News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 1 of 54 · The Research front page
Yang and colleagues study a hidden obstacle to sharing robot training data: the same command can produce different movements under different controllers. Their approach uses examples of commanded actions and resulting motions to help one policy adapt both its movements and its task strategy, without additional training at deployment.
The authors train humanoid catching policies entirely in simulation using a single egocentric camera, then deploy them zero-shot on a Unitree G1. The robot successfully catches nine objects unseen in training, including plushes, semi-deformable toys, chip bags, and high-rebound rigid-elastic objects, while respecting wrist torque limits as low as 5 Nm.
Bianco et al. demonstrate that a quadrotor can learn to traverse a moving narrow gap end-to-end, from visual and inertial inputs to thrust and body-rate commands, using differentiable simulation without a teacher, planner, or distillation. In simulation, the policy reaches a 99.3% success rate after a few minutes of training on a single GPU, and it transfers directly to a real quadrotor with an 80% success rate over 10 physical trials.
The authors propose a spatially indexed memory module for reinforcement learning that allows a legged robot to store and recall task-relevant information at specific positions in the environment. Tested in simulated and real-world mazes, the system matches exploration coverage of standard recurrent policies while uniquely preserving the maze layout, cutting reconstruction error by a factor of three.
The authors introduce Trace-WAM, a robot learning approach that augments future-video prediction with auxiliary 3D motion prediction to focus on manipulation-relevant dynamics. This yields higher success rates on standard benchmarks and real-world tasks, with the largest gains under data scarcity and distribution shifts. The inference cost remains identical to prior methods that do not predict future motion.
The authors present TAC2, a foundation model that generates granular, timestamped audio captions with speaker attribution directly from in-the-wild audio. By fusing predictions from off-the-shelf sound event detectors, diarizers, and captioners via time-registered IoU matching, they avoid prohibitive manual annotation and produce structured acoustic timelines.
The authors present CARD, a retargeting framework that explicitly preserves contact force targets from human demonstrations, bypassing the need for object models or simulators. By converting data from video, tactile gloves, and motion capture into a unified representation with contact forces, CARD maintains contact on 43–78% of frames, far surpassing kinematic methods. On hardware, a tactile tracking controller raises open-loop task success from 13% to 83% over position-only replay.
The authors introduce Octen-WebGR, a generative retrieval framework that stores web page identifiers in a language model's parameters. After indexing 201 million URLs and training on past clicks, the model outperforms embedding-based retrievers on overall navigational traffic and especially on unseen queries for pages with rich interaction histories. The 4-billion-parameter model runs on a single GPU with 10 GB memory, in contrast with the 412 GB required for exact dense retrieval.
The authors adapt FLUX.2, a pretrained text-to-image diffusion transformer, to synthesize speech from text by treating spectrograms as images. With the image encoder frozen, the model preserves intelligibility and speaker identity, and image-pretrained initialization reduces word error rate from 4.2% to 2.7% in matched comparisons.
The authors present DexSuite, a simulation framework for dexterous manipulation built on Genesis. It includes 30 tasks, 10 arms, and 12 end-effectors, along with a benchmark of 16 task-embodiment pairs and a dataset of 4,700 successful demonstrations. Policies for dexterous hands matched or outperformed gripper policies on several tasks, but precise tool use and sequential manipulation remained challenging.
The authors present a method to learn runtime monitors that predict when a humanoid robot is about to fall unrecoverably. By actively exploring its state space via an uncertainty-guided reach-avoid strategy and combining temporal-difference with Monte Carlo learning, they achieve high coverage of recoverable states and zero-shot transfer to a real Unitree G1 humanoid.
Zhao and colleagues present a speech-to-speech translator trained to decide when incomplete input provides enough information to start translating aloud. On RealSI, adjusting that decision using speech already buffered for playback improved translation quality over a low fixed threshold, with little change in first-audio latency.
