News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 24 of 54 · The Research front page
Saied and Fleuret challenge the explanation that mean-squared-error training systematically makes flow-matching velocities too small. Their analysis and ImageNet-256 experiments instead attribute much of velocity scaling’s benefit to population time lag, where sampled states resemble training states from an earlier point in the generation process.
The authors introduce Timer-M1, a multivariate time series foundation model pretrained on a pipeline that synthesizes elementary temporal patterns (trend, seasonality, regime changes) and cross-variate relations. By assembling real and generated series into forecasting episodes with varied variate roles, the model learns to generalize across domains and task configurations. On three large-scale benchmarks, Timer-M1 ranks first on FEV and TIME and second on GIFT-Eval among recent time series foundation models.
The authors introduce InfiLoop, a residual connection for looped Transformers that learns to selectively aggregate past recurrent states instead of overwriting them with each new iteration. On Sudoku-Extreme, a 7M-parameter InfiLoop model reaches 97.9% exact accuracy, and it continues to improve even beyond 20,000 effective steps, demonstrating that added loop depth can translate directly into stronger reasoning.
The authors propose Residual Advantage (RA), which treats the teacher-student probability residual as a bounded one-step reward, subtracts the student's own expected value to form an advantage, and centers it within each response. This allows the teacher to redistribute credit among tokens while the verifier sets the overall response label. Combined with GRPO or REINFORCE++, RA improved Avg@8 by 1.7–3.6 points and Pass@8 by 3.9–6.3 points across three mathematical benchmarks.
Zhao et al. study a minimal two-unit ReLU network where pretraining drives neurons toward proportionality. After source training, gradient flow predicts quick separation, but finite-batch SGD fails with high probability for a time exponential in 1/η. The failure is traced to inputs where the two ReLU gates disagree, which become exponentially rare due to weight decay during pretraining.
The authors analyze decoder attention in vision-language models and discover that early and late layers consistently attend to the same image regions regardless of the prompt, while middle layers shift attention meaningfully. They propose SAGE, a lightweight intervention that uses pretrained vision backbones to generate region-of-interest masks and steers attention away from these prompt-invariant sinks, improving visual grounding and reducing hallucinations across several VLM architectures.
The authors propose UniData, a universal pipeline that transforms simple user requirements into multi-round, multimodal instruction datasets. They also introduce UniDataset, a dataset of 20,000 entries spanning nine modalities. Experiments show that UniData achieves state-of-the-art data quality in reasonableness, clarity, detail, and relevance, and that the generated data can enhance the understanding and generation capabilities of other multimodal large language models.
The authors propose ReCal, a plug-and-play calibration method that adjusts pruning criteria based on forward KL divergence between an unpruned teacher and a pruned probe. Applied before pruning, ReCal helps preserve predictions vulnerable to pruning, enabling better recovery via on-policy distillation. Across Qwen3 models and multiple pruning criteria, ReCal consistently improves mathematical reasoning and code generation.
Zhao et al. propose Semi-OPD, a distillation method that reuses a single set of offline student rollouts sampled before training rather than regenerating them on-policy at each step. They show that Semi-OPD outperforms standard on-policy distillation (OPD) on most teacher-student pairs, especially when the teacher and student are poorly aligned. The paper provides a practical metric, the token overlap ratio, to decide which method to use.
The authors introduce PointVGGT, a zero-shot framework for multiview RGB-D registration that replaces the conventional pairwise-then-global paradigm with a foundation-then-refinement approach. Using visual geometry foundation model priors, they directly recover metrically consistent global poses and achieve robust alignment without any training or finetuning, outperforming both classical and learning-based methods across indoor, object-centric, and outdoor datasets.
The authors introduce MemoType, a framework that classifies conversational memories into three types—episodic, personal semantic, and general semantic—and applies retrieval strategies customized to each type. On three benchmark datasets, MemoType achieves up to 16.18% improvement in Recall@1 over existing memory systems.
The authors introduce WebLoop, a framework that trains a generator, critic, and refiner as roles of a shared policy for functional web generation. The critic learns to diagnose requirements without execution traces, receiving credit from downstream improvement. On WebRise and WebGen-Bench, WebLoop with Qwen3.5-9B improves the base model by 11.3 and 15.4 points respectively, and gains persist at 27B scale and to multimodal inputs.
The authors propose Conditional Residual Prediction (CRP), a training method that forces a causal video diffusion model to first predict each frame without using past frames, then let history add only a residual. In controlled experiments, this nearly closes the 6.14-point VBench quality gap between causal and bidirectional models of the same size. Scaling this recipe, they train Optica, a 2B-parameter causal model that generates 5-second 480p videos and scores 82.78 on VBench using only about 15 million training videos.
