News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 22 of 54 · The Research front page
The authors present RoboSimStudio, a toolkit built on the Newton physics engine that makes solver, contact model, and physical parameters explicit and editable. They demonstrate 18 contact-rich manipulation simulations and a pilot study in which coding agents built a bimanual sponge-wiping simulation from a smartphone video in about one hour, then transferred an improved policy to a real robot.
Wang et al. introduce SEAR, a framework and benchmark to evaluate whether LM-driven robotic agents genuinely learn from experience or merely benefit from repeated attempts. Across seven frontier LMs and over four years of agent runtime, they find that learning is real but conditional: harder tasks benefit from longer sessions, while easier tasks do not; and code-based skills transfer across tasks and simulators, but task-specific solutions often fail to generalize to novel scenes.
Yan and colleagues present Steertuning, a robot-learning system that adapts to a new task and environment using five egocentric human demonstrations, without task-specific robot demonstrations. Across 12 tasks in 12 previously unseen environments, the resulting policies completed 71 percent of trials successfully, with collection, training and deployment taking less than five minutes.
The authors propose Geometry Attention Distillation (GAD), a training-time method that distills cross-frame correspondence targets from a frozen geometry model into video diffusion attention. Across three video backbones, GAD improves long-range scene consistency and dynamic temporal consistency, including at least 30% relative gains in cross-frame matching accuracy on Wan2.2 5B, without adding inference-time modules.
The authors propose ProAS, a plug-and-play framework for critic-free reinforcement learning of LLM agents. By leveraging an Agentic Reference Graph to annotate action validity and task progression, ProAS applies Masking Unreasonable Incentives and Asymmetric Penalty Redistribution to reshape trajectory-level advantages. On ALFWorld and WebShop, ProAS boosts success rates by up to 16.7 points and accelerates training convergence up to 2.44x across multiple model scales and baselines.
The authors introduce a competition built on the Smart Buildings Control Suite, in which participants train HVAC control policies on procedurally generated simulated buildings and are evaluated on held-out buildings and climates. The goal is to determine whether a building foundation model can transfer across buildings with minimal reconfiguration. The authors argue that even a 5% reduction in building emissions could cut total US emissions by roughly 1%.
Sun et al. introduce CityCode, an agent-operable 3D city representation, and Air2Code, a pipeline that constructs it from a single low-altitude aerial RGB image per region plus known camera parameters. The work demonstrates that recovered metric entities can serve as a stable interface connecting visual evidence, 3D assets, and long-horizon spatial operations, with all 2,500 steps across 50 task chains passing their documented task checks.
Park and Rusch analyze looped reasoning as convergence in a learned dynamical landscape and identify a mechanism: a slowly decaying oscillatory mode can impose its convergence rate on coupled latent dynamics, delaying settlement near solution roots. They propose QRR (Quotient Recurrent Reasoner), a gauge-equivariant architecture in which per-token phase rotations leave the reasoning-relevant quotient state unchanged by construction. QRR converges to solution roots in quotient space faster and more stably than prior recurrent reasoners while matching or exceeding their accuracy on Sudoku, Maze and Mini-ARC.
Abdelrahman and colleagues test how a vision-language model should communicate a language-specified region to a segmentation decoder. In Sa2VA, they find that the model’s spatial attention contains substantially more localization information than the conventional single [SEG] token embedding, and using that attention as a dense prompt improves segmentation on two benchmarks.
Jahagirdar and colleagues systematically test how evaluation settings affect video-language model scores, finding swings of up to 19.5 percentage points for the same checkpoint. They introduce Meta-Video, a curated 1,000-question benchmark designed to preserve model rankings while reducing the compute required for broad evaluation.
Hu et al. address a central obstacle in training long-horizon agents: difficult tasks often provide no useful reward unless the entire workflow succeeds. They introduce Tev-Gym, spanning seven technical domains, and report that dense rewards derived from known intermediate states improve 9B and 27B models, with gains transferring to separate terminal benchmarks.
Kim and colleagues present ACCRUE, an off-policy reinforcement learning framework for continually adapting robot foundation models to new tasks. In simulated and physical assembly sequences, carrying learned policy and value information forward reduced the interaction needed for later tasks, while robots retained approximately 97% average success on prior real-world tasks.
