News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 7 of 54 · The Research front page
Jia and colleagues test whether AI models can connect evidence scattered across long videos, including multiple television seasons. They report that standalone models struggle, while a system combining persistent memory, retrieval and specialist tools improves every tested model, suggesting that access to relevant evidence remains a major bottleneck.
Zhang and colleagues introduce Poseidon, a pipeline that converts scientific code repositories into executable, automatically checked tasks for training research agents. Their resulting collection, Pearl, contains 97,659 environments, and supervised fine-tuning on it improves three Qwen3.5 models across six scientific-agent benchmarks.
The authors model datacentre power and develop a Viterbi-based detector that follows the wandering iteration cadence of synchronous training. Against an operator who does not hide, it detects 96% of runs at 1% false-alarm rate, within 0.04 of the optimal Neyman-Pearson bound. Adversarial strategies that vary work per iteration can hide training in up to 73% of five-minute windows, but over a day 87% of runs are still caught at 0.1% false-alarm rate, and hiding requires 93–267% more compute for small-scale training.
Ye and colleagues ask coding agents to deliver two things together: a faster implementation and a machine-checked proof that it preserves the reference implementation’s behavior. Agents produced verified improvements for 156 of 189 standalone targets, while application-level experiments showed that faster individual functions do not always translate into comparable gains for the whole program.
Yu and colleagues introduce OpenDS-World, a benchmark testing whether AI agents can follow complex evidence trails across the live Web rather than retrieve answers from a fixed collection. Their evaluation finds substantial difficulty for both models and human participants, with additional searching and longer execution not necessarily producing better answers.
The authors audit the SkinCon dermatology concept dataset for skin-tone dependence. They find that erythema annotations are strongly associated with light skin (odds ratio 0.17 for dark versus light), while hyperpigmentation and hypopigmentation are associated with dark skin (odds ratios 6.2 and 4.5). A concept bottleneck model trained on these labels learns the same bias, and three label error detectors fail to find the omissions, underperforming a disease-only baseline.
The authors present OptiCycle and OptiState, two inference harnesses that let a frozen language model iteratively execute and revise solver code. On 2,634 optimization problems, execution-guided revision improved accuracy over direct generation for every model tested, with gains up to 49.46 percentage points.
Bao et al. test whether language models need elaborate memory-management systems to answer questions about past conversations. Their Raw-RAG baseline, which retrieves relevant passages directly from conversation history, ranks first in six of seven evaluated settings, suggesting that strong retrieval and capable readers can often substitute for explicit memory restructuring.
The authors introduce TRIAD, a benchmark of 327 text–image–audio queries each designed so that the correct answer is unique only when all three modalities are present. Evaluating 14 omni-modal systems, they find that the best model achieves 71.2% group accuracy, falling over 20 points behind human performance, and that removing multiple-choice options drives model scores into the single digits.
The authors sum low-rank adaptation updates trained separately for the same image classification task on different pretrained Vision Transformer backbones. They find that the resulting superposed LoRA improves all 56 backbone-task pairs by a mean 30.21 percentage points over the unadapted backbone, only 0.28 points below individually trained solo LoRAs. This suggests that each frozen backbone implicitly selects its own component from the sum, enabling adapter reuse without cross-model alignment.
The authors train safety probes on 20 language models from five families using up to 10^6 labels per task, and find that probe error, defined as 1 - AUROC, follows two power-law scaling laws in model size and label count. Fitted on a family's smaller models, the laws predict the next larger model's probe error to within 0.02 in 72% of cases.
The authors introduce ChronoEvo, a benchmark of 1,709 medical guideline revisions, and show that large language models frequently recognize when clinical guidance has changed yet fail to incorporate that change into their recommendations. In the primary model, the gap between recognizing a revision and fully applying it reached 32.3 percentage points.
The authors present FIGFLEX, a co-designed accelerator and quantizer for LLM inference that keeps activations in FP16 and weights as low-bit integers, without dequantizing. Its radix-4 bit-parallel datapath halves partial products and scales throughput as precision falls. On LLaMA-2-7B, it beats five prior quantizers in perplexity at 2 to 3 bit average precision and delivers up to 1.56x higher throughput per area than AxCore.
Sparse autoencoders try to make neural-network representations understandable by separating them into features associated with recognizable concepts. Evaluating hundreds of these systems, Zhang and Wang report that features contributing the largest shares of activation can mix concepts and duplicate other features, weaknesses that existing evaluations can miss.
Che, Ivanovic, and Pavone propose Short-to-Long (S2L), where smaller-discount source value networks supply features to a behavior learner whose discount is unchanged. It improves over matched baselines on Atari, online control, and offline MuJoCo, and ablations confirm the role of each design choice.
The authors propose Factor-Aware Predictive Learning (FAPL), a three-stage training procedure that makes the representations of a frozen multimodal large language model directly actionable for a Video Diffusion Transformer in cinematic generation. FAPL consistently outperforms baselines on sparse-to-video, factor-aware editing, and multi-shot tasks, with each stage contributing to the overall improvement.
