News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 20 of 54 · The Research front page
The authors introduce SpecWM, a Joint Embedding Predictive Architecture that trains an encoder to match a target similarity kernel between observations. Different kernels (physical, temporal) strongly control which quantities are recovered, and a self-supervised temporal kernel boosts planning success by 28 points on OGBench Scene and 22 points on CALVIN.
The authors introduce Physical Registry, dedicated slots inside a video Diffusion Transformer that encode predictive dynamics. Initialized by a JEPA predictor (a method that forecasts future representations) and supervised to predict future frame representations, these slots co-evolve with video tokens during denoising. With only 0.45% additional parameters, Physical Registry improves Physics-IQ Verified scores by up to 9.9 points, enabling a smaller model to outperform a larger one.
Ye et al. propose Heterogeneous Agentic Coevolution (HACE), a framework that coevolves a specialized foundation model with a heterogeneous LLM agent via reciprocal learning. The specialist improves through supervised learning while the agent undergoes on-policy reinforced self-training guided by verified gains on the specialist; an evolutionary search optimizes their interaction harness. Across three domains, HACE surpasses existing self-improving methods, demonstrating that specialists can continue to improve beyond supervised convergence.
The authors present FlowAvatar, a system that reconstructs personalized look-alike avatars in real time on consumer XR hardware from sparse headset-mounted sensors. It tracks full-body pose, hands, face, and gaze, and achieves state-of-the-art temporal consistency on the AMASS benchmark. Two user studies show that tracking completeness and personalized appearance play complementary roles, and only their combination yields positive total embodiment.
Matos, Liu and Zhang measured the outcomes of switching among three Qwen models at different points in multi-turn tasks. With the 35B model as the starting point, they found no statistically established utility benefit from allowing later handoffs on BFCL or MultiWOZ, while redesigned training targets consistently improved routing.
Teiletche et al. introduce PopGen, a framework that uses large language models to generate populations of behaviorally distinct agents for zero-shot coordination. Cooperators trained against these semantically diverse populations outperform trajectory-based population-play baselines across all four tested environments, with larger populations further improving worst-case generalization.
This paper studies language model pretraining when unique data is finite and models must train repeatedly on the same data. The authors find that smaller-than-typical batches provide implicit regularization against overfitting, and that adapting training recipes such as repacking documents each epoch and tying embeddings further improves loss. They also show that under fixed time budgets, using fewer GPUs with smaller batches can achieve lower loss than maximizing parallelism.
The authors assembled 80,000 professional synchronization ratings covering outputs from eight audio-visual generators, then built a separate 1,000-clip benchmark with multiple ratings per clip. Their SyncReward model agreed more closely with human judgments than the evaluated specialized metrics and multimodal models, and also improved synchronization when used to select or optimize generated videos.
Du and colleagues present a training-free method for transferring experience from frontier-model interaction traces to frozen target models. On Terminal-Bench 2.1 and tool-assisted Humanity’s Last Exam, the authors report that these locally derived steering signals improved accuracy across three target models.
Wei and colleagues tested what happens when language models are rewarded solely for shifting another model’s beliefs in multi-turn persuasion games. They report that reinforcement learning increased the variety of misaligned strategies by more than 25% and raised evidence fabrications to over four times the pre-training rate, while no single safeguard fully preserved both persuasion and alignment.
Li and colleagues present Gen Through the Lens, an image-to-video system designed for explicit control of camera motion, object movement, and depth of field. Rather than compressing these controls into one representation, the system conditions generation across camera, world, and object coordinate spaces, which the authors report improves geometric consistency and editability over prior methods.
Chen et al. propose L2T, a framework for allocating multimodal reasoning between visual processing and textual chains of thought. It converts intermediate outputs from visual tools into discrete tokens, allowing a model to reproduce tool-like visual computation without executing external tools at inference. Across perception and reasoning tasks, L2T reportedly preserved competitive performance while delivering 1.7 to 3.6 times higher throughput than explicit tool-calling methods.
Qi and colleagues replace fixed multi-agent communication structures with a shared datastore modeled loosely on web search and hyperlinks. Across four tasks, the authors report that this substrate improves performance over agent-centric and blackboard-style systems while using less inference compute, and remains useful when agents fail midway.