Search agents can retrieve documents that match an intermediate query but do not help answer the user's actual question. Sharifymoghaddam and colleagues trained a four-billion-parameter model to rank documents using that broader context, improving evidence quality across several benchmarks and raising answer accuracy over search without reranking.
Li and colleagues investigated whether robot demonstrations become more useful when humans must continuously correct a learned controller’s deviations. Their shared-control approach produced data that improved or matched the success of policies trained through either imitation or offline reinforcement learning across three real-world robotic tasks.
Zhao and colleagues show that human, robot and simulation videos can provide useful pretraining signals for robot manipulation without recorded control commands. Their approach learns a shared representation of motion before adapting to each robot’s controls, with strong reported performance across simulated benchmarks and six real-world tasks.
Zhao and colleagues trained a video world model to predict what happens after actions, using a shared motion representation rather than incompatible robot-specific commands. The authors report that its generated demonstrations improve task success over training on real demonstrations alone, with benefits on simulated benchmarks and two real-robot platforms.
Zhu and colleagues introduce RynnUnify, a model that combines understanding physical scenes with generating videos of what might happen next. An additional agent trained to choose when to request those predictions improved anticipatory reasoning by 6.0 points over direct reasoning on the authors’ benchmark.
The authors introduce propose-filter-release (PFR), a general method that transforms a privacy mechanism, which only protects records that fit the remaining data, into a standard differentially private mechanism. By privately filtering out records that are atypical, PFR allows noise to be calibrated to the typical sensitivity of the data rather than the worst-case influence of a single record. They demonstrate the approach on tasks including principal subspace estimation, KL divergence, and treatment-rule selection, showing significant reductions in required noise on typical data.
The authors introduce CATT, a synthetic benchmark for evaluating data attribution methods in language models, where ground truth is unambiguously known. They find that document redundancy and lexical distractors consistently degrade attribution performance, and that standard techniques wash out localized causal evidence by averaging across all tokens. Applying negative log-odds gradients and aggregating only the top token scores improves recall by 35 to 72 percentage points across multiple methods.
The authors introduce world2vec, a method for learning mechanistic world models from observations and actions without rewards. It faithfully captures dynamics over long prediction horizons with at least an order of magnitude fewer parameters than DreamerV3, TD-MPC2, and LeWM, and achieves state-of-the-art control in continuous environments from pixels.
The authors introduce Assess to Decide to Respond, a defense where a chatbot assesses user intent, commits to an explicit safety label, and then generates a reply conditioned on it. Trained with a self-verified reward or an LLM-judge reward, this approach achieves the strongest multi-turn safety in a controlled comparison across two model families, preserving utility at a modest cost to helpfulness. Analysis confirms the decision span is a causal control point.
Reinforcement learning from verifiable rewards (RLVR) struggles on hard problems because unsuccessful rollouts yield no learning signal. The authors show that combining two forms of teacher guidance, trajectory-level intervention (correcting student paths) and token-level supervision (dense feedback on each step), dramatically improves performance. On hard logic games, PIVOT nearly doubles mean avg@8 from 19.5% (best baseline) to 35.6%.
Rezaeian et al. propose a trajectory-level evaluation protocol for tool-using AI agents that measures whether every prefix of a tool-call sequence remains executable under tool schemas, state preconditions, and data dependencies. Applying it across multiple benchmarks, they find that outcome-only rankings can diverge sharply from prefix-level reliability: unverified plan-conditioned agents often achieve higher final success while substantially lowering executability metrics, whereas pre-execution plan checking improves executable-prefix survival without sacrificing terminal performance.
The authors introduce generalized Feynman-Kac correctors (GFKC), a method for sampling from distributions defined by combining several pretrained diffusion models, with optional reward-based steering. The algorithm operates entirely in the log domain and corrects generation on the fly, requiring no retraining. Experiments on Gaussian models and MNIST confirm accurate sampling from non-trivial combinations and effective steering.
