News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 27 of 54 · The Research front page
Biyani and Dvijotham introduce SpecGuard, a system that detects and formally certifies conflicts between a coding task's natural-language description and its tests. On conflicted SWE-bench tasks, SpecGuard detects up to 72.8% of conflicts and formally certifies up to 51.1%, with a nearly fivefold lower conflict miss rate than model-based judgment, providing a pre-execution safety check against reward hacking.
Yi et al. introduce CanonQ, a quantization-aware training framework designed for simultaneous compression of multiple parts of a large language model. In the most demanding W2A4KV2 setting, meaning 2-bit weights, 4-bit activations and 2-bit key-value caches, the authors report substantially better perplexity and zero-shot accuracy than the evaluated quantization baselines.
Zhang et al. present CARE, a model-agnostic pipeline for deciding when an AI prediction can be accepted and when a human should intervene. In binary-classification experiments spanning driving, language, vision and robotics, CARE maintained human-aligned risk control while requiring 25 to 81 percent fewer human queries than the evaluated baselines.
Hu and Tang study the winner’s curse in LLM self-improvement: selecting the best-looking rewrite from several candidates whose measured performance contains noise. Across instruction-rewriting loops and established prompt optimizers, they find that selection scores systematically overstate real gains, especially when the same small evaluation set is reused.
Kong et al. present FreeEvolve, a framework in which an agent evolver decides what candidate harnesses to test, how much evidence to gather, where to spend resources and when to stop. Across four agent benchmarks, it improved the primary held-out metric by an average of 13.6 percentage points and matched or exceeded hand-designed evolvers on six of seven reported metrics.
Dang et al. trained a humanoid controller for walking and manipulating while supporting sustained upper-body loads, which alter the robot’s center of mass rather than acting as brief disturbances. In simulation and experiments on a Unitree G1, their HULK framework improved balance and command tracking under load, including tests with 10 kg applied at each wrist.
Hu et al. developed a two-stage system that learns humanoid loco-manipulation from human demonstrations recorded without a robot. On a Unitree G1, the system performed box sorting, catching and suitcase manipulation, while simulated box sorting succeeded in 77% of standard trials and 64% of trials with external pushes.
Wertheimer and colleagues introduce a transformer attention layer whose learned decay mechanism doubles as an input-dependent rule for pruning the key-value cache used during generation. In experiments with a 1-billion-parameter model trained on 50 billion tokens, the authors report more than 10-fold cache compression without degrading downstream performance, rising to 25-fold at a 16,000-token context length.
Kumar and colleagues developed SEA-LM, an audio-language model designed for egocentric, multichannel recordings from smart glasses. On the authors’ evaluation set, it localized sounds more accurately than tested baselines, improved most transcription tasks and operated across 1,211 microphone layouts without configuration-specific retraining.
The authors train over one hundred physiological waveform foundation models (20M to 2.1B parameters) on up to 36.3M hours of data and derive a scaling law that predicts downstream clinical performance. They find that compute-optimal allocation scales pretraining hours much faster than model size, and that larger models use data more efficiently, reducing both waveform and labeling requirements for a given performance target.
Chai and Xiong study language-model pretraining when a limited corpus must be reused for multiple epochs. They find that the penalty from repetition is largely predicted by extra epochs divided by unique tokens per parameter, while the optimal number of epochs depends strongly on whether the unique dataset grows with model size.
Swain and Dutta show that jailbreak attacks can succeed by suppressing safety-critical features in large language models, rather than only by activating harmful ones. They introduce Counterfactual Activation Potential (CAP), a metric that quantifies a suppressed feature's latent tendency to activate, and a filtering algorithm that discovers such features without exhaustive ablation.
Derakhshani et al. introduce LiFT, a looped generative model where each recurrent step is trained toward a distinct point on a straight path from the initial estimate to the flow-matching target. This allows the model to loop far beyond its training depth at inference, improving generation quality without retraining or parameter increase. On ImageNet 256x256, LiFT-L/2 achieves an FID 3.34 points lower than a dense DiT-XL/2 baseline while using about 60% fewer parameters and 52% fewer inference FLOPs.
