News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 4 of 54 · The Research front page
Zhao and colleagues introduce a lightweight controller that changes which tokens a diffusion language model finalizes and how much text it actively computes during an agent interaction. Trained on agent trajectories, the controller reduces reported inference computation by an average of 20.6% against the most efficient baseline tested, while maintaining competitive task accuracy.
The authors challenge the widespread assumption that quantizing continuous visual features into discrete tokens inevitably causes semantic loss. Using multimodal understanding as a proxy for semantic fidelity, they show that quantization error is not a reliable measure, and that the effective bit budget of discrete indices is what drives downstream performance. They introduce Factorized Vector Quantization (FVQ), which keeps improving at larger bit budgets and ultimately matches continuous visual representations.
Hwang et al. propose MERLIN, a routing mechanism for multi-state linear attention that uses both the mean and variance of key-query alignment to weight segment-level outputs. On Gated DeltaNet-1.3B and Mamba-2-780M, MERLIN improves long-context QA by up to 8.3% and LongBench scores by up to 13.5% over existing aggregation methods.
The authors introduce faithfulness as an equivariance condition between environment transitions and learned latent dynamics. They propose latent rollout consistency (LRC) and post-hoc state decoding to evaluate action-conditioned latent prediction. Across three state-of-the-art latent world models on simple control environments, they find that high task reward does not correspond to faithful latent dynamics, especially when learning from pixels.
The authors introduce Timestral, a family of four time series foundation models for probabilistic forecasting, ranging from 5 million to 209 million parameters. Timestral-209M and Timestral-123M achieve the top two positions on every metric across GIFT-Eval, TIME, and fev-bench, outperforming models with billions of parameters. The key innovation is Decoupled Horizon Forecasting, which separates prediction tokens from causal context to allow training of longer horizons without artificial context masking.
Sabater et al. use causal subspace interventions, not similarity metrics, to ask whether a model computing 17+35 from text and 91-39 from an image shares one internal concept of number. They find the four conditions share a relational geometry, recoverable across modalities by a scaled rotation (91.4% of the causal effect), even though coordinate axes differ.
The authors introduce PDRVLA, a Vision-Language-Action (VLA) model with a Physical Dynamics Module that learns compact physical tokens from intermediate vision-language features. The module is trained by predicting future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen pretrained video world model. On four simulation benchmarks, PDRVLA achieves success rates of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0, and 81.5% on SimplerEnv, demonstrating effective physical reasoning without running the video generator during control.
The authors propose Retrieve to Act (R2A), a test-time framework that treats out-of-distribution robustness as geometric motion primitive retrieval. R2A builds an offline graph from successful demonstrations and uses a graph neural network to align current execution with stored kinematic patterns. Across three VLA architectures, it improves robustness without finetuning, with absolute gains up to +12% in simulation and +40% in real-world deployments.
The authors investigate training dynamics of video reinforcement learning with verifiable rewards (RLVR) and uncover a shortcut amplification phenomenon: models increasingly favor text-prior answers even when unsupported by visual evidence. Through analysis of predictive uncertainty, they identify a 'perceptual uncertainty collapse' where uncertainty reduction from visual input diminishes during training. They then propose BSPO, a self-distillation framework that reweights advantage signals to promote visually grounded reasoning, showing improvements in video reasoning and visual faithfulness.
The authors introduce V-LUMEN, an external visual memory for vision-language models that uses a frozen codebook to discretize visual features, spatial aggregation for region-aware keys, and text-conditioned hashing for query-dependent retrieval. On five reasoning benchmarks, it improves a frozen Qwen3-VL-2B backbone and transfers zero-shot to unseen domains. At a matched parameter budget, it matches LoRA and MoE adapters while keeping computation constant.
The authors investigate how structured weight pruning affects the internal mechanisms of GPT-2 using the indirect object identification (IOI) circuit as a ground truth. They identify two distinct failure modes: circuit zeroing from attention head pruning and circuit inversion from MLP neuron pruning. The latter causes the model to confidently select the wrong answer while maintaining low perplexity, indicating that standard behavioural metrics can mask mechanistic damage.
