News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 12 of 54 · The Research front page
The authors introduce ArabPref, the first large-scale behavioral preference dataset covering all 22 Arab League countries across 200 cultural aspects. They show that post-training multilingual and Arabic-centric language models on this data improves cultural alignment in generation and reasoning tasks compared to baseline models.
The authors propose Gold-Conditioned Distillation (GCD), a self-distillation method for supervised fine-tuning of large language models. By using the same model as both teacher (conditioned on the gold response) and student (conditioned on the prompt only) and training via forward KL divergence on gold tokens, GCD achieves better accuracy on new tasks compared to standard supervised fine-tuning, while generally reducing performance degradation on general benchmarks.
Van Velthoven et al. evaluate four probabilistic AI weather models against the physics-based IFS-ENS ensemble for forecasting record-breaking temperatures and wind over land in 2022. The authors find that AI models outperform IFS-ENS for heat and cold records at nearly all lead times and record depths, with the advantage strengthening for deeper heat records. However, all systems show reduced skill for wind and as lead time increases.
The authors introduce noise-rejuvenated sequential Monte Carlo (SMC), an inference-time scaling method for text-to-image flow matching models. By perturbing resampled particles via source noise replacement, the method enables exploration without converting the deterministic ODE sampler to an SDE. Experiments on three state-of-the-art models show consistent improvements in reward-guided generation across multiple benchmarks.
Schaffer, Manor and Michaeli test whether decoding individual vision transformer patch tokens reliably reveals where a model represents a class. They find that this apparent interpretability is not guaranteed by the transformer architecture: it varies with the task, data distribution and training objective, even when models classify accurately.
The authors introduce MOSAIC, a multi-objective search method that learns per-layer key-value cache budgets for LLM inference. On two models and four eviction methods, MOSAIC beats uniform allocation in 146 of 154 (task, method, budget) settings, achieving gains of up to +48 points on RULER and +11.9 on LongBench. It reaches the same score as uniform allocation with up to 4x less cache.
The authors present GeoART, a training-free self-refinement method for camera-conditioned video diffusion. By generating multiple candidates at selected denoising steps and evaluating them with a geometry prior, GeoART improves camera controllability and multi-view consistency without modifying the pretrained model. Rank-calibrated probabilistic selection, rather than greedy choice, balances score magnitude and relative rank to handle noisy geometry scores.
The authors introduce Map2Vid, a task that generates egocentric navigation videos conditioned on bird's-eye-view (BEV) maps, offering a more intuitive alternative to text prompts or camera trajectories. They built a dataset of 117,800 navigation sequences across 61 indoor scenes, each paired with videos, BEV maps, trajectories, and language instructions. A curriculum training strategy gradually weakens non-map inputs, achieving robust performance with BEV maps alone.
The authors propose SteadyThought (ST), a thought-level preference optimization framework to address 'under-thinking' in large reasoning models, where models frequently switch reasoning paths even when a promising path is found. ST identifies switching points using linguistic and entropy signals, generates single-path continuations as preferred responses, and applies preference optimization to discourage unnecessary switching. The key finding: ST reduces token consumption by up to 63.8% while maintaining strong reasoning accuracy across multiple benchmarks.
You et al. show that Euclidean distance can mislead planners because learned latent representations may curve and fold within their higher-dimensional space. Replacing this distance with shortest paths through a graph of neighboring training latents substantially improved LeWM’s planning, particularly when goals were far from the starting state.
Asgari et al. decompose LLM judge strictness and find that rubric content, not the judge model, determines pass rates. They introduce a closed-form Shapley attribution for all-must-pass rubrics, revealing that model self-review deletes unique-catching checks and adds spurious requirements. A combination of attribution-based gates and provenance freeze prevents quality drift in revision chains.
Wang and colleagues show that fine-tuning data need not contain harmful content to erode an aligned language model’s refusal behavior. Across three aligned models and three safety benchmarks, benign examples increased compliance with harmful requests when they created a sufficiently large, broadly learned shift from refusing to answering.
Singh et al. test whether vision-language models (VLMs) form truly multimodal text representations that aid tasks with latent visual structure. They construct a shape-counting task that is trivial with images but difficult from text coordinates. They find that VLMs perform no better on the text version than their language-only counterparts, but a self-distillation method that aligns text embeddings with image embeddings yields substantial gains.