Wang et al. present H-SVDQuant, a framework for post-training quantization of diffusion transformers to 4-bit weights and activations (W4A4). By formulating a joint optimization problem that accounts for low-rank compensation, residual quantization, and activation noise, they derive a deflated Hessian and an activation-noise surrogate. In experiments, rank-4 H-SVDQuant outperforms rank-32 SVDQuant on SANA-1.6B and FLUX models, while reducing quantization calibration time by up to 6.25 times.
Han et al. propose iCATS, a training-free sparse attention framework for Diffusion Transformers that improves importance estimation by clustering tokens based on query-key dot-product interactions (reformulated as a quadratic form) and adapts sparsity using a signal-to-noise ratio (SNR) guided schedule. On HunyuanVideo-T2V-13B, iCATS achieves 2.03× acceleration at 31.017 dB PSNR, and on Wan2.1-T2V-14B, 1.55× at 29.301 dB PSNR, outperforming prior training-free methods.
Li and colleagues investigate whether multimodal large language models can internalize the geometric reasoning performed by tool-using coding agents. Their SpatialOPSD framework uses verified agent traces as privileged training information and, on two spatial benchmarks, outperforms supervised fine-tuning and reinforcement-learning baselines while approaching tool-assisted agents.
Deng et al. introduce SWE-Journey, a benchmark combining long-horizon software development with multi-turn, persona-conditioned interaction. Coding assistants performed substantially better when requirements came from simulated software architects, passing more than 75% of functionality tests on average, but passed fewer than 25% with simulated non-coders.
Feng and colleagues introduce Parametric Trajectory Distillation (PTD), a method for compressing iterative video generation into four model evaluations. Rather than forcing the student to reproduce difficult teacher transitions directly, PTD represents each trajectory segment as a polynomial and obtains teacher guidance along the student’s predicted path. Across two large video backbones, the authors report better naturalness and diversity than strong few-step baselines.
Cheng et al. turn a pretrained 30-billion-parameter Muse-Glimmer decoder into a multimodal bi-encoder that matches requests to independently encoded candidates. Across 11 benchmark suites and 190 tasks, the authors report stronger results than existing multimodal embedding models on every evaluated decision-making task, near-leading retrieval performance, and persistent weaknesses on problems requiring complex reasoning.
The authors present DreamTrue, a multi-view robot world model that predicts future video frames conditioned on action sequences. They address two data-quality problems: imprecise camera calibration and a lack of recorded failed interactions. With offline geometric calibration and reward-guided counterfactual post-training, DreamTrue achieves state-of-the-art action following on AgiBot and cuts a human-assessed interaction defect rate from 48.12% to 6.25%.
Hollinsworth et al. trained white-box probes on the largest deception dataset to date (FIBS) to detect when language models produce statements they believe are false. The probes achieved 98.8% AUC on the SHADE-Arena benchmark, surpassing a frontier black-box text-monitoring baseline (Opus 5.5), and scaled favorably with model capability. They also detected introspective falsehoods—lies that cannot be identified from model outputs alone—with AUC up to 99.7%.
Li et al. show that quantizing AdamW's second-moment optimizer state based on its effect on the update (preconditioner space) rather than the state value itself reduces distortion. They propose two 4-bit configurations, ZIP-SR and ZE-EDEN, which reduce the mean validation-loss gap to 32-bit AdamW at every evaluated model size, with the largest gap reduction reaching 70%.
Liu et al. introduce the task of alignment generalization prediction: forecasting how training a language model on one value alters its behavior on many other values. They benchmarked representation methods on 66 values from modern alignment targets and found that activations-based similarity significantly outperforms description-based baselines, with a correlation of 0.45 versus 0.05. Additionally, they show that value representations can predict the robustness of models trained on multi-value targets and yield an empirically grounded taxonomy of LLM values.
Ali et al. introduce the Latent Core Tokenizer (LCT), which first learns reusable linguistic units using minimum description length, branching entropy, and morphotactic constraints, then builds a surface vocabulary by recombining those units. Across 104 languages with a 200K token budget, LCT achieves lower fertility and higher MorphScore than BPE, Unigram, and parity-aware BPE, and improves aggregate downstream scores by 1.48 to 2.00 points in controlled encoder pretraining.
