablation reduces hallucination in open-ended generation by 31% — outperforming mechanism-agnostic baselines like VCD and DoLA.
The authors analyzed production traces from two companies and evaluated 14 eviction algorithms for LLM prefix caching. They found that simple Least Recently Used (LRU) performs competitively with sophisticated policies, despite substantial headroom to Belady’s optimal. The reason is that prefix reuse is dominated by the regular pacing of active sessions, making recency a strong predictor.
The authors identify answer flips—cases where nearly identical images cause vision-language models to produce different answers—and propose FlipDir, a training-free inference-time method that estimates a low-rank subspace of flip-inducing activations and selectively steers hidden states during decoding. A margin-based gate restricts steering to uncertain steps, recovering original predictions while leaving stable ones unchanged. Across 18 settings spanning scientific reasoning, robot-scene understanding, and medical VQA, FlipDir consistently outperforms existing methods on a combined metric of recovery and preservation.
Zhou et al. introduce Support-Compiled Feature Folding (SCFF), a training-free inference method for tabular foundation models that avoids full-width feature mixing by organizing columns into a core and a tail. On an 18-dataset wide-table benchmark, SCFF improved accuracy and negative log-likelihood across all six evaluated frozen backbones, while cutting median peak GPU memory by 2.09-2.36x. The method also used saved memory to retain more evidence, yielding accuracy gains of 4.06 and 3.72 points over widest feasible single-leaf baselines.
The authors introduce COILD, a parallel corpus of 1.16 million human-verified sentence pairs covering 20 Indian language pairs from four language families. Fine-tuning IndicTrans2 and NLLB-200 on COILD produced higher translation quality across all metrics and domains.
The authors introduce RECLAIM, a benchmark of 100 NeurIPS 2025 papers with predefined reproduction targets and GPU-hour budgets. They test four AI agents once per paper; the best agent reproduces 41% of Run-tier papers (code, data, weights provided), 27% of Retrain-tier (no weights), and 15% of Reimplement-tier (no code). Failed attempts use only 29% of their budget on average, and the most common error is implementing methods without verifying against the paper's reported numbers.
Ginjala et al. examine how post-training compression (pruning, quantization, distillation) of Whisper ASR models redistributes error burden across demographic groups. On FairSpeech, pruning Whisper-large-v3 widens the temporal-taxation differential between Black/AA and Asian speakers by 111%, while INT4 quantization on West African accents increases catastrophic transcript loops by factors of five to seven. The authors conclude that fairness audits on full-precision models do not capture the deployment-time burden compression imposes on already-marginalized speakers.
Hansen et al. present InsertionWM, a framework that trains a single world model on up to 90 geometrically diverse insertion tasks using wrist-mounted camera images and robot proprioception. The model achieves 56% zero-shot success on unseen objects with unknown geometry, versus 7% for a model-free baseline. Finetuning the generalist model on held-out objects improves data-efficiency and asymptotic performance.
Neitemeier et al. address two distinct forms of imbalance in mixture-of-experts models: persistent skew across an optimizer step and uneven expert use within individual microbatches. In experiments on 7.5-billion-parameter models trained for up to 500 billion tokens, their methods improved global and local balance while maintaining, and in some settings improving, downstream performance.
Sunagad, Jesslen and Kortylewski dissected the training objectives behind DINO-style self-supervised vision transformers by removing or modifying individual components. They find that aligning two geometrically displaced global crops from the same image is sufficient for strong semantic correspondence, whereas local crops and patch masking do not independently create comparable structure.
The paper identifies a flaw in multimodal learning: optimization yields asymmetric gains in predictive certainty between strong and weak modalities, driving imbalanced contributions. The authors propose Max Confidence Regularization (MaxCR) to dynamically intervene in modality confidence using a nonlinear sparsity measure, penalizing overconfident strong modalities and encouraging underconfident weak ones. Experiments on several datasets show MaxCR improves overall performance over state-of-the-art baselines.
The authors develop in-situ representation refinement, where support labels guide changes to the episode's representations during a forward pass. They propose RefineICL, an attention-gated, FFN-free contextual stack that achieves 0.93836 OVR-AUC and 0.87173 accuracy on AMLB29 and 1644.8 Elo on TabArena, 31.4 above TabPFN-3. A direct intervention confirms that evolving support states are essential: removing one intermediate support update increases final query cross-entropy in all 72 tested episodes.