Eitan, Yehudai, and Maron analyze the expressive power of efficient hybrid sequence models, which pair a full-attention layer with cheaper recurrent or sparse layers. They prove that existing designs, including sparse top-κ access to the attention cache, need Ω(√n) capacity for two-function composition. Their proposed Multi-Hop Retrieval Attention (MuRA) block computes compositions of m functions at fixed size with m hops of retrieval, establishing a strict expressivity hierarchy.
Jebraeeli et al. exploit an industrial accident where roughly 53,000 prompts were labeled under two different safety policies (moderator block vs. harmful content). Training ten guardrail models at two scales, they find that benchmark scores are heavily influenced by the labeling policy: HarmBench-copyright varies by 0.99 across arms, and over-refusal benchmarks vary by 53 percentage points. They also show that training corpora often contain benchmark prompts (e.g., 88% of JailbreakBench-benign is in training data), and validation accuracy tracks policy agreement (rho=+0.74) rather than detection capability (rho=+0.34).
Tan et al. introduce MemBubble, a benchmark of 1000 query-history pairs built from real user-assistant conversations, to measure how agent memory affects response diversity. Across six LLMs and five memory mechanisms, they find that memory processing reduces response entropy by 34.7% on average compared with providing the full interaction history. Simple user-side interventions fail to close this diversity gap.
The authors trained smaller language-model agents by on-policy distillation from stronger teachers and found a counterintuitive pattern: students often exceed the teacher mid-training, then regress. They attribute this to competing effects of decision repair and erosion, and introduce SPOD, which preserves successful student choices from the teacher's conflicting supervision. SPOD consistently beat ordinary distillation (up to 15.2 percentage points) and even surpassed the teacher by up to 20 points.
The authors introduce Conjugate Compilation, a method that reframes speculative decoding as a cross-model compilation problem by jointly optimizing the small draft model and the larger target model within their shared execution on a single GPU. A self-improving agentic system couples an LLM with Monte Carlo tree search, proposing and evaluating cross-model plans based on actual co-located GPU performance, achieving speedups up to 1.59x and improving GPU utilization by up to 1.50x.
The authors show that a non-looped pretrained LLM can be converted into a recurrent-depth model through post-training upcycling. Using layer-wise looping with learned gates on residual branches, they break the common saturation at two loops. Starting from Qwen3-4B, UpLoop improves average accuracy across 20 benchmarks by 6.17% at three loops.
The authors systematically evaluate when additional acceptance length in speculative decoding is worth the computational cost. By remeasuring released checkpoints and retraining six drafter recipes under shared conditions on two target models, they find that published method gaps shrink substantially. They identify two routes to improve acceptance—nearly free context/objective changes and more expensive token-to-token conditioning—and show that throughput gains depend on task entropy and serving concurrency.
The authors introduce WorldStateMedia, an open dataset of 118 million state media articles in 114 languages, built by linking Fineweb web crawl data to outlet-level state control annotations. Auditing these corpora, they find editorially controlled state media has grown sharply in English crawls and is overrepresented in low-resource languages. A retrieval-augmented generation experiment shows that retrieving country-specific state media biases LLM answers toward that country's territorial claims.
Arab, Wu and Bai introduce ReDT, a feed-forward transformer for reconstructing 3D scenes from multiple high-resolution images. Across nine datasets spanning standard definition to FHD inputs, the authors report better reconstruction accuracy and substantially lower computational costs than existing visual geometry transformers, including a 16× speedup on ETH3D at 2K resolution.
Wang and colleagues introduce BiDex-100, a MuJoCo benchmark for controlling two multi-fingered robotic hands across 106 manipulation tasks. The suite pairs every task with a controlled out-of-domain variant and uses progress-based scoring to diagnose generalization, instruction following, low-data learning, long-horizon control and memory.
Tian and colleagues developed RARE, a reinforcement-learning method that updates rubric criteria as a model’s capabilities change. On three video-script types, its best sample-level variant improved average evaluation scores by 3.82 points over training with fixed per-query rubrics, while largely preserving performance on earlier requirements.
Luo and colleagues propose EG-GRPO, a reinforcement-learning method that turns evidence annotations in synthetic search data into rewards for individual retrieval steps. An LLM judge links each required claim to the first citation that supports it, allowing the method to track how an agent's knowledge develops during a search. Across five deep-search benchmarks, the authors report generally stronger performance than standard GRPO and several reward-enhanced or step-level alternatives.
