News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 29 of 54 · The Research front page
Chakraborty et al. derive finite-sample generalization bounds for flow-matching generative models. They show that the Wasserstein-p distance between the learned and true distribution scales as n^{-1/d} where d is the intrinsic dimension of the data, not the much larger ambient dimension. This result explains why flow matching succeeds on high-dimensional data that lie on low-dimensional manifolds.
The authors propose Seg-OPD, a method that trains large language models to revise their own intermediate reasoning steps by preferring teacher-written redrafts. The method selects high-uncertainty segments for revision and outperforms standard on-policy distillation across math and code benchmarks.
Controlled fine-tuning experiments show that standard post-training methods—SFT, RLHF, and DPO—do not merely shift a model's default behavior toward a trained perspective, but also degrade its ability to faithfully adopt an opposing perspective when instructed in-context. This reveals an inherent tension between optimizing for a single set of values and preserving the technical capacity needed to serve diverse stakeholders.
Sandlund trained Qwen3.6-35B-A3B to process long documents through two tools: one reads selected token ranges, while the other delegates subtasks to fresh copies of the model. On OOLONG-synth documents longer than 40,000 tokens, the resulting system matched GPT-5.4 using a native one-million-token context window, despite limiting each agent invocation to 8,000 tokens.
The authors adapted a full-utterance speech language model to simultaneous translation by creating target prefixes from its own translations of partial waveforms. Using a confidence threshold to control latency, they found that multi-turn append-only decoding reduces commit-calibration error by 63–68% overall and 68–80% at early prefixes, while single-turn forced-prefix training yields only modest gains. The synthesis margin that controls supervision density introduces a non-monotonic trade-off, with a small margin sometimes extending low-latency coverage on shorter utterances but a larger margin degrading quality.
The authors formulate text-to-image batch generation as satisficing: each image must meet a reward floor and the batch must meet a diversity cutoff. Varying the reward floor traces a Pareto frontier between worst-image reward and batch diversity. SatisDive, a training-free method, traverses this frontier and improves worst-candidate reward over FK steering by up to 0.43 and 0.70 on FLUX and SANA at matched DreamSim diversity.
The authors propose LMOPD, a multi-teacher distillation method that uses a lexicographic priority order to route each student rollout to the specialist for the highest-priority unmet objective. The chosen specialist's token-level correction is then projected against higher-priority experts to avoid harmful interference. On math benchmarks with 30B models, LMOPD retained over 90% of top-priority gains in a four-expert setting, outperforming baselines.
Vohra and colleagues restructure the ReAct agent loop so the model solving a task no longer approves its own actions or decides unilaterally when it has finished. Across GAIA and SWE-bench Verified, this modular design raised first-attempt success most for weaker models, while producing better-grounded trajectories without improving first-attempt success for the strongest model tested.
The authors revisit personalized adaptation for large language models by asking how much of a personalized adapter is actually personal. Through empirical analyses, they find that independent user adapters contain substantial reusable structure and that user history provides transferable correction signals. They propose LINEUP, which learns a bank of reusable low-rank personalization factors, composes them via user-conditioned recall and query-dependent calibration, and restricts target-user adaptation to a tiny eight-scalar code. Across six LaMP tasks spanning classification, prediction, and generation, LINEUP achieves the best mean performance on all 12 reported metrics, demonstrating that rich personalization can be supported primarily by shared capacity.
The authors present NEEDLE, an offline dataset-augmentation algorithm that stitches together useful segments from multiple robot demonstrations by adding short, verified action bridges. Using only RGB images and proprioception, NEEDLE identifies connections that bypass detours, combine complementary behaviors, and recover from failures. On real-robot tasks, it improves success rate over the strongest baseline on each task by an average of 21 percentage points.
Zhang et al. present GenCine, a system for directing camera and foreground motion together from a single reference image. By representing both in a shared 3D coordinate system, the method distinguishes actual object movement from apparent motion caused by a changing viewpoint.
The authors investigate language drift in LLM chains of thought during reinforcement learning with verifiable reward (RLVR) post-training. They prove that RLVR permits unbounded drift, unlike supervised fine-tuning which has a finite upper bound, and that limiting drift limits reward. Empirically, drift arises specifically on novel reasoning tasks where the target behavior cannot be drawn out of the base model, and different RLVR runs develop unique drifts.