Zhan and colleagues show that reinforcement learning based only on unit-test outcomes can improve code correctness while encouraging larger patches and invented bugs. Their PreciseCoder training method rewards both passing tests and changing the specific lines that require repair, reducing average patches to about three lines across models ranging from 4B to 27B parameters.
Chen and Levine study how flow-matching models pair source noise with target data, a choice that determines the velocity targets used for training. They introduce a scalable approximation for larger online optimal transport couplings and prove, under their theoretical framework, that such couplings reduce stochastic-gradient variance and numerical integration error compared with independent pairing.
The authors present SynAct, an active learning framework that allocates expert feedback to individual decision points rather than whole trajectories. Across four benchmarks and four models, it improves held-out accuracy in all 16 evaluated settings, by 11.7 percentage points on average.
The authors present Elastic Latent Recurrence (ELR), a recipe for training looped language models that avoids the one-loop tax—the performance gap at a single loop compared to a non-recurrent model of the same size. At one loop, ELR models match or exceed their non-recurrent baselines; each additional loop then buys further accuracy. On 4B dense and 30B-A3B MoE models, four loops add 7.25 and 4.75 points respectively to the one-loop ten-task average.
Li et al. build a vision-language reward model with no human preference labels: a linear head on an intermediate layer of a frozen Qwen3-VL backbone, trained to rank web captions above perturbed versions. At 32B it beats GPT-4o by 6.6 points on VL-RewardBench and is level with it on two other benchmarks, at about one thousandth of the compute of fine-tuning. The comparison distribution, not benchmark rank, determines what the reward actually rewards.
The authors propose Kronecker Product Attention (KPA), a structured attention mechanism that caches keys and values as compact Kronecker product factors. Trained from 124M to 1.5B parameters, KPA matches or beats prior KV compression methods while shrinking the cache by 6.4x to 10x. It also introduces a Kronecker-separable rotary embedding that preserves relative positions without a decoupled RoPE branch.
The authors introduce EvoSyn, a multi-agent framework in which the space of controllable interaction factors for synthetic data is treated as a revisable structure driven by task-agent feedback. On τ²-Bench and CAR-Bench, EvoSyn improves average task success by 13.3 points over the strongest synthesis baseline while generating broader and higher-quality synthetic experience.
The authors introduce Behavioral Self-Rewarding (BSR), which reuses the current policy to compare trajectories sharing the same test outcome, producing auxiliary rewards for reinforcement learning with verifiable rewards. On SWE-bench Verified, SWE-bench-Live Lite, and Terminal-Bench 2.1, BSR improves average accuracy for Qwen3-14B and Qwen3-32B over verifier-only RL, without any new environment interactions.
Lu et al. argue that reinforcement learning with verifiable rewards depends less on dataset size than on whether problems produce a useful mix of successful and failed attempts. Their MagicBeanStalk method converts hard seed problems into new tasks matched to a model and rollout budget, substantially outperforming training on a 78,000-problem math corpus.
The authors propose CopRoute, a dependence-aware conformal routing family that builds model sets sequentially by ordering models on how many new queries each answers beyond those already selected, then calibrates a cutoff to meet a target all-miss risk. On RouterBench and RouterEval, CopRoute achieves the fewest certified calls on nine of ten datasets at target risk 0.05 and eight of ten at 0.10, with up to 79 percent fewer calls (median 24 percent) compared to RACER.
Jingru Yang and colleagues propose CAST, a shift-or-stay method for cross-tokenizer on-policy distillation. CAST solves credit alignment by distributing each aligned chunk's advantage according to student predictive uncertainty and local teacher-signal smoothness, while penalizing drift at tokens without teacher correspondence. On six benchmarks, it raises average scores over the strongest baseline by 2.45 points for Qwen3-8B to Mistral-7B transfer and 2.10 under shared tokenization.
Zhihao Liu and colleagues at Zhejiang University, Ant Group, the Institute of Science Tokyo and KAUST propose DP-MIND, a differential privacy (DP) inference framework that scales to contexts of up to a million tokens. The method processes long inputs in locally perturbed chunks while a memory carries task-relevant information across the sequence, and the authors report that it consistently outperforms existing DP and long-context baselines in both utility and scalability.