The authors revisit REPA-E, a prior approach for improving latent diffusability, and find that removing the diffusion objective while retaining semantic prediction from corrupted latents gives comparable results. This motivates a general principle—robust prediction—which they instantiate in RoAR-VAE using two complementary objectives: RobustAlign (predicting pretrained semantic features) and RobustRecon (reconstructing the clean image). RoAR-VAE improves average generation benchmark performance by 2.19 points over the strongest baseline and boosts small-text reconstruction accuracy by 16.93 percentage points over FLUX.1-VAE.
Zheng et al. distinguish between unlikely errors produced during sampling and physically implausible behavior that a video model nevertheless considers likely. They use this decomposition to build HA-Sampler, which adds external physical guidance during latent refinement, and HA-Scorer, which estimates model likelihood to rank generated videos. Tests on Physics-IQ and VideoPhy show improvements over the compared sampling and scoring baselines, although the supplied text does not report effect sizes.
The authors identify a Combination Gap: directly attaching existing adaptive reuse methods to a compressed diffusion transformer degrades sample quality (FID from 5.56 to 10.40–13.18 on ImageNet 256×256). They propose TamedFlow, which trains a compact DiT-B/2 student via cross-architecture feature alignment, residual velocity correction, and randomized rollout supervision, then deploys a 4M-parameter Learnable Trajectory Estimator to accept reliable student drafts or trigger teacher rescue. On ImageNet 256×256, TamedFlow achieves FID 2.58 at 2.25× measured speedup over the DDIM-50 teacher, and outperforms training-free adaptive baselines at comparable latency.
The authors introduce KV-Surgeon, a compensation-aware method for offline KV-cache compaction that adjusts remaining entries after each deletion. KV-Surgeon achieves state-of-the-art accuracy on question-answering and code retrieval tasks, outperforming prior methods by up to 4 percentage points at 50x compaction.
The authors present UtilityBench, a 15,000-instance benchmark that assigns sample-specific costs to errors using Best-Worst Scaling, and UtilityAUC, a metric that measures how well model confidence aligns with the utility of outcomes. Evaluating 20 state-of-the-art language models, they find that greater accuracy and more advanced architectures do not reliably translate to higher UtilityAUC, revealing a gap between problem-solving ability and responsible decision-making.
The authors present Maestro, an agentic framework for Lean proof development that replaces local repair of failed subgoals with global coordination. Workers provide structured failure reports, and a single coordinator decides on appropriate responses, such as guiding a stuck prover or replanning a flawed subtree. Maestro proves 34 out of 48 theorems from the June 2026 arXivLean release with Claude Opus 5, compared to 19 for the native agent, and outperforms all published provers on PutnamBench, Lean-IMO-Bench, and IMO 2026.
Guanda Li and colleagues propose interventional step credit (ISC), a causal measure of how much each chain-of-thought step influences the final answer, and use it to train models via interventional guidance policy optimization (IGPO). On two model scales across nine benchmarks, IGPO reduced average response length by up to 62% while improving macro-average accuracy by up to 8.4 percentage points over base models.
The authors propose ActionRoPE, a parameter-free interface that converts action-driven camera displacement into world-coordinate rotary positions for planar-camera game world models. This approach reduces revisit drift by 66% over the strongest baseline and generalizes across model sizes and out-of-distribution scenes, solving a core mismatch between learned feature conditioning and the geometric structure of top-down and isometric games.
The authors introduce EvalRx, an open-source agentic framework that discovers systematic failure modes in foundation models, formulates falsifiable hypotheses, and verifies them before applying targeted repairs. In evaluations across 12 model family-modality combinations spanning Gemini, Qwen, Gemma, and Nemotron, EvalRx improved performance by +1.4 to +10.9 percentage points, with an average gain of +5.8 points.
Xue et al. systematically evaluate 37 public security skills for coding agents (Claude Code with four LLMs) on 100 real Linux kernel vulnerabilities. Across 14,800 runs, only 11.1% of detections were fully correct (vulnerable function, statement, and class). The authors propose SKILL-SIF, a self-improving framework that iteratively refines skills from missed vulnerabilities, improving fully correct detection from 2% to 16% on unseen vulnerabilities.
Zheng, Li, Deng and colleagues target a persistent weakness of video generators: smoke, fire and explosion debris often look detailed but move without coherent turbulence. Their system combines supervised training on Houdini simulations and real-world footage with preference optimization guided by predicted fluid velocity, producing more structured vortices, advection and fine-scale motion than the tested alternatives.
