News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 23 of 54 · The Research front page
Yang et al. introduce CRBugBench, which tests whether review agents can identify bugs that humans later confirmed and fixed, rather than whether agents reproduce human review comments. The best evaluated configuration caught 71.9% of bugs in open-source pull requests within its top five comments, but performance declined for complex changes, non-local bugs and proprietary code.
Guan et al. repeatedly sampled answers from intermediate points along fixed reasoning trajectories, measuring how accuracy changed as more reasoning was retained. They report that 64.3% of thinking tokens occurred after the earliest measured point of maximum empirical accuracy, while completing the full trajectory reduced accuracy by 9.03 percentage points on average.
The authors introduce Power-Law Random Feature (PLRF) models—simple quadratic models with heavy-tailed, power-law distributed inputs—that exhibit the same intermittent training spikes seen in large language model (LLM) pre-training. They develop a layer-wise simulation algorithm that predicts when spikes will occur in deep neural networks, and in a case study on a Transformer trained with Adam, they derive a beta2 hyperparameter schedule that avoids spikes and improves performance.
The authors identify two failure modes in long-horizon machine learning agents: long-horizon deceleration, where reasoning intensity drops as a session stretches, and harness staleness, where custom tooling helps less as base models improve. Their workflow, Humanize MLE Agents (HMA), alternates two agents on a shared workspace with context resets. On 75 Kaggle tasks in MLE-bench it reached a 78.2% any-medal rate, 8.0 points above the strongest open-source baseline.
The authors propose BranchOPD, an on-policy distillation method that selects informative fork points using student uncertainty and teacher-student disagreement, then resamples student continuations from shared prefixes. Across mathematics and agent benchmarks, BranchOPD outperforms vanilla OPD and seven-rollout OPD while generating roughly 19% fewer student tokens on ALFWorld and 31% fewer on WebShop.
The authors introduce AgentSWE, a benchmark for agent software engineering in which coding agents are commissioned to build, edit, or optimize a complete agent from a natural-language requirement. Across 25 tasks, they evaluate five models in the Codex harness plus other pairings, grading frozen submissions against requirement-derived rubrics on held-out cases. The strongest builders average 19.1 on creation and 31.6 on editing, and none closes more than 5 percent of the remaining gap in optimization.
Hu and Chong train a general-purpose verifier to assess an entire coding agent’s execution history, rather than judging only its final patch or returning a scalar score. On held-out CodeForge-Preview trajectories, supervised fine-tuning followed by reinforcement learning improved classification accuracy by 21 percentage points over the base model.
Wang et al. introduce T2Mem, which defers memory construction to inference and builds query-specific memory as the LLM reasons. On LoCoMo and LongMemEval, it outperforms strong baselines across two backbone LLMs under comparable token budgets while reducing construction overhead. The authors show that memory constructed at test time adapts to each query's information needs.
The authors identify a store-read gap in Mixture-of-Experts (MoE) models, where a fixed-precision checkpoint stores weights at one rate but decode reads them at a much higher rate (2.95× for BitsMoE). They propose DuoRate, which stores one uniform tier of a prefix-readable weight stream offline and then, for each workload, selects per-expert read tiers via a Lagrangian rate-distortion optimization that cancels routing frequency. On Qwen3-30B-A3B, DuoRate reading 2.3 bpw outperforms every 2-bit-class baseline by 6.3–47.0 average accuracy points, and at 2.9 bpw it matches 4-bit formats that read 1.4–1.6 more bpw.
Lee and colleagues developed Prompteina, a model that jointly generates protein sequences and all-atom structures conditioned on natural-language descriptions of desired properties. Across structural-control and binder-design benchmarks, it generally outperformed models requiring specialized input formats, and nine of 89 experimentally tested designs bound a previously unaddressed target.
Chen and colleagues introduce Rollout-Time Intervention (RTI), a reinforcement-learning method that turns selected failed agent rollouts into successful training trajectories using evaluation rubrics. Across five benchmarks and two model backbones, the authors report that RTI outperforms standard reinforcement-learning baselines and eventually increases unaided task success, rather than merely improving performance through intervention at test time.
