News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 25 of 54 · The Research front page
Liu et al. present FreeMatching, a single dense-correspondence model designed to identify matching pixels or regions even when image generation or editing breaks smooth motion and rigid geometry. The authors report substantial improvements on image-editing and reference-guided-generation pairs, while maintaining competitive results on classical optical-flow and geometric-matching benchmarks.
Zhang et al. identify inefficient exploration as a bottleneck in massively parallel robot reinforcement learning: uniform sampling spends substantial experience on situations that are already mastered or currently impossible. Their adaptive sampler instead emphasizes the frontier of a policy’s abilities, enabling difficult locomotion and assembly behaviors without demonstrations or manually ordered curricula.
Dong and colleagues trained video world models on progressively larger subsets of a 30,000-hour egocentric manipulation corpus, then separately measured how faithfully the models reproduced human motion and its effects on objects. More data improved both, but object interactions remained substantially less accurate and saturated at a level the authors judge inadequate for simulation.
The authors introduce Long-WAM, a framework that scales the context window of causal world-action models for real-time robot control. They find that longer visual history improves success only when the video foundation is pretrained autoregressively: on RoboCasa GR-1, extending context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas bidirectionally pretrained models show no gain.
The authors introduce RoboJEPA, a latent world model based on the Joint Embedding Predictive Architecture (JEPA), and train it on a large-scale multi-embodiment robotic dataset. They demonstrate that the model's latent forward-prediction error obeys a scaling law with training compute, and that this error is strongly correlated with downstream planning performance on real and simulated robots, enabling predictable allocation of resources for model training.
The authors introduce SGF+, which assigns separate parameters to context writing and denoising in autoregressive video diffusion. Trained on 5-second rollouts, it outperforms Self Forcing and Self Gradient Forcing on VBench and generates continuously for up to 24 hours without long-video fine-tuning.
Li et al. demonstrate that block-causal attention and few-step sampling can be acquired independently and combined by simple weight addition at inference, avoiding the sequential distillation typically required. The resulting model generates joint audio-video at about 26 fps at 480x832 on six GPUs, matching the bidirectional teacher on appearance and outperforming one chained baseline on most metrics.
Using 32 layer-skipping and repetition programs on two language models (Qwen3-4B-Base and Llama-3.1-8B) across 4,413 multiple-choice items, the authors compare oracle gains with gains from input-blind controls at the same sites. They find that with shared option order, these controls produced 10.2-11.8 and 15.6-19.4 percentage points of headroom—exceeding the real programs' 9.0 and 10.1 in all random-direction draws per model. This indicates that substantial headroom can persist without establishing a benefit specific to the selected layer computation.
The authors introduce AttenFeed, a module designed to subsume the functions of both Attention and Feed-Forward Network (FFN) layers, and use it to build uViT, a Vision Transformer with no strict Attention-FFN separation. Comparing uViT against standard ViTs across datasets and model scales, they find that the conventional alternating structure imposes a rigid parameter allocation that actively hinders performance at smaller model scales.
The authors curate 400 verified jailbreak failures across 8 open-source instruction-tuned models and design a local evaluation protocol that tests both harmful variant correction and benign prompt preservation. They find that strong harmful correction can coincide with severe benign regressions, even when MMLU remains unchanged, showing that aggregate metrics mask important collateral damage.
Guo et al. developed VPP2, a 14-billion-parameter world action model that predicts an intended visual future and converts that prediction into robot actions. On open-ended video tests, real-world ALOHA tasks and three post-trained benchmarks, the authors report that improving and preserving video prediction produced more reliable manipulation than competing video models and robot policies.
Alrawajfeh et al. introduce PrivTab, a tabular foundation model that embeds differential privacy directly into its architecture. Pretrained on simulated datasets, PrivTab uses in-context learning to compress sensitive rows into compact private summaries, then predicts without further access to raw data. Under moderate-to-strong privacy, PrivTab outperforms private linear and neural-network baselines, shows negligible membership leakage, and reduces dataset fitting time by 10,000x, requiring only a single forward pass.
Using random matrix theory, Hayase and Karakida establish Gaussian equivalence for multi-head self-attention: replacing softmax attention with rescaled scores plus Gaussian noise preserves the limiting spectral law of the centered output. This allows precise study of how head count, projection widths, sharing, and temperature affect representations and linear-readout learning.