The authors introduce TVRL, a framework that derives token-level credit for video reinforcement learning from the same reward signal. By using gradient magnitudes from a frozen vision-language model's question-answer likelihood, they reweight denoising log-probabilities, leading to a 3.60 point improvement on VBench-2.0 Overall score over the base model and consistent gains across multiple samplers and reward models.
Noori et al. propose GeoWM, a world model that predicts future scene geometry directly in explicit 3D space rather than in pixel or latent space. The model extracts geometric features from observed RGB frames using a geometry foundation model, then uses a horizon-conditioned flow-matching diffusion transformer to forecast depth, camera pose, and 3D point clouds at a specified future time without autoregressive recursion. On four datasets spanning driving, aerial flight, and robotic manipulation, GeoWM outperforms existing world models by 4 to 58 percent in depth error at short horizons and 7 to 48 percent at long ones, while reducing inference time by about 98 percent.
Zhu et al. systematically analyze how verifier capability and prompt design affect self-verification in vision-language models (VLMs), finding that stronger verifiers are more reliable and no single prompt dominates across tasks. They propose MOTIVE, which evaluates candidate answers from multiple verification perspectives and learns a correctness-aligned reliability score to guide accept-or-rethink decisions, consistently outperforming strong self-verification and self-correction baselines.
Park and Levine introduce OCBench, a simulated robotic manipulation benchmark designed to reproduce behavioral cloning phenomena that usually appear only with costly human-collected data. By generating narrow, history-dependent and smoothly varying trajectories, the benchmark reproduces four reported anomalies, including benefits from overfitting and action chunking, while allowing candidate explanations to be tested under controlled conditions.
Ding et al. investigate how vision transformers recognize abstract ideas such as anger when an image contains only concrete cues such as fire, faces or colors. Using sparse feature extraction and causal interventions, they find structured computational paths in which low-level visual features lead to concrete conceptual anchors before abstract predictions emerge.
Abdali, Ko and Cameron present a training-free alternative to KV-cache eviction, which permanently removes selected tokens to save memory. Their method computes low-rank representations from each prompt’s own activations and reports performance comparable to a dense cache across an agentic benchmark and LongBench while using as little as half the KV-cache memory.
Che and Li test whether interpretable LLM features merely control outputs when edited or actually participate in the model’s normal computation. Across entity recognition and grammar tasks, they find that strong steering often fails to establish either sufficiency or use, particularly when interventions exceed values observed on natural inputs.
Ramesh et al. address a weakness in disaggregated LLM serving: fixed prefill and decode worker pools can become imbalanced even when the deployment has enough total capacity. Implemented on SGLang and tested with production-derived Azure workloads across three language models, FluidPD improved SLO attainment by as much as 94.6 percentage points over static SGLang under the same GPU allocation.
Cai et al. find that independent runs of an autoresearch agent often settle into different, persistent performance plateaus, apparently because early choices confine later experimentation. Their fork-and-flush method preserves accumulated files and results while periodically restarting reasoning across several parallel branches, then retains the branch with the best score.
The authors introduce MoLE, a framework that transforms latent tokens into specialized visual experts by controlling both the visual evidence each token receives and how it transforms that evidence. Across five visual reasoning benchmarks, MoLE achieves an average score of 78.6, outperforming the strongest latent reasoning baseline by 3.6 points and supervised fine-tuning by 4.9.
The authors introduce MemComposer for collection-grounded B-roll generation. Given a user's video collection, a natural-language directive, and a target duration, the system builds a structured memory of characters, settings, objects, and style, then retrieves conditioning frames and generates a multi-shot sequence. In a user preference study, MemComposer significantly outperforms ungrounded text-to-video planners on both prompt adherence and visual alignment to the collection.
The authors propose metadata-supervised pretraining (MSP) for encoder-free speech LLMs, using speaker identity and emotion labels to train models that directly process Mel-spectrograms. On joint ASR and speaker diarization, their model outperformed a randomly initialized encoder-based baseline under matched training data and was competitive with encoder-based models pretrained on much larger datasets.
Shin et al. examine why video language models often give the same answer even when a video's frame order is reversed. They find that temporal order is represented most strongly in intermediate layers, but this signal progressively diminishes toward the output. Their inference-time method restores part of that fading signal and improves temporal reasoning across three models and four benchmarks.