The authors propose PoSTCache, a training-free framework that makes diffusion caching postdictive: predictions for skipped steps are first forecast, then revised using discrepancies observed at later full-computation steps. They also introduce Posterior Variance Reduction (PVR), an anchor selection strategy based on offline-profiled cross-timestep dependencies. On representative image and video generation models, PoSTCache improves reconstruction fidelity over existing caching methods at comparable or higher acceleration.
Han and colleagues introduce DeepSEA, a professionally translated benchmark covering mathematical and scientific reasoning, question answering, summarization and machine translation in 12 Southeast Asian languages. Their evaluation finds that leading frontier systems remain comparatively robust, while smaller models deteriorate sharply on low-resource languages and often perform better when using English as an intermediate language.
The authors present CUA-Lite, an open platform that standardises the components needed to develop computer use agents (CUAs)—agents that control desktop, browser, and mobile applications. By defining three abstractions (a unified environment interface, a unified supervised data format, and per-model harnesses shared across evaluation, RL, and SFT), the platform integrates 12 benchmarks, provides lightweight VM-free desktop sandboxes with over 30,000 verifiable tasks, and converts major public CUA datasets. The authors release reference scores for 15 models and show that supervised fine-tuning and reinforcement learning through the same harness improve open-weight models.
The authors identify Procedural Exposure Bias (PEB): increasing the exposure frequency of an action pattern in a skill shifts a language model's action preference even when state and relevance are unchanged. Rendering the same procedural content as an image substantially attenuates this shift, and visually represented skills improve task success across agent benchmarks.
The authors introduce FinHopper, a benchmark for LLM-based agents on complex financial queries requiring 1–6 reasoning hops, each conditioned on the exact answer from the previous hop. In tests with 20 agent configurations, performance dropped sharply with hop depth, and process-aware scoring revealed partial reasoning progress that final-answer-only metrics would miss.
The authors propose Embodied World Agents (EWAs), agents that perceive and act across diverse open-world 3D environments, and release EmbodiedWorld-200K, a dataset of over 200K video samples with multi-level annotations. They demonstrate that offline pretraining on this data improves embodied planning, and that further RL post-training using a video world model as a scalable environment proxy yields clear gains on out-of-distribution scenes.
The authors introduce LEAPS, a collection of 12 novel datasets designed to evaluate long-context language models on state tracking and pattern learning, capabilities often overlooked by existing recall-focused benchmarks. Testing 30 API and open models at context lengths from 8K to 512K tokens, they find that models scoring near-perfect on recall tasks vary dramatically on these other abilities and degrade significantly as input length increases.
The authors test whether nonparametric self-evolving LLM agents, which adapt by rewriting prompts, memory, or code without updating model weights, can function as continual learners. Across four drifting task streams built from BIRD, BFCL, and ScienceWorld, they find that self-evolution improves aggregate performance on most streams but systematically sacrifices stability for plasticity, and that both self-modification and similarity-based retrieval suffer from interference or overwriting when task distributions shift.
He and colleagues introduce Swift, a training method that uses a frozen, undertrained denoiser to modify the regression targets seen by a stronger model. Across CIFAR-10 and ImageNet-256 experiments, the resulting models generally produced better samples than conventionally trained counterparts, without requiring the weak model or additional computation during inference.
Wang and Chen show that reinforcement learning with verifiable rewards fails in recognizable, localized ways: a small set of phrases and operations strongly predicts whether a reasoning rollout will end incorrectly. They use those diagnostic n-grams to retrieve just 35 million tokens from existing training corpora, then continue pretraining before rerunning reinforcement learning, outperforming equally sized random-data controls in 10 of 11 settings.
The authors introduce Spherical Fixed-Point Reasoners (SFPRM), which perform a damped fixed-point loop on the hypersphere of normalized transformers. On Sudoku-Extreme, Maze-Hard, and Mini-ARC, SFPRM achieves test accuracies of 96.6%, 87.9%, and 59.1%, outperforming Fixed-Point Reasoners (FPRM) and other baselines while halting up to 2.7 times earlier at equal or higher accuracy. The angular geometry also provides interpretable representations: token-input angles encode puzzle structure and predict answer correctness.
Sun, Wang and Li address a weakness in Group Relative Policy Optimization for video: rollouts normally see identical frames, so they may all succeed or fail and provide little comparative learning signal. Their FrameGRPO method varies frame density within each rollout group and selectively rewards successful lower-frame trajectories when every full-frame reference rollout fails. The authors report consistent gains over standard GRPO across seven video benchmarks.