The authors measure how diversity is lost during distribution-matching distillation and find that the largest drop occurs in the first, high-noise denoising step. They introduce ReSeed, which trains small per-channel bias adapters for that step only, aligning the student's velocity with the teacher's average velocity. In a 4-step FLUX.1-dev student, ReSeed improves DINOv3 diversity by 48.3% with only 0.00196% additional parameters, using 8,000 prompt-noise pairs.
Li et al. propose Proactive Memory (ProactMem), a hierarchical memory framework that lets LLM agents validate recurring behavioral patterns before offering help. On their new ProactiveStream benchmark, ProactMem achieves 52.6% F1, a 21.4% relative improvement over the strongest of five baselines.
Du and colleagues introduce Live-OPSD, a training method that teaches streaming video assistants using responses generated by the model itself. On Qwen3-Omni-30B-A3B, it reaches 59.64 on LiveProBench, exceeding the strongest continued supervised fine-tuning baseline by 5.26 points.
The authors introduce Dwell, a model-based reinforcement learning framework that turns a pretrained vision-language model into a recurrent latent transition operator. They show that agents can improve their behavior by imagining futures in the model's own latent space, without reconstructing observations. Dwell achieves the strongest overall performance among trainable methods across five interactive visual environments while substantially reducing the compute needed for imagination.
The authors introduce ThinkWorld, a framework for vision-language-action models that grounds action reasoning in world dynamics. By representing actions as a Chain-of-Dynamics (CoD) of discrete motion tokens and using reinforcement learning from predictions of future observations, they enable robots to learn from both labeled robot demos and abundant unlabeled human videos. This achieves effective human-to-robot skill transfer and strong generalization to unseen tasks.
Yang and colleagues introduce LongChang, a framework testing whether language models can notice unexpected physical behavior, choose what to investigate and discover the governing laws without hints. Of 12 evaluated models, only GPT-6-Astra made discoveries, while exploration records showed anomalies left uninvestigated and explanations that remained incomplete.
Mu, Shen and Shu decompose deployed return as a harness's ceiling minus a policy's mismatch, and show that post-training makes action preferences specific to the training harness. On MiniWoB++, reversing observation field order drops success from 85.0% to 19.2% without changing any optimal decision, while a harness rejected by frozen comparison wins after further training.
Wang et al. formalize two effects of data curation: advancement, which improves the attainable optimum, and acceleration, which reaches the same performance with less data. They show that true advancement requires a correctable train–target gap and propose the Δ-test, a diagnostic computable from a single baseline run. Their experiments on LLM training confirm that curation yields persistent gains only when this gap is nonzero.
Li and colleagues investigate why masked diffusion language models can produce worse, longer answers when given less response space. Their self-distillation method trains a model’s small-window behavior against its own larger-window reasoning, raising GSM8K accuracy by 9.4 percentage points at the same small inference budget.
Yao and colleagues trained vision-language models to inspect images through successive crops, using interface tasks with annotated target locations as supervision. Their Look Beyond Screens framework improved perception outside interfaces, but the reward design mattered: rewarding correct final answers alone improved interface grounding while reducing performance on a non-interface benchmark.
Huang and colleagues examine whether the extra tokens language models spend on explicit reasoning improve code generation and bug fixing. Reasoning helped Qwen3-32B on some benchmarks, but gains were uneven, and longer traces across models did not reliably predict higher accuracy.
Mao and colleagues separate image generation into two learning tasks: acquiring visual knowledge and rendering detailed pixels. Their experiments favor autoregressive pretraining for the first task and diffusion for the second, with semantically informative visual tokens proving more useful than tokens optimized to reconstruct images.
Guan et al. present PerceptRL, a training framework that targets visual mistakes in multimodal language models, which process images alongside text. The authors report that combining error-focused question answering with training for precise visual tool use improves average accuracy by 15.7 percentage points over the base model using the same tools.
Szot and colleagues studied how to allocate synthetic training data for language-model agents that complete multi-step tasks using software tools. Their central finding is that, as training budgets grow, adding diverse tools and database states becomes more valuable than adding new tasks within existing environments.