Ekbote and colleagues study language models that repeatedly apply the same layers, adding computational depth without storing a separate set of weights for every step. They find that the best looping design changes with the resource budget, and that additional repetitions eventually stop helping or can make training unstable.
An AI agent can miss evidence that another agent has already retrieved simply because their workflow does not connect them. Gong and colleagues introduce SPLICE, which adds a targeted connection and examines the resulting changes before deciding whether to keep the new answer. Across five benchmarks, the authors report improvements over the original fixed workflows without changing the agents themselves.
Peng Xie and Amr Alanwar tested whether language models grade items independently when many items are placed in one prompt. They found that changing which items share a batch flips 11-22% of pass/fail verdicts on harm detection and short-answer grading, far more than changing instruction wording or item keys. The instability is concentrated on ambiguous items and persists even when averaging over multiple batch orders.
Zhang et al. introduce DynaRev, a benchmark of 181 web tasks where user goals change at semantically defined milestones. Testing five models, they find that agents often complete only part of the revised task. Rewriting instructions while preserving browser state improved quality and efficiency, while browser resets degraded performance.
The authors present TextSeal2, a distortion-free watermark for large language models that combines two or more watermark keys into one via a Gaussian copula, avoiding the signal loss of its predecessor TextSeal. The scheme adds no latency, preserves output diversity, and maintains the acceptance rate of speculative decoding. At equal token diversity, it detects 8 to 22 percentage points more 100-token responses than TextSeal at a false-positive rate of 10⁻³.
The authors propose Multi-View MoE, a modification to mixture-of-experts language models that uses several independently initialized routers to score experts and reuses their relative preferences to mix learned input views for the selected experts. On a 10-billion-parameter model trained on 708 billion tokens, the full method raises the average over nine downstream tasks by 1.18 percentage points over the baseline, while preserving the active-expert count, expert FLOPs, and communication volume.
The authors show that the conventional sequence of training stages for AI models, from pretraining through fine-tuning and reinforcement learning, can be expressed as a single maximum-likelihood algorithm. Their unified trainer, which mixes static datasets with model-generated data in one continuous process, improves accuracy and task coverage over staged training at the same compute budget, while preserving broader capabilities.
The authors built TOME-Math, a corpus that converts 2,118 university-level mathematics textbooks into roughly 380,000 proof-oriented question-answer pairs while preserving chapter order, proposition-proof links, and dependency references. Fine-tuning a 30B language model on this corpus lifted its mean score across six mathematical benchmarks from 29.17% to 45.83%, a gain of 16.7 percentage points, and beat three public instruction-data baselines in aggregate. Controlled experiments show that the structural signals, not just the extra data, drive much of the gain.
The authors investigate how AI agents track the roles of messages in conversations. They find that the language models inside these agents rely on two sets of cues, content (writing style) and structure (chat formatting tokens), to infer who said what, and that these mechanisms are easily exploited. Role spoofing attacks that combine fake content and structure can hijack any model tested, with over 95% success impersonating the user in agent environments.
Zhu and colleagues study whether AI agents can solve extended tasks more effectively by deciding what to remember and when to clear their accumulated history. Their approach preserves selected progress notes across context resets, improving reported benchmark performance while sharply reducing the amount of text the model processes.
The authors present Wizard, an agentic game world model that uses structured proxies such as skeletons for characters and geometric maps for scenes to guide video generation. An agentic framework maintains explicit game states and updates these proxies in response to player inputs, enabling real-time, fine-grained control over character motions. The system generalizes zero-shot to unseen scenes, characters, and actions, resulting in consistent and visually coherent gameplay across diverse scenarios.
Gourevitch and colleagues show that discrete diffusion models, which generate data represented by separate symbols such as tokens, admit a deterministic counterpart to their usual sampling process. Their construction reproduces the original model’s distributions over time, while providing trajectories that support training models to generate samples in fewer steps.