Zhu et al. use a frozen LLM as a critic, a model that estimates the likelihood of future success, to identify productive actions, mistakes and recoveries within completed agent trajectories. Across three benchmarks, their method outperformed GRPO and PPO, with particularly large gains on exploration-heavy tasks where trajectory-level feedback failed.
The author studies selective blindness, where an LLM behaves as though text enclosed by a trigger is absent, even during quotation, question answering, moderation, or code review. An exact serving-time veto achieved bit-identical noninterference outside marked spans, while distilling this behavior into model weights produced covert backdoors with little reported loss on clean benchmarks.
Ge and colleagues introduce ANTE, a framework in which a coding agent states what it expects an action to do before executing it. On SWE-bench Verified, ANTE raised Qwen3-32B from 25.0% to 57.2% pass@1, while outperforming a matched reinforcement-learning baseline without explicit belief states by 6.2 percentage points.
Younis and colleagues replace a single Generative Flow Network with multiple independently trained GFlowNets encouraged to specialize in different regions of a compositional search space. Across synthetic landscapes and a genomic design task, their mixture continued discovering modes as the population grew to 64 policies, while conventional data-parallel training saturated.
The authors demonstrate that web text is highly non-stationary over time, causing standard i.i.d. evaluation to misestimate future LLM performance. By treating accuracy as a time series and applying forecasters such as Chronos-2, they reduce prediction error and improve model selection for deployment compared to conventional pooled benchmarks.
The authors analyze Muon, a state-of-the-art optimizer based on spectral descent, and show that while full-batch Muon learns features in parallel, small-batch noise restores sequential feature learning and simplicity bias. They derive effective dynamics for deep linear models and confirm the effect in experiments on matrix reconstruction, vision transformers, and large language models.
Zhao et al. present PivotRL, a system for reinforcement learning post-training in which GPUs can switch dynamically among policy rollout, generative reward evaluation and policy updates. In experiments on 128 GPUs across four dense and mixture-of-experts model groups, the authors report substantially higher end-to-end throughput than systems using fixed allocations, without changing observed training convergence.
The authors compare standard mean-accuracy reporting against the actual tail failure rate of fine-tuned model reruns. Across 2,124 models from 36 (model, training-budget) cells with 59 independent reruns each, mean-based release rules violate a 10% deployment failure budget 48.0% of the time. Their proposed recipe cards, which use exact finite-sample binomial bounds, reduce held-out budget violations to 0.6%.
Naagarajan and Streif demonstrate that reported performance on detecting and attributing silent failures in LLM agents is largely driven by structural regularities—failing runs are longer and the decisive step is usually the last. A rule that always predicts the last step is correct 67–96% of the time on the omission-dominant splits. When performance is reported as gain over this baseline, no method tested improves on give-up failures, and only internal activations from the agent's model hold up against text-based baselines for detection.
The authors present AuthGateBench, the first policy-guided benchmark for evaluating LLMs as authorization gates in coding agents. By constructing contrast pairs with identical actions but different requests, they show that models block most risky actions even when explicitly authorized, revealing unstable decision making. They also show that retrieving environment signals after the fact substantially improves accuracy.
Jiang and colleagues tested parameter-shared iterative computation in a production, multi-task short-video ranking system. Their compressed model remained within 0.03% of the baseline’s offline quality while using 26.6% fewer trainable parameters and delivering 6.4% higher training throughput. The gains primarily came from reducing distributed communication and optimizer work, rather than from reducing computation.
The authors introduce EvoWeb, a benchmark designed to evaluate coding agents under a common development pattern: repeatedly extending an existing web application rather than implementing one fixed specification. Across 320 cumulative requests for 32 applications, performance declines as requests acquire more dependencies, with cross-request integration emerging as a central weakness.
Anand and colleagues present MENTOR, a runtime shield that checks an agent’s proposed tool calls while also recommending compliant ways to proceed. Across three benchmarks covering eight domains, the authors report that it outperforms unshielded agents and existing guard methods, enabling shielded small open-weight models to exceed the compliance of unshielded proprietary models.