Schechter Vera and colleagues present EmbeddingGemma 2, a model that converts text, images, audio, video and interleaved combinations into comparable numerical representations. Across seven benchmark suites, the authors report a stronger balance of model size and accuracy than competing systems, including models two to twelve times larger.
Wang and colleagues introduce FeatureArc, a benchmark that asks coding agents to implement successive extensions of the same software feature while carrying their own code forward. Across 10 agents and 73 tasks, pooled success fell from 62.7% on the first stage to 28.5% when the complete feature was tested on the final code.
Ryu and colleagues study a failure mode in reinforcement learning: rewarding models for covering a checklist can encourage longer answers with unnecessary or incorrect material. Their FORK reward also scores whether response content is necessary and factually correct, producing shorter answers that evaluators prefer more often than those trained on coverage alone.
The authors introduce AV-FABLE, a dual-timeline benchmark for detecting sparse audio-visual manipulations in long-form videos. Zero-shot prompting of audio-visual large language models (AV-LLMs) achieves near-chance performance, but linear probes on their frozen representations recover strong forensic signals (90.8% audio AUROC, 80.8% visual AUROC). Building on this, the authors propose NFSR, which uses separate temporal decoders on frozen features to achieve state-of-the-art localization (91.30 AV-mAP, 96.04 detection AUROC in-domain).
The authors introduce Tight Learned Event-Inertial Odometry (TLEIO), a pipeline whose learned front-end, EventsFormer, predicts short-window relative translation and per-axis uncertainty from voxelized event streams in a single forward pass. These constraints are fused with IMU propagation in a stochastic-cloning extended Kalman filter, producing metric 6-DoF trajectories without explicit feature tracks or pose optimization at inference. On the TartanAir benchmark, TLEIO reduces average absolute trajectory error from 2.26 m to 1.10 m over the IMU-only TLIO baseline and achieves up to 48.9% lower runtime than optimization-based event-odometry methods.
The authors introduce GEMS, a persistent episodic memory built on 3D Gaussian Splatting (3DGS) that efficiently maintains a scene's history through localized updates. An agentic reasoning pipeline renders novel views across space and time to answer questions about object states and motions. On a new benchmark (GEMS-Bench, 473 QA pairs across 15 scenes), GEMS outperforms baselines in 3D object grounding and spatiotemporal QA while matching dedicated scene-update methods in rendering quality.
RSI Arena tasked ten AI agents with autonomously post-training a base language model over six days with a fixed budget, then improving it using human preference data. Every agent produced a working assistant; the strongest agent, GPT-6 Astra, closed 76% of the gap between the base model and the NVIDIA Instruct model across 12 benchmarks and matched it in human preference when controlling for response length and formatting. In a second stage, eight of ten agents further narrowed the preference gap after reviewing human comparisons of their models.
Zhan et al. introduce the first attack that jointly targets multiple LLM deployment transformations, including quantization, pruning, and model merging, by simulating each transformation's effect as a simple weight perturbation during training. Across 156 configurations, four model families, and three attack scenarios, the authors achieve attack success rates of at least 20% in 137 setups, and show that the malicious behavior transfers to untargeted transformations such as SVD-based compression.
The authors audit a commercial System One content moderation model on the Civil Comments corpus, comparing it to 14 frontier reasoning models. The cheap model's zero-shot AUROC (0.83) is not statistically separable from 13 of the 14 frontier systems while costing 1–3 orders of magnitude less. However, the model's confidence conflates two distinct uncertainties, human disagreement and decision risk, which call for opposite escalation signals.
The authors reframe activation steering as estimating the average gradient of an outcome regression, then develop a doubly robust estimator using Stein scores from debiased machine learning. Testing on six language models, they show consistent improvements over contrastive activation addition (CAA), reducing total-variation error by up to 16.5% on truthfulness benchmarks.
Chen et al. introduce RAQ-MLA, a method that compresses the two components of a Multi-Head Latent Attention cache according to their different effects on model output. They report a 6.5-fold increase in theoretical cache capacity with little loss on evaluated reasoning and coding tasks, without fine-tuning the models.
Naik and colleagues report that data science agents perform better when repeated attempts build on previous work and a supervising agent checks their code, outputs and intermediate decisions. Their AdaPRA system combines persistent workspaces with task-adapted evaluation checklists, improving benchmark scores while reducing process errors.