Sun and colleagues study how to check whether a computer-using AI agent actually completed an instruction, rather than merely attempted it or claimed success. Their verifier gathers additional evidence after the agent finishes, improving agreement with reference labels and rejecting more false completion claims than a vision-language model judge in the online study.
Lu and colleagues evaluate whether AI models can distinguish how many times something appears in an image from how many physical objects those appearances represent. On their PhysCount benchmark, counting appearances is substantially easier than linking them to physical identities, and even the strongest model answers every related question correctly in only 13.36% of groups.
Ke and colleagues examine a blind spot in reinforcement learning that rewards language models for verifiably correct answers: the update does not account for how confidently the model produced each response. Their confidence-aware adjustment improved 14 of 16 model-and-training combinations and increased the share of problems solved within 32 attempts.
Schmidgall and colleagues trained a language model through simulated psychiatric conversations, then compared it with its identically prompted, untrained counterpart. The trained model scored higher in 76.7% of matched encounters with psychiatrists role-playing patients, while separate safety tests showed fewer failures to recognize the urgency of acute risks.
Randomly changing a language model’s weights can improve reasoning when selected variants vote on an answer. Dufour and colleagues reproduce these gains but find that changing weights is unnecessary: adding random biases to a frozen model’s output scores often works just as well, challenging the explanation that the procedure uncovers nearby task specialists.
Zamir and colleagues present a reinforcement-learning algorithm whose expected regret, the reward lost while learning rather than acting with full knowledge, matches the best possible worst-case rate up to a constant. For sufficiently many episodes, their guarantee removes the extra logarithmic factors that remained in earlier results.
Xu and colleagues present a training-free way to accelerate video diffusion transformers, which generate video through repeated denoising steps. On Wan2.1 and HunyuanVideo, they report end-to-end speedups of 1.84–2.23 times on a single H20 graphics processor, with better fidelity to full-attention outputs than competing sparse-attention methods.
The authors profiled compute time and energy consumption for LLM inference across many input and output token lengths. They found that output length scales linearly with cost, but input length has a quadratic component, making long contexts disproportionately expensive. They also show that compute time is a reliable proxy for energy, enabling estimates for proprietary APIs without direct measurements.
Li, You and Huang investigate whether simpler adversarial training can compress modern text-to-image models into generators that need only one evaluation. Their method beats the original multi-step models on a statistical image-quality measure, using a single training objective and reportedly a fraction of the training compute required by distribution-matching methods.
The authors show that transforming well-specified skill operations into executable programs reduces reliance on LLM reasoning and improves task performance. Their framework, SkillCompiler, uses a compilation strategy to identify compilable operations and a reuse mechanism to avoid redundant program synthesis. Across extensive experiments, it consistently outperformed state-of-the-art methods, achieving up to a 32.33% improvement.
The authors introduce ATLAS-Bench, a benchmark of 1,200 local search queries that distinguish mandatory from optional and objective from subjective requirements, along with a reproducible sandbox called ATLASBox. Testing 14 configurations of 11 LLMs, they find a trade-off between constraint satisfaction and query coverage: the highest-accuracy agent (GPT-6-Luna) satisfied 71.5% of constraints in its answers but found an acceptable option for only 67% of queries, while a model with lower per-answer accuracy (GLM-5.3) succeeded on 85% of queries.
The authors evaluated Google's AI Overviews (AIOs) for 3,611 health questions, comparing their sources to traditional search results and assessing statement grounding and accuracy against physician-written rubrics. They found that AIOs cite only 44.4% of the top five traditional results, overrepresent commercial publishers like WebMD and Healthline at the expense of academic and government sources. An estimated 56.7% of cited statements were supported by their source, while 20.8% of all statements were factual claims with no citation; overall, AIOs scored 38.2% of available points on physician rubrics, and 22.2% of AIOs where a harmful claim was anticipated contained that content.