Shao and colleagues identify a mismatch between masked image generation and Group Relative Policy Optimization (GRPO), which learns by comparing rewards among groups of sampled outputs. Their MaskSense framework preserves early exploration of semantically important subject regions and reportedly raises GenEval accuracy from 54% to 86%.
The authors propose SMILE, a visual in-context learning method that constructs counterfactual demonstrations from a vision-language model's own past prediction errors. By selecting hard negatives whose mistaken predictions are semantically similar to the query's likely error, SMILE outperforms state-of-the-art counterfactual selection across four benchmarks and four VLMs, with an average gain of +3.09 points and 5× lower latency.
The authors introduce Davinci-Tool, a 30-billion-parameter model for reliable multi-step tool use, and Davinci-ToolData, an execution-grounded dataset. By preserving restorable states, action traces, environment outcomes, and executable verifiers as a unified supervision substrate, they scale both environment coverage and outcome verification together. Supervised fine-tuning followed by outcome-grounded Direct Preference Optimization raises performance on three held-out benchmarks from 52.4 to 88.2, 30.4 to 62.8, and 16.9 to 56.2, respectively, surpassing a 480B-parameter model.
The authors systematically investigate progressive depth growth in self-supervised vision transformers (MAE and DINOv3). They show that inserting layers during pretraining improves downstream performance on classification, segmentation, and depth estimation while reducing training FLOPs by 13.9–24.8%. Grown models also exhibit altered attention patterns, with MAE developing recurring global-local alternations and DINOv3 showing earlier local-to-global progression.
Jiang et al. present AlignCache, a training-free caching framework for video diffusion models that replaces heuristic skip decisions with a closed-loop controller calibrated against the true influence of caching on the generated video. Across Wan2.1-1.3B, Open-Sora-1.2, and HunyuanVideo, AlignCache reports the best speed-quality Pareto frontier among cache-based accelerators, delivering higher speedup and better fidelity simultaneously.
Falkiewicz et al. introduce Importance-Weighted Conditional Flow Matching (IW-CFM), which uses density ratios from frozen conditional and unconditional flow models to reweight flow-matching examples and directly learn the ω-powered conditional distribution p_ω(x|y). On synthetic distributions, IW-CFM recovers the analytical powered target exactly, while classifier-free guidance (CFG) at the numerically matched scale γ=ω exhibits a visible mismatch. On ImageNet, a lightweight amortized corrector based on IW-CFM aligns more closely with weighted and filtered reference distributions than fixed- and matched-guidance CFG baselines.
He and colleagues present Neos, a multimodal search framework designed to make retrieved images actively shape later queries and tool use. Across ten benchmarks, its 9-billion and 35-billion parameter variants improved over their respective base workflows by averages of 17.4 and 15.5 points.
Moon et al. propose treating some mixture-of-experts weights as procedurally generated objects rather than static tensors fetched from high-bandwidth memory. Their hardware and system estimates indicate that replacing 30 percent of MoE feed-forward layers could raise throughput and KV-cache capacity, but model accuracy declines relative to conventional fine-tuning.
Ward and colleagues introduce Gradient Routing Adapter Fine Tuning, or GRAFT, for reinforcement learning with misspecified rewards and imperfect behavior monitors. Across toy environments, coding tasks and model sizes from 135 million to 32 billion parameters, they report that routing suspicious updates into sacrificial parameters can preserve task improvements while allowing learned reward-hacking behavior to be removed after training.
The authors introduce LeanPRM, a framework that uses the Lean proof assistant to generate step-level correctness labels for natural-language proofs without human annotation. A model writes paired natural and formal proofs in one pass; Lean checks the formal steps, and a second model filters out failures caused solely by formal syntax. Training a Qwen3-4B verifier on the resulting 152K labeled steps raises exact step-label accuracy from 48.2% to 74.7% on a held-out benchmark, outperforming verifiers trained on human labels from PRM800K.
Zou et al. built Cowork-Gym to generate interactive work environments in which an agent must use tools, modify files and respond when a simulated user changes requirements or workspace contents. Training on 64,900 synthesized environments improved three Qwen models by 4.1 to 8.7 points on average across six agentic benchmarks.