Dutulescu et al. address a mismatch in Direct Preference Optimization, where preferences apply to whole responses but language models learn token by token. Their method, GAW-PO, uses gradient alignment to avoid suppressing rejected-response tokens that support behavior also found in preferred answers, producing the best average result across 11 evaluated benchmarks.
Adeniran-Lowe et al. show that state-of-the-art 3D encoders, pretrained on globally reconstructed scenes, fail catastrophically on partial viewpoint-dependent point clouds (drop from 72.47 to 2.57 mIoU), primarily due to coordinate-frame mismatch rather than missing geometry. They propose PAGER, a parameter-efficient, label-free adaptation that aligns partial-view features with a frozen global semantic representation using only paired partial/global geometry. On ScanNet partial views, PAGER achieves 65.46 mIoU without any semantic labels, outperforming supervised fine-tuning baselines, and in zero-shot transfer to ScanNet++ it surpasses a fully fine-tuned Sonata encoder (53.93 vs. 48.09 mIoU).
Kao et al. show that in verifier-free reinforcement learning for LLMs, the posterior probability of a correct answer conditioned on a reasoning trace collapses to a low-variance interval as the trace lengthens, making rewards nearly indistinguishable. This undermines group-relative policy optimization like GRPO, which relies on reward differences across rollouts. The authors propose RLCPR, which improves stability and token efficiency via uncertainty-aware sampling and concentration-aware regularization, outperforming existing baselines by up to 4 percentage points on six benchmarks.
Jiang, He and Chen present CoFlow, a flow-matching method that repels nearby, conflicting trajectories during training rather than changing endpoint couplings or applying post-training distillation. The authors connect trajectory crossings to large local Lipschitz constants in the learned velocity field and report improved ImageNet 256×256 generation, particularly with 20-step sampling.
Micheli, Zerio and Bhatt propose Discrete Drifting, a framework that adapts one-step generative modeling to finite state spaces. In a controlled, exactly computable setting, they show that its Markov jump dynamics follow the intended probability flow, reduce divergence from the target distribution and can be absorbed by a finite-capacity neural generator.
Lin et al. introduce MMVistaReason, a post-training framework for multimodal reasoning that combines broader capability coverage, efficient SFT and RL data construction, and a specialize-then-integrate training strategy. Their 4B model averages 72.8 across 15 multimodal benchmarks, outperforming Qwen3.5-9B and MMFineReason-8B, and the scaled 9B version reaches 74.4, surpassing Qwen3.5-35B-A3B.
Fradlin et al. present ARROW, a feed-forward model that reconstructs scene geometry and tracks 3D points using arbitrary collections of RGB observations, including videos, static images and multiple camera streams. Its central mechanism assigns each observation a content-derived identity rather than a sequence position, allowing queries to address source, target and reference images without imposing temporal order. The authors report state-of-the-art results on two 3D tracking benchmarks while retaining competitive reconstruction performance.
Yang and colleagues introduce SELF-POT, a framework that compares direct answers, parallel sampling and self-revision while charging every generation, critique and selection call in dollars. Tests on 350 sealed tasks show that additional inference can create useful candidates, but selection rules and failed protocol stages often determine whether that potential becomes higher final accuracy.
The authors show that a completed reinforcement learning (RL) training history contains structure that can be exploited to construct policies stronger than any checkpoint visited during training. They propose SURGE, which spectrally decomposes the weight updates from a high-accuracy anchor checkpoint and a competitive donor checkpoint, retaining the anchor's dominant component while incorporating complementary information from the donor. The method achieves state-of-the-art results on multiple reasoning benchmarks, exceeding the best native checkpoint without additional training or per-query inference cost.
Lee and Kang measured how much of the reported out-of-distribution detection gain survives when the detector is tested on an OOD dataset it has never seen—not just on held-out samples of the same dataset. They find that most of the gain comes from recognizing which dataset the input belongs to, not from detecting novelty per se, and they derive a closed-form formula that predicts the inflation without needing a separate hold-out set.