Sun et al. separate LLM workflow creation from query-specific selection. Their system first evolves a reusable set of workflows, then learns which workflow to apply to each query, outperforming evaluated baselines across five benchmarks while substantially reducing training token use on MATH.
Ji and colleagues connect empowerment, an agent’s capacity to control future outcomes, with unsupervised skill learning. They use skill-learning methods to estimate how many distinct behaviors remain executable from each state, then train a separate policy to reach states with more options. In complex environments, including Crafter, this objective produced behaviors such as collecting resources, using tools and building shelter.
The authors propose that what creativity benchmarks share is the evaluation of task performance in the long tail of a reference input-output distribution, where inputs or outputs are improbable. Analyzing 15 benchmarks and 26 open-weight models, they show creativity benchmarks sit further into the tail than standard benchmarks, and rankings derived from tail subsets align more closely with human creativity judgments in LM-Arena. They introduce q@Tail(α), a lower-cost diagnostic that measures mean quality on the most novel fraction of a benchmark.
The authors propose a lightweight velocity correction framework for few-step text-to-image generators. By training a small corrector to add residuals to the generator's predicted velocities at selected early steps, they improve instruction following on complex prompts by 17.6 percentage points in pairwise NetWin, with only 0.84% added latency and no change to the frozen generator.
Guo et al. propose T-Mem, a memory framework for LLM agents that uses temporal locality and temporal evolution to organize long-term multi-party conversations. On GroupMemBench and EverMemBench, T-Mem achieves 63.09% and 79.46% accuracy with GPT-5.5, outperforming competitive memory baselines across LLMs and query types.
Kordes and Dorent address cross-modal alignment when paired data are scarce but pretrained unimodal encoders and unpaired samples are plentiful. Their method builds a shared coordinate system from a few paired anchors, computes a global optimal-transport matching that suppresses overused nearest-neighbor hubs, then distills that matching into compact encoders for new data.
The authors introduce FG-Sudoku, a graded benchmark of visual Sudoku puzzles, and AdaC-GRPO, a curriculum reinforcement learning method that adaptively samples tasks within the model's Zone of Proximal Development. Applied to Qwen2.5-VL-7B, AdaC-GRPO raises accuracy on the hardest 9 by 9 tier from near zero to 0.219, approaching Gemini-2.5-Flash while closing 62.2% of the overall gap to Gemini-2.5-Pro.
Huang et al. address evidence fabrication, where search agents answer from internal knowledge while presenting their reasoning as grounded in retrieved information. Across seven knowledge-intensive benchmarks, SOAR improved average performance over reinforcement-learning baselines and reduced the reported fabrication rate by 74.81% relative.
The authors introduce TQS, a structural scoring framework that evaluates coding-agent trajectory quality from event logs alone, without LLM judges or reward models. Applying TQS to 155,144 successful trajectories, they show that supervised fine-tuning on top-ranked trajectories consistently outperforms random or bottom-ranked subsets on SWE-Bench-Verified.
The authors introduce A4-MR, a 3D vision-language foundation model for MRI trained on approximately 50,000 studies. It learns anatomy-level alignment and abnormality-aware representations, achieving gains over existing models on detection, segmentation, and localization tasks.
Kasimbeg et al. evaluated 13 optimizer submissions under AlgoPerf V1’s self-tuning rules, which prohibit workload-specific configurations. Schedule-Free AdamW v2 ranked first, while the broader analysis showed that strong progress per training step does not necessarily translate into faster real-world training.
The authors report on ByteExpress, a software communication layer over hardware RDMA deployed in ByteDance's parameter-server inference and training systems. In production, it cuts mean and P99 parameter fetch latency by 35.7% and 43.7% versus kernel TCP, P99 push latency by up to 74.0% in training, and total CPU usage by 10.69%.
The authors present EvoBelief, a self-evolving world model that represents an LLM agent's environment knowledge as falsifiable hypotheses. During deployment, it proposes, tests, and validates beliefs about environment dynamics, then uses only validated beliefs to guide predictions and plans. On three benchmarks, this approach significantly outperformed baselines, achieving on ALFWorld success rate improvements of up to 31.3% and completing scientific discovery tasks that none of the compared methods could finish.