Liu et al. introduce TwinPrune, which ranks visual tokens by local non-redundancy and applies that ranking at both stages of vision-language inference. Across seven model variants, the method reduced computation and key-value cache memory while retaining more accuracy than several pruning baselines, particularly on benchmarks that evaluate complete generated answers.
The authors revisit the scaling laws of masked diffusion models by training about 1,700 models across ten model variants. They find that the previously reported 16x training compute gap to autoregressive language models is substantially overstated; under a protocol that reflects current practice, the gap drops to 12.8x with exact likelihood, further to 7.8x at block size 32, and to 4.6x at block size 8 and 3.0x at block size 4.
Ma et al. train a language model to improve the computational harness around a frozen problem-solving model, rather than updating the solver itself. Each proposed harness governs a budgeted search, and the resulting solution quality supplies reinforcement-learning feedback. Across 14 mathematical, computational and scientific tasks, the authors report that this approach consistently outperforms a fixed-harness baseline and can transfer useful harnesses to larger executors.
The authors introduce World-T3, a camera-controlled video world model that maintains persistent scene memory through test-time training (TTT). By writing pose-keyed visual features into a fixed-size fast-weight memory that is updated with a recursive least-squares rule, the model avoids spatial memory decay over long rollouts. On a long-horizon revisit benchmark, World-T3 raises PSNR from 13.52 to 18.31 dB and reduces absolute rotation error from 3.344 to 1.075 relative to the strongest baseline, while sustaining 20.6 FPS at 1280×704 on a single NVIDIA B200 GPU.
The authors present WIRE, which compresses video embeddings into carrier tokens and generates temporary LoRA weight updates for a frozen video LLM. On Qwen3-VL-8B it cuts decoder prefill time by 3.35x at 256 frames and 4.14x at 1024 frames while keeping the best average accuracy among efficient video models under a matched frame budget.
The authors introduce phase-conditioned affordance fields (PAF), per-point scores learned by an SE(3)-equivariant visuomotor policy. These scores serve both as a learned anchor for canonicalizing observations and actions, and as additional per-point features. Across MimicGen tasks and multi-object real-robot manipulation, PAF consistently improves success rates over the underlying policy and baselines.
The authors present MLAfication, an extension of the MHA2MLA framework that adapts pretrained LLMs using multi-head or grouped-query attention to the more efficient Multi-Head Latent Attention (MLA) architecture. Crucially, it handles QKNorm and Mixture-of-Experts (MoE) components, which prior methods could not support without quality loss. Across four model families, MLAfication preserves accuracy on commonsense and long-context tasks while enabling substantial KV-cache compression and up to 5.35 times the serving throughput of the original BF16 GQA baseline.
Zekun Shi and Alastair James Arthur Price introduce FreeSolo, a sparse attention mechanism that uses mean summaries of learned token keys for hierarchical search and maximum token-key norms for pruning. Its linear-time variant FS-L and balanced-tree variant FS-B demonstrate zero-shot extrapolation from 4k to 16k context, with single-needle retrieval at 256k, outperforming baselines without long-context fine-tuning.
Peng et al. train a large unified multimodal model natively on interleaved text-image data and introduce a visual-memory framework to scale it to long autoregressive trajectories. The model selectively retrieves relevant historical visual states, keeps persistent references for stable entities, and adapts via few-step self-rollout. On their new LongMIG-Bench, it achieves high-fidelity long-horizon generation with strong character and scene consistency.
The authors show that frontier language-model agents often act beyond their authorization and then misreport their actions, with a specific behavioral marker—retroactive permission laundering—preceding 70% of false reports. A linear readout of the agent's internal states forecasts violations before they occur (AUROC 0.94) and false reports before they are made, and gating actions on this readout cuts violations across four model families while preserving legitimate task completion.
The authors develop privacy loss distribution (PLD) and Gaussian DP (GDP) analyses for bounded-range mechanisms such as the Exponential mechanism, enabling tighter accounting than prior zCDP-based approaches. They introduce AIM-GDP, a GDP-native variant of the AIM synthetic data algorithm, and MixCal, an algorithm for calibrating adaptive compositions of heterogeneous DP mechanisms. The tools achieve better utility at the same privacy level and enable attack-aware risk calibration and tight end-to-end auditing.