Meng and colleagues investigate why Diffusion Transformers, or DiTs, can generate strong images while producing weaker visual representations than U-Net diffusion models. They report that adding long skip connections improves the quality of intermediate activations used for vision tasks, and propose that these connections let diffusion noise bypass the representation-forming core of the network.
Binz and Schulz built a single interpretable model from established cognitive principles, then fitted it jointly to 7,036,442 choices from 137 experiments. The model outperformed the tested black-box alternatives and, when simulated, reproduced 50 of 67 previously reported behavioral effects.
Zhang et al. trained a single reinforcement-learning policy from scratch on all 20 tasks of the NIST Assembly Task Board #1, using point-cloud observations and binary success rewards. The policy reached 98.2% overall training success, while separate experiments showed that policies could complete ten assembly pairs in one episode and select their own order.
Ye et al. developed an automated real-to-sim-to-real pipeline for teaching a mobile manipulator to pick up a user’s objects in a specific environment. Starting from a 20 to 30 minute phone scan and object images and dimensions, their strongest adapted policy improved from 0% baseline success to 62.5% in the target scene, with similar performance in simulation and on a real Stretch 3 robot.
The authors studied five representative LLM-based agent approaches for tier-1 SOC alert triage and found that all missed at least 40.4% of attack-related alerts. They then designed AIDA, a multi-agent framework that requires an explicit proposed decision, independent challenge, and stronger evidentiary requirements before dismissal, achieving an F1 score of 0.958 (compared to 0.371–0.744 for baselines) and reducing the false-negative rate to 3.1%.
Balepur, Tomlinson and Schnabel introduce CABRA, a framework that generates controlled coding tasks from call graphs and varies their difficulty along specific dimensions. Across 6,840 tasks, eight language models and six coding agents, the authors find that tools compensate for many model weaknesses, but tasks requiring detailed comparison and decomposition of code still cause agent accuracy to decline.
The authors introduce LLM-IDEA, an agent that couples a large language model proposer with a deterministic identifiability engine to triage plateaus in mechanistic world model learning. The engine computes the sensitivity matrix rank and certified exhaustion, enabling a three-way verdict: capability limit, resolvable within design class, or certified exhausted. On the Alien Universe testbed, LLM-IDEA under an identifiable protocol reached discovery depth at least three on 8/8 seeds versus 1/8 without the identifiability engine.
Jiang et al. propose V-CoLA, a training-free token compression framework specifically designed for linear attention within vision-language models. The method identifies critical vision tokens using a uniqueness-aware importance criterion and merges them adaptively, achieving strong performance retention and significant inference speedups across multiple benchmarks.
The authors introduce OmniDex, a massive simulation benchmark for dexterous grasping in cluttered scenes, created via a scalable seed-and-filter strategy that avoids costly scene-level optimization. They also propose an end-to-end grasping model that uses Soft Winner-Takes-All learning and human-inspired physical constraints, achieving state-of-the-art performance on the benchmark without the latency of post-optimization.
The authors introduce VAMR, a video agent that answers multiple questions about one recording in a shared tool-use trajectory instead of restarting exploration for each question. On LVBench it reaches 62.1 percent accuracy, 4.3 points above the prior VideoARM agent, while cutting reasoning rounds by 85.9 percent and processed frames by 61.4 percent.
Kwon et al. test whether language models using stored trajectory-reward pairs actually learn from rewards at inference time. They find that rewards are recognized but exert little influence: most improvement comes from the presence and task relevance of trajectories, resembling ordinary in-context learning more than reinforcement learning.
The authors present MiDashengLM-Spatial, the first open-source end-to-end audio-language model that jointly handles general audio understanding and spatial awareness. By augmenting an existing monaural LALM (MiDashengLM) with a spatial encoder (Spatial-Dasheng) via a hierarchical conditioning module, the model achieves superior performance on spatial reasoning benchmarks while remaining competitive on standard monaural audio tasks.
Gu and colleagues present a diffusion-based system that jointly separates a single RGB image into albedo, shading, surface normals, roughness and metallic maps, while also learning to render those maps back into an image. Across three synthetic scene benchmarks, the authors report competitive or improved estimation and rendering, with greater consistency among the predicted properties.