The authors present Juno, a framework that repurposes a single action-conditioned joint-embedding predictive architecture (JEPA) as a control-aligned backbone, a predictive teacher, and an adaptable dynamics model across the vision-language-action (VLA) lifecycle. On SimplerEnv, Juno raises average success from 60.9% to 68.5% over the strongest baseline, and to 72.7% with test-time adaptation. On a real robot, it retains 70-75% success under background, height, and object shifts where the base policy collapses to 0%.
The authors built Bolzano, a multi-agent system where parallel prover agents propose informal proofs and a verifier agent checks them, all maintaining a human-readable research state. After initial expert-guided case studies produced eight verified results, they ran the system without problem-specific guidance on approximately 3,800 open problems extracted from four sets of papers, solving about 200 of them. Among these, four solutions to questions raised in papers accepted to STOC 2026 were confirmed by the original authors.
The authors propose Painter-Thinker (PaTh), which couples a small recursive reasoner (the Thinker) to a frozen diffusion model (the Painter) via ControlNet adapters. The Thinker refines a latent state within each denoising step, enabling iterative constraint satisfaction from pixels without symbolic supervision. On visual reasoning tasks, PaTh dramatically outperforms standard diffusion models and prior specialized methods, with margins that widen as problems become harder.
The authors introduce DisParQ, a method that learns spatially grounded, discrete part concepts from a frozen DINOv2 backbone without class labels or language supervision. On ImageNet linear probing, DisParQ achieves 83.2% top-1 accuracy, within 0.2 percentage points of its teacher, and its concepts are more consistent across images than language-aligned models. The method also enables cross-category part retrieval, outperforming self-supervised baselines on a new benchmark.
The authors collected 4,782 agent sessions of software engineers in JetBrains IDEs and analyzed the long sessions with at least three user messages, which account for 33% of the sample. They find that real task flows mix and shift among types such as explanation, planning, review, and refactor, unlike issue-derived benchmarks, and propose SWE-TaskFlow to calibrate benchmark interaction flows. A pilot on 700 SWE-Bench Pro tasks shows that changing the interaction protocol alone can roughly double agent cost without stable gains in resolve rate.
Yu et al. propose SafeEvo, a circuit-level interpretability framework that extracts refusal circuits from pretrained-only LLMs and tracks them through alignment. They show that causally ablating these circuits eliminates refusal behavior in base models, that circuit structures become less stable as alignment proceeds, and that constraining safety updates to these circuits yields stronger safety, less over-refusal, and better utility.
Kim, Cho, and Han propose Collaborative Reasoning Distillation (CRD), a multi-teacher distillation framework combining interactive cross-feedback, step-level quality assessment, and coherence-aware chain stitching. Trained on only 50K examples, their CRD-4B scores 97.3% on MATH-500 and 70.3% on AIME'25, beating same-parameter baselines.
Samuel Tetteh and Cody Fleming from Iowa State University audit a frozen CLIP prompt-margin safety score in a driving RL setup. Across 180 episodes, the score decreases before contact, but mechanism controls reveal it mainly tracks resemblance to the shared scene described in its caption bank, not the presence of an approaching hazard. Policy gains do not require the score to perceive hazards.
The authors present Unpaired CCA (UCCA), the first extension of canonical correlation analysis that learns maximally correlated projections when no paired samples are available. They prove a connection to the Quadratic Assignment Problem, derive a practical algorithm from it, and show on real-world multimodal datasets that UCCA outperforms existing unpaired alignment baselines and approaches the accuracy of paired CCA.
The authors introduce OmniCam, an autoregressive model that generates camera pose sequences from a single panoramic image and a textual trajectory description. The model uses a panoramic point-cloud encoder, a hybrid tokenization scheme (absolute rotations with sign-consistent quaternions and relative translations), and separate geometric and semantic conditioning streams with a 3D target anchor. On the new OmniCaT dataset of 267,700 trajectories across four camera behaviors, OmniCam reduces average trajectory error by 43.0% and collision rate by 62.3% compared to the best baseline.
Phan et al. trained purpose-built 4B and 9B user simulators on human conversations, explicit reasoning traces, and 13 behavioral patterns drawn from real interactions. Their 9B MIMESIS model better matched human users than frontier assistant models, while agents trained against the frozen simulator performed better across eight environments and nine unseen user models than agents trained with GPT-5.5.