The authors introduce a flow-matching generative model whose velocity field is parameterized by a permutation-equivariant Graph Meta Network. This allows them to learn directly from collections of independently trained neural networks without the need for approximate neuron alignment. The method generates new networks that closely reproduce the joint statistics of accuracy, functional similarity, and weight similarity observed in training collections, and a single conditional model can produce task-specific networks for heterogeneous architectures and previously unseen hidden-width configurations.
Xing and colleagues introduce CinematicVQA, a benchmark for reasoning about how cinematographic choices shape what viewers perceive and how a scene functions narratively. Their evaluations reveal a semantic gap: large vision-language models describe visible presentation more reliably than they identify the filming techniques behind it, while targeted fine-tuning improves narrative and multi-step reasoning.
Chen and colleagues survey how autoregressive video generators can preserve relevant history after it leaves their limited context windows. They argue that effective memory requires more than storing information: retained states must remain accurate, retrievable, and causally influential during later generation.
The authors propose an interaction-centric framework that uses a parameterized universal gripper abstraction to transform RGB-D observations into a canonical gripper frame. This enables zero-shot cross-embodiment and cross-viewpoint transfer for two-finger gripper manipulation tasks, demonstrated in both simulation and real-world experiments. The approach simultaneously achieves competitive benchmark performance and extreme generalization to heterogeneous robot platforms.
Zhang et al. introduce Endpoint-Constrained Optimization (ECO), a training-free post-processor for end-to-end driving models that fixes the vehicle's executed history and the policy's predicted endpoint, then optimises the intermediate waypoints for physical plausibility. Across two closed-loop simulators and five datasets, ECO improved the aggregate score of every evaluated policy, with gains up to 71% in HUGSIM score for the VaVAM model.
Tallini et al. introduce LocUS, which uses the unembedding matrix to define a property-specific subspace and selects a sparse subset of attention heads. Steering is applied only within that subspace and on those heads. The method matches or outperforms standard steering baselines on three tasks while intervening on under 6% of parameters and better preserving general capabilities.
Cestola et al. examine position-independent key-value cache reuse, a technique intended to accelerate retrieval-augmented generation by reusing cached representations of text chunks at different prompt positions. They find that common aggregate accuracy metrics and existing datasets can conceal the accuracy loss caused specifically by reuse, then introduce Boxoffice to generate more demanding evaluations.
The authors show that scoring predicted outcomes by their distance to the final goal can break latent world model planning even when dynamics are exact and short-horizon search is globally optimal, whenever the route to the goal initially moves away. They introduce Anchored Planning, a training-free method that reuses the frozen model's own offline trajectories to select an intermediate target. Across Cube, PushT, Reacher, and TwoRoom, it outperforms the LeWM planner and additional final-goal search on long-range goals.
The authors introduce ExplorationBench, a benchmark that evaluates AI systems' ability to explore unknown environments by forming hypotheses, designing experiments, and iterating on results. Its two executable sandboxes make every answer exactly checkable while ensuring that pretraining recall alone cannot solve the tasks. Evaluating ten AI systems, the authors find that the strongest can acquire and apply these unfamiliar rules, though performance varies across trajectories and continued exploration can stall or reverse earlier gains.
Hussein et al. present FAST-CAP, a causality-aware framework for LLM-based simultaneous speech-to-speech translation. They introduce a factorized architecture (FAST) that separates lexical and acoustic modeling, and a causality-aware adaptive policy (CAP) for read/write decisions. Experiments on Spanish, German, and French show improvements of up to +1.2 BLEU and 26% latency reduction over fixed policies.
The authors reformulate distributionally robust optimization (DRO) as a problem over transport maps and prove that optimal maps are cyclically monotone. They show standard adversarial training violates this property, wasting transport cost. They propose multi-start particle ascent and input-convex neural network parametrizations that enforce cyclical monotonicity, achieving improved robustness on regression, image classification, and control tasks.
Zhang et al. introduce Trajectory-to-Step Policy Optimization (T2SPO), a method that extracts step-level supervision from historical successful trajectories to guide reinforcement learning of LLM agents. By pairing each state in a successful trajectory with its remaining number of steps, they train a frozen TabPFN regressor to predict progress in new rollouts; the change in predicted distance between consecutive states provides auxiliary credit for each action. Experiments on ALFWorld and WebShop with 1.5B and 7B models show that T2SPO consistently improves task success over GRPO, a baseline that only uses outcome rewards.