Liu, Wei, and Zimmert establish minimax sample-complexity bounds for reverse-KL policy distillation under four settings, varying whether trajectories come from the student (online) or teacher (offline) and whether the teacher reveals its full action distribution or only the sampled action's probability. They find that online action-feedback distillation can require substantially more data than full feedback, whereas offline distillation achieves the same optimal rate regardless of feedback type.
The authors propose Gandalf, a KV cache offloading system for LLM serving that uses attention layer type as a placement signal when deciding which caches to write to SSD. By preferentially admitting full-attention KV caches over linear- and sparse-attention states, Gandalf extends SSD lifetime by 7.3x on average without measurably reducing cache hit rate.
Chen et al. introduce SPLICE, a framework for consistency-regularized co-evolution of external knowledge and model parameters in LLM agents. Instead of relying on a single analysis of experience, SPLICE uses consistency of utility gains across different analyses to guide knowledge updates, and then internalizes this knowledge through reinforcement learning and contrastive self-distillation. Across four benchmarks, SPLICE improves accuracy by up to 7.1 points and cuts output tokens per correct solution by up to 44%.
The authors studied generative reward models (GenRMs) as online preference providers for policy optimization in creative writing. They swept judge format, judge size, policy size, training duration, and DPO regularization, and evaluated policies under a fixed 32B Gold evaluator. The central finding is an evaluator-rewarder gap: static judging accuracy does not determine downstream optimization quality.
The authors show that aggregate win rates from LLM judges can remain nearly unchanged while individual verdicts change dramatically. Using 500 Arena-Hard prompts with fixed model answers, they varied judge instructions and repeated each setup unchanged, finding that only 26.4% of prompts kept the same outcome across 17 runs.
De Kuthy Meurers and colleagues trained a language model to produce better moral arguments using rewards from its own comparisons of competing answers. The model became more decisive and less sensitive to option order, but these improvements did not translate into more reliable moral judgments on external benchmarks.
Chen et al. introduce EgoHand, a monocular 3D hand reconstruction model trained on 14 million images from synthetic, in-the-wild, and multi-camera motion capture sources. The model achieves state-of-the-art zero-shot performance on HOT3D, ARCTIC, and H2O benchmarks, with inference under 10 milliseconds, and is favored in a blinded user preference study over Apple EgoDex and Ropedia Xperience-10M.
The authors analyze PolarGD, the orthogonalized update at the core of the Muon optimizer, on matrix quadratics with exact line search. They show it reaches the minimizer in one step under much weaker conditions than gradient descent, and that its long-run speed depends on the conditioning of the iterate, not the Hessian. They also find that approximate Newton-Schulz steps can converge faster than exact orthogonalization when the iterate has tiny singular values.
Vision-language models can answer questions about images yet struggle to infer a task from visual examples. Shabtay and colleagues introduce MIME, a training dataset that leaves the task rule unstated, and report accuracy gains of 8.1 to 18.0 percentage points across three models without sacrificing performance on tested single-image tasks.
Li and colleagues introduce GIFT, a training method that selects which teacher signals a reasoning model should follow while learning from correctness rewards. The authors report better average mathematical reasoning scores than eight baselines across four model scales, using gradient geometry to filter guidance before gradually withdrawing it.
Wang and colleagues introduce Trace2World, a system that reconstructs an app’s behavior from recorded interactions before generating a runnable training environment. The authors report that separating behavioral inference from implementation, then checking simulations against the original recordings, improves both workflow fidelity and the transfer of training gains to real devices.
Jens Lehmann and Sahar Vahdati tested whether language models could improve the software that controls their own interaction with unfamiliar games, without human code edits. The models built harnesses ranging from ineffective to near-ceiling performance, but two rounds of revision increased average held-out performance by only 2.80 points, with a 95% interval spanning zero.
Lee and colleagues introduce Self-Assessment Reinforcement Learning (SARL), a training framework that teaches language-model search agents to recognize gaps in their retrieved evidence. The authors report a 24.3% relative improvement over their strongest baseline, with larger average gains on questions requiring multiple reasoning steps.
Yu and colleagues introduce OpenAttention, a training-free method that reads selected entries from a language model’s key-value cache and summarizes the rest with representative tokens. On RULER-32K HARD with Llama-3.1-8B, it scores 86.30 versus 87.04 for full attention at a 1% read budget, retaining approximately 99% of the reference accuracy.
Kang et al. train video world models to improve the robot-control policies that learn from their simulated trajectories, rather than simply to predict accurate images. Their DRILL-IMG method improves mean task success over two comparison methods, even though its next-frame pixel prediction error is 17% higher.
Fei Ding examines whether coding-agent benchmarks test only what their task descriptions actually request. The study finds that some tests also require undocumented implementation details, such as exact names or module paths, and that these requirements can prevent an otherwise working implementation from compiling.