Yuan et al. reframe method selection for formal theorem proving as an applicability problem, where each candidate method generates a problem-specific proposal before ranking. Using 82 Method Contracts built from Putnam 2000-2014, their self-advertisement approach achieves 95.0% hit@5 on Putnam 2015-2025 and 91.7% on IMO ProofBench, outperforming retrieval and reranking baselines.
The authors introduce ScribbleEdit, a benchmark for image editing in which the user's only input is a scribble drawn on the image. Across 23,772 synthetic training pairs and 250 manually labeled evaluation pairs, existing vision-language editing models fail to correctly identify the intention of the scribble. The authors show that understanding the scribble, rather than generation quality, is the main bottleneck.
Li and colleagues study how a video generator should respond when viewers redirect a story after part of it has already been shown. Their VIS-Ground framework explicitly models dependencies among the grounding source, the rendered video and the new request, improving a composite measure of grounded continuation by an average of 10.3 points over the strongest baseline for each of three generation backbones.
The authors analyzed why vision-language models (VLMs) miss small objects, deriving a theory based on two quantities: S (visual tokens across an object's side) and L (content per call). They found fundamental token budget limits and proved that splitting an image into overlapping views is safe (no recall loss) if no view zooms out relative to the whole image and views overlap by at least one object. Their experiments, involving about 177,000 requests on 797 images across 11 models, largely confirmed the theoretical predictions, with only a few failures for certain OpenAI models given extra pixels.
Wang et al. identify a software supply-chain attack that begins with community-shared files such as AGENTS.md, CLAUDE.md and .cursorrules. They show that injected instructions can make autonomous coding agents import a specific attacker-controlled package, and that prompts optimized on a local surrogate agent can transfer to other models and agent frameworks.
Han et al. tested whether adversarial images can control what a vision-based web agent actually does, rather than merely altering its model output. Their WebMirage framework achieved a 91.9% average attack success rate across 2,250 tasks, substantially exceeding prior attacks when success was measured by the browser action executed.
The authors propose Consistent Distribution Matching Distillation (CDMD), a method that uses only a frozen teacher and a trainable student with one objective, avoiding distillation datasets, teacher rollouts, and auxiliary networks. On ImageNet 256x256, CDMD achieves an FID of 2.04 with a single function evaluation and 1.37 with four evaluations after 40 epochs of training, surpassing prior data-free distillation baselines.
Yiwen Kou and Yimeng Wang establish the first polynomial lower bounds on the number of sequential rounds required for diffusion sampling when only approximate scores are available. For smooth near-isotropic Gaussian mixtures in d dimensions they prove a Ω̃(d^{1/3}) lower bound, and for anisotropic axis-aligned boxes inside the unit ball they prove a tight Ω(d) lower bound, both holding for arbitrary randomized algorithms with polynomially many queries per round and inverse-polynomial score error.
Kasa et al. systematically measure how serving-context factors—batch size, batch composition, padded length, precision, quantization, device, and inference backend—affect the outputs of transformer text classifiers. They train 180 models across discriminative, pseudo-generative, and fully generative formulations, and evaluate each under 3,960 serving contexts, holding the checkpoint and text fixed. The authors find that label stability does not imply score stability: changing padded length alone flips about 10% of labels, while bf16 precision moves up to 56.7 percentage points of probability mass without altering labels in some settings.
The authors introduce EDiS, a framework that performs a one-time decomposition of a graph into edge-disjoint subgraphs, then efficiently recombines these subgraphs per epoch to produce training graphs with varying edge budgets. On 19 node classification benchmarks, EDiS achieves the highest mean accuracy/ROC-AUC and lowest average rank among 17 baselines under a 30% edge budget, with the largest gains at tight budgets.
The authors show that adversarial soft prompts and images can be trained to retain influence on a language model after the input is masked from attention or removed from the context. By optimizing for persistence in the key/value (KV) cache, they achieve up to approximately 90% target success on the Qwen3-VL-8B-Instruct model and demonstrate survival after context compaction that retains only a summary cache.
The authors introduce SPIN, a training-free method that sparsifies the indexer in sparse attention by predicting block importance from past scores. On DeepSeek-V4, SPIN achieves 30–40% block sparsity and improves vLLM output throughput by up to 14.9% and median inter-token latency by up to 13.2% without degrading accuracy on long-context and agentic benchmarks.