Authors from Meta Reality Labs present FaceGAN, a single-pass GAN for audio-driven facial animation that runs at 43 ms latency and 12x real-time on one GPU. They identify that a causal generator driven by i.i.d. noise cannot produce smooth motion without collapsing innovation variance, and introduce acausal noise shaping to dissolve this limitation without adding audio latency. FaceGAN matches or surpasses diffusion models in fidelity while avoiding their iterative sampling cost.
The authors introduce AgentHorizon, a benchmark of 1,373 instruction-trajectory pairs built from 166 hours of human-recorded computer-use sessions across three operating systems. By swapping instructions between closely related tasks, they create negative examples that test whether an AI judge can spot subtle failures in trajectories that look complete. Evaluating eleven judge models, they find that the best, GPT-5.5, achieves 80.9% balanced accuracy on the hardest split, and that providing tools helps some models but hurts others.
Prasanna, Tay and Westover test two ways to use model distillation, in which a weaker student learns from a stronger teacher, as an AI safety tool. In controlled model organisms, distillation made secret-keeping models more willing to disclose hidden behavior, while modified training procedures transferred mathematical capabilities with substantially less transfer of a proxy for misalignment. The work supports each strategy separately, not the broader claim that one must always succeed.
The authors prove an exact finite-sample correspondence between Distribution Matching Distillation (DMD) and Drifting Models, two approaches for one-step generative modeling. By converting the velocity-field from pretrained diffusion models into attraction force fields and estimating repulsion from the generative distribution, they show that DMD's optimization is equivalent to drifting model training. They then introduce Multi-Bandwidth DMD (MBDMD), which improves convergence and generation quality on image datasets.
The authors present LVSPM, a feed-forward model that jointly estimates camera poses and renders novel views from long, unposed image sequences. Trained on RGB images and pose supervision only, it scales to 256 views and outperforms prior pose-free and pose-dependent systems on three benchmarks.
The authors evaluated four LLMs as annotators for two machine translation quality schemes, MQM and ESA, comparing their agreement with human annotators across a 70-language-pair test set and WMT23/WMT25 data. LLM-human agreement matched or exceeded human-human agreement for some tasks, but both varied widely across annotation schemes, language pairs and domains, and remained unreliable for most settings.
The authors propose Nullify, a training-free, non-destructive method for LLM unlearning that operates entirely in activation space during inference. It achieves forget quality comparable to or better than existing gradient-based and preference-optimization methods on the TOFU and MUSE benchmarks while keeping utility near-lossless, all at a small fraction of their compute and memory requirements.
The authors show that minimizing reconstruction loss on calibration data during weight-only post-training quantization (PTQ) can hurt model performance, even on the very same data. To address this, they propose Distributionally Robust Quantization (DRQ), a post-hoc refinement that minimizes worst-case reconstruction loss over a constrained set of input activation distributions, and demonstrate gains across six representative PTQ methods on dense and mixture-of-experts LLMs.
GATOR reconstructs complete, textured 3D objects and their scene-relative pose from one or a few casual images. The authors combine a generative cascade with agentic refinement: the cascade produces an initial asset, which anchors a multimodal agent that edits geometry and texture against the input observations. Results show strong fidelity on synthetic objects, cluttered tabletops, and indoor scenes.
Chen et al. introduce WorldFact-Bench, a dataset of 1,274 real–fake image pairs spanning ten domains, each containing a factual conflict while preserving visual plausibility. They also propose PERSIST-Agent, an iterative verification system that maintains a persistent state linking candidate facts, observations, evidence, and verification statuses. On 8B backbone models, the agent raises pair accuracy by up to 10.4 percentage points over direct judgment, but overall performance remains modest, highlighting the gap between visual plausibility and factual correctness.
The authors introduce OpenProblemBench, a benchmark of 82 open problems drawn from the mathematics and theoretical physics literature. Each problem supplies research context, assumptions, and prior progress. Four evaluator models independently assess submissions for correctness and progress without reference solutions. Across seven evaluated configurations, GPT-6-Astra attains a mean judged solve rate of 14.0%, compared to 2.4–6.7% for other models.
Vishwanath and colleagues compared eight months of real-world use of an institutional language model assistant with 58 public clinical AI benchmarks. They found a substantial mismatch: clinicians mostly requested help with documentation, administration and knowledge retrieval, while benchmarks emphasized diagnosis and assumed questions could be answered from the information provided.