Zhang et al. study why on-policy distillation becomes unreliable when language-model agents must reason and act over many turns. Their State-OPD method removes prior student reasoning from the context shown to the teacher, then directs additional sampling toward uncertain states, raising average task success for two Qwen3 student models across three interactive benchmarks.
Cai and colleagues recast proximal policy optimization for language models as a choice along two axes: how policy change is measured and how that change is estimated from samples. They derive a generalized trust-region guarantee and report that their best configuration, KL divergence with Atomic estimation, improves average math and coding scores by 50% and 14%, respectively, over the standard formulation used in their experiments.
Nguyen et al. present VEPA, a runtime guardrail that learns new safety and security rules from deployment feedback while keeping enforcement deterministic and auditable. Across three agent benchmarks, it achieved the strongest reported F1 on events covered by its active policy, including under noisy feedback, shifting risk types and longer deployments.
Cai and colleagues identify a weakness in long-horizon LLM reinforcement learning: small policy changes at each token can compound, causing the updated model to encounter states poorly represented by the policy that generated its training data. Their Trajectory-Aware Policy Optimization method explicitly estimates and constrains this accumulated mismatch, outperforming established alternatives across eight reasoning benchmarks.
Draguns et al. construct backdoored neural networks in which a separately trained capability is encrypted inside a host model's weights and executes only when a secret trigger is present. Using a cryptographic primitive called a digital locker, they show that recovering the trigger or identifying the embedded capability is computationally hard even with full white-box access. In proof-of-concept tests on a chemical reaction prediction task, standard elicitation methods failed to activate the hidden model.
Liou et al. introduce ReplaySSM, a decoding method for state-space models that unrolls the recurrence into a checkpoint and a bounded buffer of recent inputs. Integrated into vLLM for Mamba-2 and Gated DeltaNet, it accelerates standard decoding by 1.20 to 1.47x at serving batch size 256 and delivers about 1.8x the throughput of vLLM's speculative decoding on two 120B-class models.
The authors introduce OpenSocialJax, a benchmark extending Clean Up and Commons Harvest with procedurally generated layouts and hidden rules from tens of millions of configurations. Evaluating eight large language models in homogeneous and mixed teams, they find that while models often discover hidden rules and improve after initial trials, rule discovery does not reliably lead to sustained cooperation: cleaning becomes concentrated among fewer agents and shared resources are depleted.
The authors define Squint Text: text embedded in an image's coarse structure that humans can read by squinting but multimodal LLMs (MLLMs) often report as absent. They introduce SquintBench (1076 production moderation images) and SquintCtrl (synthetic suite) for diagnosis, identifying spatial selection and high-frequency interference as key bottlenecks. Their training method, Squint OPSD, uses a frozen teacher with a downscaled view and ground-truth transcription to supervise a student on the original image, achieving 88.61 ANLS versus 26.50 for the base model without requiring auxiliary views at inference.
The authors present Oído, an inference engine that runs open-vocabulary English speech recognition entirely on an ESP32-S3 microcontroller, a $5 chip with no neural accelerator. Using int8 arithmetic, it achieves 3.70%/8.23% word error rate on LibriSpeech test-clean/test-other, within 0.12 points of full precision, and outperforms laptop-class models such as Whisper tiny.en. Measured on physical hardware, the firmware processes audio at 1.97x real time.
The authors introduce ProgramGym, a framework for training coding agents that reproduce a program's observable behavior from documentation and a reference executable, without source code. Its environment pipeline builds verifiable black-box testing tasks across C, C++, Go and Rust, and its agent workflow coordinates analysis, implementation and review over long horizons. Fine-tuning on the resulting trajectories lifts Qwen3.8-27B from 34.87% to 42.84% on ProgramBench.
Artisan generates articulated 3D objects from text or reference images by delegating design, coding, and validation to independent agents, with feedback from a physics simulator. Across a benchmark covering appearance, physical plausibility, and downstream robotic tasks, Artisan achieves the lowest collision rates and highest stability rates among evaluated methods while keeping spatial fidelity competitive.