The authors systematically evaluate whether token-level certainty scores from LLMs predict correctness, distinguishing between identifying questions a model will likely answer correctly and distinguishing correct from incorrect responses to the same question. They find that certainty is generally better at the former, and that the timing of useful information differs: question difficulty appears early, answer correctness appears late. A test-time compute strategy built on these findings raises accuracy from 78.71% to 79.54% while cutting generated-token cost by 82.4%.
Viennot, Baudart, and Lelarge treat the interface between an AI agent and a proof assistant as an optimization target. Their evolutionary process built rocq-mcp-evolve, a set of agent-facing tools that outperformed both a compiler-only interface and an established Rocq server across four language models.
Huang evaluates retention, which determines what remains in bounded memory, separately from selection, which chooses stored information after a query arrives. Across 300 matched episodes, query-aware selection improved required-fact recall by 15.5 percentage points when access was held constant, while a confounded comparison suggested a much larger 68.7-point gain.
Yang et al. introduce Zero2Repo, a benchmark that gives coding agents an empty workspace, a product requirements document and an interface contract, then asks them to deliver an entire software repository. On 11 tasks, the strongest evaluated agent completed 10, while failures were usually nearly correct implementations undone by one omission or uncommon rule.
Jin et al. developed SMART, a multi-agent subtitle translation system that adapts its workflow to each series without changing the underlying language models. It achieved the lowest translation-error penalty across all 15 directions in the authors’ new benchmark, improving by 6.9% on average over the strongest competing agent system.
Mathur et al. introduce DSB-DG, a benchmark for testing whether real-time voice agents answer faithfully from supplied documents. Across cascaded and speech-to-speech systems, the authors find that grounding deteriorates under longer contexts and sustained dialogue, with failures often producing unsupported claims rather than explicit abstention.
Huzaifa, Mavali and Eisenhofer test whether latent communication, where agents exchange internal representations rather than text, creates a distinct safety risk. Across several attacks, they find that lightweight links connecting otherwise aligned agents can be optimized to elicit harmful responses while retaining useful performance, and can also be repaired without retraining the agents.
The authors propose Density-Aware Reward Aggregation (DARA), a method that calibrates reward contributions in multi-reward reinforcement learning for LLMs. DARA dynamically adjusts weights based on the frequency of active learning signals, leading to faster convergence on format and length compliance goals. In experiments, DARA achieved up to 26% fewer training steps on tool calling and up to 65% fewer steps on math reasoning compared to the baseline GDPO.
The authors propose APIVIS, a training-time framework that integrates finite-budget Gumbel search into reinforcement learning with verifiable rewards (RLVR) for mathematical reasoning. By combining direct and searched responses in each rollout group and applying selective supervision to search-improved tokens, they achieve substantial gains over existing search-based methods across benchmarks and model scales.
Duarte et al. present ABSENTIA, a security scaffolding that directs LLM agents to systematically audit web application backends for broken access control. The agents build a route–code graph and apply invariant falsification to infer intended authorization properties, reporting routes where they are not enforced. On a new benchmark of 30 real-world advisories (BAC-Bench), ABSENTIA recalls 19 vulnerabilities with 17 under strict credit; CodeQL and Semgrep recall none.
Yang et al. introduce an automated method for improving the runtime harness around a GUI agent without changing the underlying model. Across six backbones on OSWorld-Verified, the authors report consistent held-out improvements, including a 12.33-point gain for Qwen3-VL-32B-Instruct. A harness optimized on one benchmark also transferred to WindowsAgentArena, where it improved GPT-5 by 13.87 percentage points.
Li and colleagues evaluate language models together with the software harnesses that expose tools, manage context and return errors. Across 66 configurations and 6,204 scored trajectories, they find no consistently best model or harness: rankings often reverse when either the harness or task changes, while higher spending and vendor-native pairings do not reliably improve results.
Mehta, Ritchie and Chen post-trained Kimi K2.7 Code using reinforcement learning with verifiable rewards, without a supervised warm-up. The resulting model achieved higher pass@1 on every evaluated benchmark, including later-released task sets, unfamiliar agent harnesses and multi-hour software projects.