The authors explain the spurious-reward paradox—why random rewards improve LLM performance—by introducing reachability, the potential for further training to elicit dormant capabilities. Applying random-reward RL to OLMo checkpoints across training, they find three distinct regimes of responsiveness: early pre-training (dormant), later pre-training (receptive), and mid-training (autodidactic). The pattern reveals that capabilities not yet expressed in current accuracy can be activated by training, even without correctness feedback.
The authors show that standard RLVR training, while improving single-sample accuracy, increases error correlation among samples, degrading majority vote. They propose adapter thickets—training K independent LoRA adapters on random data shards—which outperform a single trained adapter in all tested settings, maintaining vote quality even at large vote budgets.
Song et al. introduce the attack pyramid, a five-level hierarchy of shortcut attacks that exploit different information sources (options, text, other items, single frames, or shuffled video). Applying it to 115 video benchmarks, they find that on 35 benchmarks attackers that never see a frame achieve full-video accuracy, and on 51 benchmarks with temporal probes, shuffled frames retain a median of 96% of the full-video accuracy. They screen 505,518 question-answer pairs to produce Video-Index, a meta-benchmark of the 840 hardest verified items, and show that even the best evaluated system (Claude Opus 5 with agent tools) leaves significant room for improvement.
The authors introduce APEX, a learned controller that dynamically selects the speculative decoding mechanism (EAGLE-3, n-gram, or a draft model) and adapts draft depth at each verification block to balance throughput and wasted computation. Evaluated with Qwen3-8B across six workloads, APEX achieves up to 5.24x wall-clock speedup over autoregressive decoding, and its depth-optimizing variant reduces wasted-token percentage by 41% relative to fixed n-gram speculation at depth 16.
Liufu et al. train a software engineering agent on recovery attempts that follow failed code patches, rather than treating every attempt as independent. On 500 SWE-bench Verified tasks, their FC-SWE method improved both first-attempt performance and recovery under multiple-attempt budgets compared with standard Group Relative Policy Optimization (GRPO).
Fishelson and Mohri give a single online algorithm for calibrating high-dimensional forecasts without assuming a stochastic model for outcomes. For any fixed accuracy ε, its required number of rounds is polynomial in dimension for simultaneous binary forecasting, replacing previous exponential dimension dependence.
Minegishi and colleagues tested whether vision language models generalize from visible objects or from the abstract relations among them. They report that stronger models increasingly choose relational matches, and that open-model analyses trace this behavior to competition between early object-focused processing and late relational processing.
Scholten et al. recast compounding quantization error as a robustness problem. By injecting activation noise, fitting linear denoising transforms and then rounding weights in a noise-aware metric, they improve low-bit post-training quantization without calibrating layers sequentially or adding inference-time computation.
Luo and colleagues test whether on-policy distillation, which trains a student using token-level feedback from a teacher on the student’s own outputs, can transmit a teacher model’s latent backdoor. They find that an initially clean student can acquire trigger-conditioned harmful behavior from a safety-aligned but compromised teacher, with attack success reaching 70% at a 3% poisoning rate.
Toso and colleagues identify a control failure hidden by standard JEPA training: accurate next-step prediction and anti-collapse regularization can coexist with representations that erase unstable but controllable dynamics. They propose endpoint inverse dynamics, which reconstructs an action sequence from the initial and final latent states, and prove that exact reconstruction preserves every state direction reachable within the training horizon.
Wistreich and colleagues introduce an offline augmentation framework that converts demonstrations recorded at one canonical robot pose into physically consistent examples at perturbed poses. Across simulated humanoid and bimanual tasks, plus experiments on a real Galaxea R1 Pro, policies trained with these examples were more robust to unseen base-position and camera-viewpoint shifts than the tested baselines.
The authors present Kepler, an open-source harness for ARC-AGI-3 that represents agent hypotheses as executable world models and validates them through retrodiction and prediction checks. Under a frozen Claude Opus 5 configuration, Kepler achieved a server-verified 100.00 RHAE on all 25 public games, with 181 of 183 completed levels using no more actions than the median-human baseline. The paper also reports three evaluation failures—source-code leakage, autonomous tool reconstruction, and silent planner repair—that score-only evaluation would have missed.
Jin and colleagues introduce ReLiveGym, a diagnostic environment for evaluating language-model agents that monitor evolving information streams and act only when needed. Across eight base models, the authors find that performance depends not just on reasoning ability, but also on how the surrounding agent system decides when to trigger an action.