The authors introduce EpiReal-Bench, a benchmark of 10,000 false-claim prompts and generated images spanning 10 claim categories and 10 visual formats. They find that four leading commercial image generation models produce credible visual misinformation from over 70% of such prompts, and their proposed attack EpiReal-Attack pushes the success rate above 95%.
The authors introduce a verification-centric perspective for LLM-assisted peer review, focusing on error detection rather than mimicking human reviews. They present a scalable benchmark built by synthetically inserting logical contradictions into conference papers, and propose a Multi-Layered Review (MLR) framework that conducts multiple passes over the manuscript before generating feedback. Across evaluations, MLR shows strong alignment with human review scores, achieves high error detection performance, and offers complementary reviewer focus, though it remains vulnerable to adversarial manipulation.
The authors propose AVE, which first edits a sparse set of keyframes using a strong image editor, then generates the full video via motion-guided image-to-video diffusion, treating edited keyframes as anchors. On IVEBench and VIE-Bench, AVE achieves state-of-the-art performance in instruction following, temporal consistency, and content fidelity, with the quality strongly tied to the image editor used.
The authors test whether a pretrained language model can improve its visual understanding by mid-training on raw video frames with no captions or text loss. They encode frames as continuous visual tokens and train the model to predict the next visual token, then apply identical instruction tuning to this model and a baseline. The mid-trained model outperforms the baseline on video and image benchmarks while matching text performance.
The authors unify recent Muon optimizer variants as chained linear minimization oracles (LMOs), compositions of norm-ball projection steps such as the matrix sign. They show that many practical chains lack standard LMO guarantees and can diverge on convex objectives, yet in linear associative memory chaining provides a provable learning advantage. Their proposed TensorChain optimizer stacks compatible weight matrices across layers into a 3D tensor and normalizes across axes, achieving 9.6% average token savings over Muon at matched validation loss in Qwen3 0.6B and 1.7B pretraining.
The authors propose SPLIT-RL, a staged reinforcement learning method that trains a vision-language model's visual reasoning (VR) and language reasoning (LR) capabilities in separate phases. By using the VR output as a bottleneck and applying claim-level advantage computation, they achieve accuracy gains of 1.4 to 6.1 points over standard GRPO across five model families, while also improving both VR and LR individually.
The authors introduce CePU, a preference-driven unlearning method tailored to few-step distilled (FSD) text-to-image diffusion models. Instead of relying on noise-prediction errors, CePU uses consistency error as a reward signal, demonstrating strong forgetting of identities and NSFW content while preserving other capabilities, all in roughly 15 minutes on a single GPU.
The authors introduce Plan-and-Patch, a framework that uses a diffusion language model (dLLM) to generate structured plans and repair them by infilling affected regions while preserving surrounding steps. On Natural Plan benchmark without task-specific training, the diffusion planner achieved 53.7% repair success versus 27.0% for autoregressive planners, and after task-specific training on ALFWorld and TextCraft, diffusion reduced plan-generation latency by 39-46% while maintaining similar success.
Zhao et al. identify a source of wasted work in sparse attention: neighboring queries often select many of the same key-value (KV) entries, yet existing kernels process each query separately. Their QUILT execution mechanism exploits this overlap, cutting average sparse-attention kernel latency by up to 55.1% and time to first token by up to 36.8% in the reported experiments.
Benlalah et al. present GHARP, a system that builds a subject-specific 3D Gaussian head avatar from a few images, then animates it from facial-expression codes. On Ava-256, the authors report state-of-the-art visual quality while using eight times fewer Gaussians and running up to 13 times faster than FastGHA on an A100 GPU.
Lu and colleagues address a weakness of reinforcement learning with verifiable rewards: fixed training tasks eventually become either trivial or effectively impossible, leaving little useful feedback. Their VICO framework rewrites structured visual problems as the model learns, producing automatically relabeled tasks calibrated to remain near its current capabilities.
Bhandari et al. separate KV-cache storage precision from the number of bits transferred during each language-model decoding step. Across six base models, their progressive codec read four bits per scalar on average from an eight-bit cache while increasing C4 perplexity by no more than 0.66% relative to an uncompressed cache.