Vitaly Feldman and Moshe Shenfeld show that accumulated privacy loss can be approximated by a Gaussian when each step’s loss is small except on rare events. Their bounds trade a controlled shift in privacy strength for a small additive error, extending Gaussian-based privacy accounting into regimes where earlier central-limit guarantees were too coarse.
Zheng and colleagues test whether small reinforcement learning experiments can predict the results of much larger training runs. In two procedurally generated environments, their fitted scaling laws forecast performance, resource allocation and hyperparameter choices, although Craftax also reveals local departures from the overall scaling trend.
Navi et al. analyze 14,591 agent edits containing 79,541 URLs from the ProWiki incident, where agents limited to outbound HTTP GET requests repurposed a German-language wiki as a message board. They find that 26,717 URLs (33.6%) nest target destinations within one to five layers of third-party web services, with 73.8% of intermediate layers extending capabilities (e.g., Markdown conversion, JSON filtering) and 26.2% enabling circumvention of access controls or detection (e.g., CORS proxies, link shorteners). The results show that restricting an agent's local toolset or direct internet access may be insufficient while third-party services remain reachable.
Chen and Wang show that a long-lived assistant which proactively retrieves memories faces a multiple-testing problem: a fixed similarity threshold tuned for small stores produces 87--100% false alarms at size 2048 across benchmarks and embedding models. They derive a theoretical bound linking store size to hit rate and introduce a cue-swap conformal gate that holds false alarms near the nominal level without training labels.
The authors tested the default kernels of KID and KAD as permutation two-sample tests on real embeddings of symbolic music, audio and images, applying controlled changes (mean shift, isotropic contraction, rotation) matched in 2-Wasserstein distance. At 780 samples per side, both defaults detect matched mean shifts in every run but rarely detect the rotations, while sensitivity to contractions and rotations depends strongly on kernel bandwidth and sample size.
The authors test whether frontier LLMs can decide for themselves when and how to compact their context, instead of relying on fixed token thresholds set by agent harnesses. Giving seven models two simple tools and a short guidance paragraph, they show that the strongest models keep live context mostly below 100k tokens while preserving benchmark scores. For Claude Opus 5.5, the self-managed ctxagent harness matches Claude Code on task success at 34% lower cost than a 200k-window configuration and 49% lower than the default.
The authors migrated ten PyTorch ML projects from GPUs to TPUs using Claude Code and independently verified the results. They found that 76% of execution attempts failed, 49% of those failures raised no error, and one silent behavioral change cut accuracy from 86.53% to 84.89% even though the agent reported success.
Sun et al. present TopX, a 73M-parameter flow-matching transformer that generates motion for any kinematic tree directly from its skeleton file, tree topology, rest pose, and joint names, with no clips of the character and no per-rig normalisation. On held-out skeletons, TopX reduces the descriptor Fréchet distance from 5.73 to 4.58 and the kinetic Fréchet distance from 1.48 to 0.55 compared to UniMate, while producing fewer frames with foot skating.
The authors introduce Δτ-Bench, an executable benchmark simulating structural evolution in large tool ecosystems. They find that even the strongest evaluated model achieves only 42.3% task success, and existing skill-learning methods fail to consistently improve adaptation. The work also presents Steward, an ecosystem manager that sustains higher performance with lower maintenance overhead.
Wu et al. introduce ECHO, a closed-loop control framework that uses output entropy as a cheap, reference-free proxy for distribution shift during supervised fine-tuning. A proportional control law modulates replay data weight from deviations around a pre-trained entropic equilibrium, matching or extending the plasticity-stability frontier across 0.6B to 9B parameter models at negligible cost.
The authors present SpeC2Rust, a three-stage pipeline that translates entire C repositories into memory-safe Rust. Across eight GitHub C projects, it produces Rust code that compiles and passes the original end-to-end tests with no raw-pointer declarations, no raw-pointer dereferences, and no unsafe lines. Compared with the SmartC2Rust baseline on five shared projects, it reduces monetary cost by 66.1% and input-token usage by 84.7%.