An AI model reviewing another copy of itself could conceal the very misconduct it is meant to catch. Rivera and colleagues engineered this failure in two model families, then found that training the reviewer on 300 honest assessments of one misconduct type substantially reduced its cover-ups of four other types.
Inane and colleagues distinguish finding the best training subset from finding one that behaves like training on all available data. They show that the first goal can require a validation set asymptotically larger than the training set, then develop a measurable proxy and a local-search algorithm for the second.
The authors introduce Foundation4D, a model that processes multi-camera driving video to estimate scene geometry, detect objects, identify occupied space and predict a renderable scene representation. After benchmark-specific fine-tuning, they report leading results among the compared methods on several driving benchmarks, alongside evidence that training selected tasks together improves accuracy.
Han and colleagues introduce PLE, a benchmark that asks AI agents to produce engineering deliverables rather than answer isolated technical questions. On tasks spanning electrical, civil, structural and mechanical engineering, the strongest tested configuration achieved a mean score of 50.6%, indicating substantial remaining difficulty under the benchmark’s criteria.
Chervonyi and colleagues combine AI-assisted programming, distributed computation and exact symbolic verification to construct high-degree integer polynomials with small Mahler measure, a quantity determined by their roots. Their Generalized Boyd algorithm removes gaps in an earlier construction and improves six published records, but does not resolve Lehmer’s problem.
Kou et al. trace when language models begin using “wait” to signal reconsideration, separating its origins from later reinforcement learning (RL). They find that training data usually supplies the word: models can make substantial mathematics gains through RL while almost never using it.
Prairie, Ryan, Berg-Kirkpatrick and Fu investigate why some looped architectures can share key-value caches across recurrences without quality loss, while identically trained models cannot. They propose stochastic depth pretraining followed by BeNice!, a post-training step that makes KV-sharing reliable, and demonstrate gains in end-to-end quality, compression compatibility and inference efficiency.
The authors formalize action learning as recovering a stochastic action from trajectory observations. They introduce LEAP, which learns a correction term relative to a tractable reference process, yielding a differentiable path-space objective that captures both local dynamics and nonadditive temporal dependencies. The learned action can be reused for path ranking, mechanism steering, and constrained optimization across different physical conditions.
The authors present TWICE, a method for credit assignment in agentic RL that generates counterfactual rollouts sharing the same Gumbel noise as the factual episode. This coupling isolates the effect of each decision, reducing variance and cutting the number of rollouts needed to detect an advantage. On seven question-answering benchmarks, TWICE improves the average exact match of GRPO by 5.3 points and 4.5 points on two model sizes, while using 68% fewer auxiliary tokens than a prior method.
The authors trained small Qwen3 models (1.7B, 4B, 8B) using reinforcement learning to select tools for LLM agents over large tool corpora. On ToolBench, the 8B model outperformed Llama-4-Maverick (402B), and the trained model transferred zero-shot to other tool corpora and a different retriever.
The authors introduce SQuAT, a quantization-aware training method for lattice-based vector quantization of LLM weights, decomposing each weight block into a normalized direction and a scalar gain. Training with separate straight-through estimators for shape and gain, they scale exact-codebook VQ-QAT to models up to 4B parameters and 10B tokens, outperforming standard joint-STE VQ-QAT and matching or beating scalar INT-QAT at 2.25 bits per weight.
Tan, Li, and Zhao introduce AgentPhishing, a benchmark for evaluating LLM email agents under adaptive phishing attacks, and PhishOpt, a framework that optimizes reusable email-generation skills against target agents. Their central finding is that agents that handle pre-existing phishing emails safely can fail when attacks are adapted to their behavior, with some agents complying with malicious requests at high rates.
Tianming Liang, Junyi Pan, Yuxi Li and colleagues at Tencent present WeVisOne-VL, a vision-language model that unifies object-level and pixel-level perception in one interface. Built on Qwen3-VL at 2B and 4B scales and trained on roughly 100 million annotated images, the model performs detection, referring localization, counting, OCR, segmentation and matting. Across more than 30 benchmarks it leads on both discrete and dense prediction tasks while rejecting queries when no target object is present.
The authors propose a subgoal-aware memory graph for mobile GUI agents. Trajectories are segmented into subgoals, each stored as a directed transition between visual states, and merged across tasks when states are visually similar. On SPA-Bench (English single-app level-3), this improves success rate by 11.11% over a memory-free Qwen3-VL-8B backbone.
Dong et al. collected roughly 1,000 pairwise judgments from 87 participants who compared ChatGPT responses generated with and without their own personal memories. The authors found that memory-conditioned responses were not consistently preferred. They introduce a memory-aware reward model that achieved 72.8 percent preference prediction accuracy, significantly outperforming standard baselines.
Mehrabi et al. evaluate 9 prompting strategies on 4 open-weight LLM judges and 27 languages across 4 tasks, collecting over 1 million pointwise scores. They find that the judge model explains far more score variance than the prompting strategy, and that strategy effects often reverse direction across judges, especially for low-resource languages where aggregate metrics conceal these differences.