Lee and Kim introduce directional Fréchet distance, which measures how much each feature-space direction contributes to the scalar distance between generated and reference distributions. Across image, video and protein case studies, they show that a small number of interpretable directions can dominate the metric and clarify results that the aggregate score obscures.
Khorrambakht and colleagues introduce FLEX-WAM, a joint generative model of states and actions designed for efficient, open-loop simulation over long horizons. By combining block-causal attention with training that explicitly preserves sensitivity to commanded actions, the model produces stable rollouts for thousands of steps and supports planning entirely through imagined trajectories.
Fenoglio and colleagues introduce Distributed Subliminal Learning (DSL), in which participants probe locally adapted models with random, task-unrelated inputs and send the resulting outputs instead of parameters. Across language-model composition and federated image classification, DSL preserved useful local behavior while reducing communication, although its effectiveness depended on the model interface and carrier coverage.
Liu and colleagues evaluate whether research agents can autonomously improve a common base model on image, audio, video, joint audio-video and image-grounded software tasks. Across six agent configurations, more than half of model-task averages fell below the base model, and agents often failed to preserve gains or submit their strongest evaluated candidate.
The authors study Online Agentic Test-Time Training (OaTTT), where a deployed LLM agent executes each task once and uses the single verified trajectory as the only learning signal. They identify that directly imitating or reinforcing these trajectories destabilizes the policy. They introduce ASCENT, which self-distills verified experience by using a frozen copy of the LLM as a teacher, conditioned on the successful trajectory as privileged information, to produce stable next-token distributions that are distilled into persistent LoRA weight updates. Across ALFWorld, WebShop, and AppWorld, ASCENT improves task success and interaction efficiency as experience accumulates, outperforming online adaptation methods and transferring to held-out scenes.
Hu and colleagues introduce Red-TTT, an automated red-teaming method that updates an attacker’s parameters while it attempts to jailbreak a language model. On HarmBench, this behavior-specific adaptation raised average attack success from 55.9 percent to 72.4 percent under the same 120-sample attacker-generation budget as a frozen Best-of-N baseline.
Habib et al. present PixReenact, a head-avatar system that animates a reference portrait from a live driving video without relying on keypoints, 3D face models, or other specialized motion representations. The causal diffusion model generates frames incrementally, and its distilled student emits four frames per update with a reported mean latency of 239 ms while maintaining identity over long streams.
Xie et al. introduce MechHypoBench, the first benchmark designed to assess whether AI agents can generate mechanistic hypotheses from empirical data. By pairing paper-derived mechanisms with real datasets containing 17.98 million records, the authors find that current AI scientists fall short: the gap between generated hypotheses and true underlying mechanisms remains substantial.
Bruner and colleagues formulate semantic camera motion control as a video-to-video task in which the model must infer how to move the camera relative to an evolving subject. On a 109-scene benchmark, SemCam executed the requested behavior in 68.6% of cases, compared with 45.3% for the strongest evaluated baseline, while preserving subject identity at a comparable level.
Cao et al. separate routine search judgments from the slower reasoning and generation performed by a full language model. Their specialized decision model produced faster, better-calibrated judgments on a new benchmark, while a dual-system agent improved both search speed and answer accuracy on BrowseComp-Plus.
On Tai Tang studies online reinforcement learning for multimodal deep-research agents at unprecedented scale: 128k context and 75+ tool-interaction turns. Directly extending conventional RL to this regime proves inefficient due to weak early exploration and premature entropy collapse. The authors propose Long-MDR, a training recipe combining On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue, which stabilizes learning and yields a 9B-parameter agent that ranks first on five of six benchmarks at a 50-turn evaluation budget.