The authors introduce 'prior barriers' to measure how strongly a pretrained model favors competing concepts over a target concept. They observe these barriers follow a long-tail distribution, meaning frequent concepts need little additional evidence while rare concepts demand more instructions. They propose PASS, an adaptive instruction selection method that allocates budget to under-represented concepts, and show it outperforms seven baselines across four backbone-budget settings.
The authors demonstrate that language models can act as Research World Models (RWMs), predicting the outcomes of experimental interventions across multiple research environments. Using 2,653 experimental records from nine pretraining, post-training, and inference environments, they show that research knowledge acquired from real experiments improves prediction accuracy for unseen interventions within the same environment and can transfer to different environments, reducing selection regret by 78% in one case. This supports the idea that experimental data can be reused to guide future research decisions.
The authors introduce VibeEdit, an image editor that follows canvas instructions—circles, scribbles, arrows, and optional short notes placed directly on an image—to perform object addition, removal, replacement, attribute change, and movement. They train the model on 1.55 million edit pairs using layer-decoupled conditioning and rubric-guided reinforcement learning. On a 419-case benchmark emphasizing target selection among similar objects, VibeEdit achieves a VLM rubric score of 79.9 and an outside-region PSNR of 32.8 dB, substantially exceeding the best text-instructed baseline (FireRed: 67.4 and 24.0 dB).
The authors propose Multi-Agent Self-Supervision (MASS), a method where a single model generates, evaluates, and optimizes multi-agent workflows, then fine-tunes on those trajectories. Over two cycles with a Qwen3.6-27B model, they achieve 1.2-1.6x higher performance per output token on four benchmarks, and multi-agent trajectories are more training-efficient than single-agent ones, suggesting that jointly learning orchestration and subagent execution provides effective signal for recursive self-improvement on non-verifiable tasks.
The authors present SuperNav, a navigation framework that equips a frozen pretrained multimodal large language model with an agent harness for tool use, progress tracking, and context management. Without any navigation-specific fine-tuning of the MLLM, SuperNav achieves 78.00% success on single-object navigation and 59.50% on demand-driven tasks, outperforming four baselines. It also reaches 73.33% success on a category-level HM3D benchmark and was deployed on a real quadruped robot.
The authors propose Hippocam, a memory architecture that organizes an LLM agent's interaction history as a tree of nested intents, consolidating completed tasks and recursively abstracting older unused experiences. Evaluated across eight measures of memory, knowledge use, and learning, Hippocam led or matched the best compared method on seven of them, demonstrating that agents can learn and evolve capabilities through their own experiences without parameter updates.
Liu et al. propose QGMem, a memory framework that stores dialogue histories as event-indexed atomic memory units and consolidates related units into dynamic memory traces. When a query arrives, hybrid retrieval and query-aware reranking select a compact working memory, which is then organized as a local graph and encoded into a graph token for the LLM. Across six benchmarks, QGMem consistently outperforms flat and graph-based memory systems in retrieval, multi-hop evidence composition, conflict resolution, and ultra-long dialogue reasoning.
The authors propose EvoAlloc, a resource-allocation agent that revises its strategy based on accumulated search experience. Across coding and agent-harness benchmarks, EvoAlloc requires 59–82% fewer full evaluations and 61–89% fewer LLM tokens to match baseline performance, and achieves 8.7–12.0% higher final performance under the same evaluation budget.
The authors propose ResOT, a training-free, inference-time method that mitigates object hallucination in large vision-language models by repairing, rather than suppressing, hallucination-related hidden representations. It aligns each token state toward a faithful distribution via Gaussian optimal transport in a localized residual subspace. Across three LVLMs, ResOT reduced hallucination while improving caption quality and multimodal performance.
Yang and colleagues present a benchmark for diagnosing how multimodal language models describe continuous audio-visual content. Across 786 densely annotated videos, their structured evaluation indicates that frontier models can recognize many local details but struggle to preserve identities, temporal relationships and audio-visual alignment over longer sequences.
The authors introduce Memento 3, a system that enables a frozen large language model (LLM) to continually learn explicit world models by storing revisable hypotheses in an external natural-language rulebook. The rulebook is compiled into executable code for planning and prediction, and the agent iteratively refines both representations through a reflective loop of observation, rule revision, and verification via exact replay. On ARC-AGI-3, the single-model agent cleared all 25 public games with a mean Relative Human Action Efficiency (RHAE) of 100.0 and used only 44% of the human action count; in an Atari Pong case study, it learned a feedback controller that won 21:0 in each of three evaluated episodes without additional LLM calls.