The authors compare Adam, Lion, Muon, SOAP, and Shampoo across batch sizes in language model pretraining, independently tuning each optimizer at every batch size. They find that optimizer rankings reverse: SOAP performs best at small batch sizes, while Shampoo performs best at the largest batch size tested. They also show that for Muon, no theoretically derived hyperparameter scaling rule consistently preserves performance across tasks or batch sizes.
Gao and colleagues introduce a benchmark that tests whether generated videos obey measurable physical relationships rather than merely look plausible. Evaluating 1,280 videos across mechanics, optics, fluids, heat, phase changes, electromagnetism and surface tension, they find large differences by task and persistent inconsistencies even in the best-performing model.
The authors introduce a calibration pipeline that recalibrates the DROID dataset into a metric 3D corpus (DROID-3D) and train DepthWorld, a video diffusion world model that jointly predicts multi-view RGB and depth. The central finding is that depth supervision not only provides metric depth but also improves RGB prediction by 1.48 dB PSNR over an identical RGB-only baseline at equal training budget.
The authors provide a deterministic algorithm for online inverse linear optimization that achieves the optimal O(sqrt(d)) regret for any horizon T, and runs in time polynomial in the dimension d and T. This resolves an open question from Sakaue, who obtained the same regret but with a randomized, improper algorithm requiring exponential time per round. The new algorithm is a variant of variable-metric methods with careful metric revocation.
The authors propose Coordinated Multi-Agent Diffusion Steering (CMDS), which treats independently trained diffusion models as fixed generative primitives and learns a control that steers their joint reverse processes toward high-reward assembled outputs. They show that this amortised control recovers known targets, satisfies spatial constraints, separates degraded mixtures, and coordinates agents in maze navigation, robot planning, and human motion synthesis.
Nishat et al. profiled three small language model families on ARC-TGI, a benchmark that generates controlled abstract grid transformation tasks. In-distribution accuracy was often high, but it was uneven across task families and a poor predictor of transfer: performance collapsed when grid scale changed or when models faced ARC-AGI tasks.
Vishwanath et al. tested how large language models handle incidental information in clinical documentation and reasoning. They found that models frequently wrote small-talk exchanges into medical notes and that background speech from separate encounters contaminated transcripts. Quality scores failed to flag these errors, suggesting a need for safeguards against distraction.
The authors propose RFF-GPA, a plug-and-play Transformer attention module that treats attention as a Gaussian process posterior with a kernel approximated by random Fourier features. The low-rank formulation computes predictive mean and covariance in time linear in sequence length, versus cubic for exact GP attention. On image and text classification, including CIFAR-10-C under distribution shift, it improves calibration while maintaining accuracy.
Garg introduces a normalization framework for interpretability probes, defining a floor from declared simple inputs and a ceiling from all available information, with the gap between them as the headroom a probe can explain. In a controlled meta-analysis task, the framework shows that falling probe scores under distribution shift reflect data losing information, not representations degrading. A reanalysis of four influential LLM probing studies finds some claims survive the baseline test, while others are largely explained by the input text alone.
The authors investigate cross-tokenizer on-policy distillation (OPD) for language models. Across three heterogeneous teacher-student pairs on math and code tasks, they find that strict 1:1 token alignment covers 85–97% of tokens and retains nearly all probability mass in the shared vocabulary. A compact student-selected top-16 vocabulary subset matches full shared-vocabulary OPD, losing less than 4% of the improvement. Adding mean squared error (MSE) supervision on mismatched spans—which achieves complete alignment coverage—consistently worsens accuracy; gradient diagnostics show that span gradients conflict with strict gradients and grow in relative magnitude.
The authors introduce VETTA, a credit assignment method that jointly models turn- and token-level advantages for multi-turn LLM agents. Using separate value heads on a compact critic, VETTA combines these advantages in PPO updates, achieving success rates of 95.5% on ALFWorld and 76.0% on WebShop with Qwen2.5-7B, outperforming standard PPO by 37.5% and 22.3% respectively.
The authors introduce GeoPID, a training-free framework that decomposes multimodal representations in vision-language models (VLMs) into redundant, modality-unique, and synergistic components using geometric relationships between visual and textual subspaces. By selectively amplifying vision-unique representations during inference, they achieve an average relative accuracy gain of 7.63% across 22 VLMs and 14 benchmarks. The method requires no parameter updates and improves visual grounding in tasks that demand visual evidence.