The authors propose RollVerify, a lightweight framework that augments partial-rollout reinforcement learning for large language models. By using a new off-policy shift metric (OPS) to detect stale trajectory segments, RollVerify performs sequence- and token-level verification to truncate and regenerate off-policy suffixes. In experiments on mathematical reasoning, it achieves accuracy comparable to fully on-policy training while reducing GPU training cost from 98.1 to 57.5 GPU days.
Yang and colleagues introduce Bag-of-Tokens Group Relative Policy Optimization, a modification of GRPO that uses token-level process rewards without learning a value network. Across React code generation and AIME mathematics, the method reached target performance in fewer training steps than GRPO and several variants, although most experiments used only one random seed.
Liu et al. introduce a benchmark of 432 prompts requiring image generators to render exact text across realistic documents, signs, packaging and interfaces. Tests of 24 model configurations show that legibility can improve while textual accuracy declines, and that performance falls sharply as the amount and distribution of requested text increase.
Zhang and colleagues study whether vision-language models direct visual evidence toward the answer candidate that the evidence actually supports, rather than merely producing the correct final answer. Their RIVET interface improved both candidate-level effect transfer and task accuracy compared with matched baselines, showing that evidence utility and evidence ownership are distinct properties.
Zheng et al. present a reconstruction-free method for separating individual hidden factors in learned world representations. Their theoretical result shows that DSReg can recover each latent up to sign and ordering, provided an existing representation identifies the latent state up to a linear transformation and the latents satisfy a new Structural Diversity condition.
Choudhary et al. formalize two ways hostile interface content can subvert computer-use agents: changing the intended action and redirecting its execution to the wrong interface target. Their Secure-CUA design addresses both while achieving 53.55% task success on WebArena, 1.57 percentage points below an unsecured baseline.
Fan et al. designed CoMoE for Mixture-of-Experts models running across commodity GPUs connected only through PCIe and host memory. On RTX 5090 systems, the authors report up to 1.46× higher inference throughput than SGLang, reaching 85 to 90 percent of a stronger A800 configuration while requiring at most 23.4 percent of its equivalent GPU investment.
Park et al. design a transformer that can execute Lloyd’s k-means algorithm with embedding dimension d + ceil(log2 k), rather than the previously demonstrated d + k, where d is the data dimension and k the number of clusters. They also train such models across distributions of clustering tasks and analyze how architecture and smoothing of the discrete objective affect convergence and in-distribution generalization.
The authors propose CIPO, a framework that enables LLM agents to adaptively select between atomic tools and composite skills. By mining frequent tool subsequences as executable skills and using counterfactual branch rollouts to compare outcomes of granularity choices, CIPO improves task success and decision efficiency across multiple benchmarks.
Chen and Han introduce EDISCO, the first discrete diffusion model for Euclidean combinatorial optimization problems (ECOPs) that guarantees exact E(2)-invariance of the generative distribution over solutions. The model consistently outperforms prior learning-based solvers on synthetic TSP and CVRP benchmarks, using fewer training instances, and generalizes better under spatial distribution and constraint shifts.
Ye et al. demonstrate that watermarking large language model outputs can cause factual hallucinations even when the correct information is present in the input context and the unwatermarked model produces accurate answers. Using a retrieval-augmented generation setup, they quantify an average factual accuracy drop of up to 10% across six watermarking methods and attribute the effect to token-level perturbations and cumulative attention drift. Two plug-in interventions reduce the factual error by about 90% at a detection rate of 0.90 true positive rate at 1% false positive rate.
The authors introduce mRewriter-R1, an agentic framework that treats multilingual query rewriting as a multi-turn sequential decision process. Trained with reinforcement learning, it dynamically chooses among language, structural, and semantic rewriting operators and consistently outperforms static rewriting baselines on three multilingual reasoning benchmarks.
The authors introduce ExStereo, a stereo module that adds 3D perception to pre-trained 2D Vision-Language-Action (VLA) models. By reconstructing metric depth from stereo images, rendering multi-view observations, and using an attention mechanism to condition actions on 3D features, ExStereo significantly boosts performance across simulation and real-world tasks, outperforming RGB, point-cloud, and implicit-stereo baselines.