Zhang et al. introduce FAITH, a model-free reinforcement learning framework that decouples safety enforcement from task optimization using a learned safety critic and feedforward filter. The task policy is trained through the filter on task reward alone, and when no safe action exists, the filter selects the action with minimum predicted peak harm. In simulations and on a physical Unitree G1 humanoid, FAITH achieves high safety rates while retaining most of the unfiltered task return.
The authors introduce FastBench, a benchmark of 306 QA pairs across eight domains designed to evaluate streaming video language models on high-dynamic content. They find that the strongest model, Gemini-3.5-Flash, scores only 50.7%, and that denser sampling improves accuracy but saturates, suggesting current models cannot autonomously decide when fine temporal detail is needed.
The authors introduce WOVEN, a training source and benchmark for visual transition reasoning, and show that training on it improves multimodal LLMs across many external benchmarks, even with small subsets. They propose a training recipe based on reasoning operation rather than scene or action.
The authors audited HELM Safety's HarmBench dataset for construct validity of 'harmful refusal'. They found that HarmBench does not measure a single attribute: multidimensional item response theory required several latent dimensions, and differential item functioning flagged items where developers affect scores beyond refusal ability. This undermines the assumption that safety benchmark scores reflect unidimensional constructs.
Wang et al. introduce SparseDecoding, a pruning framework for LLM inference that addresses two problems: the distribution shift between teacher-forced calibration and autoregressive decoding, and the lack of efficient sparse matrix-vector (SpMV) kernels for N:M sparsity. By using activation statistics from the model's own decoding (excluding prefill) to guide pruning, and by implementing a bitmask-indexed N:M SpMV kernel, the authors achieve consistent quality improvements over standard fixed-text calibration on long-form generation benchmarks and deliver practical decoding speedups on A100 GPUs.
Kara et al. introduce BudgetPix, a compute-adaptive tokenization framework for pixel-space diffusion models that allocates tokens based on visual complexity rather than a uniform grid. On text-to-image generation, BudgetPix matches the fidelity of baselines MiniT2I-L at 512² and PixelDiT at 1024² using only 25% of the original token budget. For class-conditional generation with a MeanFlow backbone, the method produces near-zero quality degradation at 60% of the full compute budget.
The authors propose VINCIE-NExT, a video editing framework that leverages in-context visual demonstrations from image editing pairs to guide video edits. By decomposing video editing into a chain of sub-tasks that route through the image domain, the model can be trained jointly on abundant image editing data and video data. On the OpenVE-Bench benchmark, VINCIE-NExT achieves state-of-the-art performance across various editing types without requiring large-scale paired video editing data.
Kumar, Rathore and Moitra audit whether LLM judges produce verdicts that are reproducible, unaffected by candidate order and correct. They find that identical queries can yield different answers even at zero temperature, while swapping response order can reverse decisions at high rates. The authors propose the trustworthy verdict rate, which measures all three requirements jointly.
The authors introduce ViMoD, a framework that couples learned visual memory with adaptive access to fine-grained tokens, targeting a low average active visual token budget during decoding. At a 20% target budget, it improves mean normalized score by 39.0% over the strongest one-shot baseline on eight reasoning benchmarks, with only 0.0546% additional trainable parameters.
Yang and colleagues address a mismatch between conventional text-to-SQL training and deployment, where agents must interact repeatedly with a database before submitting an answer. Their framework trains on verified, full interaction trajectories inside the same SQL harness used at runtime, then applies reinforcement learning based on execution correctness. On Spider 2.0-SQLite, this raised execution accuracy from 15.5% to 45.2% for Qwen3-8B and from 22.2% to 54.8% for Qwen3-14B.
Ding examines why reference-policy KL regularization, which discourages a trained model from drifting too far from a baseline, can impair group-relative policy optimization. The paper identifies seven ways KL updates can become misaligned with reward learning, then proposes Zero-Sum Calibrated Policy Optimization (ZCPO), which incorporates reference information by calibrating relative rewards rather than adding an independent penalty.
Lin and colleagues replace the fixed timestep schedules used by conventional few-step flow generators with repeated refinement in image space. Under matched distribution matching distillation training, their approach performed better than standard flow discretizations and matched or improved upon schedule-searched variants on ImageNet at 256 by 256 resolution.
Cheng et al. examine whether an agent must generate a new chain of reasoning before every action. Their method, RACE, uses changes in the likelihood of reference actions to identify reasoning steps that can be removed, cutting reasoning-token use by 32.3% to 80.2% across four benchmarks while maintaining or improving reported task performance.