Khafizov et al. replace iterative diffusion denoising with next-scale autoregression for zero-shot novel-view synthesis from a few posed images. On three object datasets, NAMVIS produced more accurate target views than the evaluated diffusion baselines and generated a 256 by 256 view in 0.6 seconds, about 3.3 times faster than the fastest baseline.
Baroudi, Bredin and Marxer introduce SepRQ, a self-supervised speech model designed specifically for mixtures containing multiple speakers. Instead of reconstructing masked audio, it learns to separate pseudo-sources represented by discrete targets from frozen random-projection codebooks, while training at several temporal resolutions. The authors report leading results on multi-speaker SUPERB and TS-SUPERB tasks, including difficult three-speaker separation, with 85.68 million inference parameters.
The authors benchmarked two neural surrogate architectures, Transolver and BSMS-MGN, as end-to-end approximations of the final-time particle concentration for the two-dimensional linear radiation transport equation on canonical Lattice and Hohlraum benchmarks. They found that both surrogates generally outperform coarse conventional solvers, but that architectural choices such as Fourier features and region-weighted loss must be tuned per model family for optimal accuracy.
Tiwari, Chordia and A.P. combine on-policy distillation with Group Relative Policy Optimization (GRPO), a reinforcement-learning method that compares rewards among multiple sampled solutions. Their DiffGate objective uses correctness feedback to target teacher guidance at failed solutions, especially when every rollout fails, improving the chance that Qwen3 students produce at least one correct answer across eight attempts.
Mukhopadhyay and Nair test whether AI agents can conceal attacker-selected information inside benign-looking responses, store it in persistent memory, and retrieve it during a later session. They report exact recovery in 20.1% of trials, with 14.4% of all trials both completing the primary task and passing independent safety oversight.
Chen and colleagues formulate inference configuration as constrained multi-objective black-box optimization, seeking Pareto-efficient settings that balance model quality against memory or computation. On Qwen2.5-7B and DeepSeek-V2-Lite, InferOpt found layer-specific configurations that outperformed uniform allocations and three general-purpose search baselines under matched resource budgets.
Liu and Han held language models fixed while comparing Claude Code, mini-SWE-agent and OpenCode on SWE-bench Verified, a benchmark of real repository repair tasks. They report little reproducible evidence that changing harnesses improves task completion, but find large and systematic differences in token consumption and cost.
Qi et al. evaluate five existing mitigation methods for tool hallucination in LLM agents across multiple tool configurations and discover hallucination escape: these methods reduce errors in their tuned setting but increase them elsewhere, cancelling gains. They trace this to intrinsic tool-use tendencies that models hold, which existing methods reinforce. The authors propose EscapeGuard, a training-free method that detects conflicts between model tendencies and the current configuration and redirects attention to the runtime tool set, reducing tool-selection hallucination by 9.0 percentage points and lowering cross-configuration mean hallucination by 23.7 pp.
Yang et al. examine whether commands emitted by an LLM agent reach the underlying program with their intended meaning intact. Across production shell calls and a purpose-built benchmark, they find that wrappers, interpreters and other intermediate hops frequently alter calls, causing failures that conventional trajectory analysis often blames on the model.
Wang and colleagues introduce QiYao-I, a foundation model for forecasting multivariate time series whose measurements are unevenly spaced and not synchronized across variables. The authors report that it outperforms both general time-series foundation models and specialized irregular forecasting systems, including in zero-shot and few-shot evaluations, although the supplied text does not include numerical results.
The authors introduce Clean, a second-order optimizer that uses randomized Nyström sketching to approximate the preconditioners in SOAP, reducing memory complexity from quadratic to linear in model dimensions. They also propose Q-Clean, a quantized variant. Clean matches SOAP's performance while using less memory than AdamW, converges 26% faster than AdamW, and enables pre-training a 13B parameter model on a single 80GB GPU.