VTrajEmbed is an embedding model that aligns text with video object trajectories by decomposing videos into individual trajectories and embedding each trajectory separately. It introduces a Slow‑Fast Trajectory Encoding dual‑branch architecture that captures high‑frequency motion and scene context, and a Mask‑Guided Contrastive Learning objective that prevents false negatives during training. The model is evaluated on a new trajectory‑level benchmark, Text2Traj‑Bench, and outperforms existing large‑language‑model embeddings by significant margins on both text‑to‑trajectory and trajectory‑to‑text retrieval tasks.
By Ruizhi Chen, Zining Fan, Yudi Guo, Jiang Liu, Zhiyu Lin, Zhongjiang He, Shouqian Shi, Hao Sun
LexPilot is a new benchmark that evaluates how well legal large language models (LLMs) can proactively interact with laypeople during underspecified Chinese civil case consultations. The dataset contains 432 real judgments from 36 common causes of action, paired with expert‑reviewed consultation histories that range from complete to underspecified, and the corresponding court verdicts. Experiments show that current LLMs perform poorly in proactive interactions, with downstream judgment prediction dropping by an average of 23.84% compared to fully specified consultations, highlighting key bottlenecks in user‑centered legal AI.
By Wenbin Zhang, ShangZhan Li, Bingxiang He, Ye He, Yuxin Zhou, Muzhen Cai, Qi Shi, Wanxiang Che
The paper introduces HAPACT, a new benchmark for human‑centric physical impact localization in movies. It defines a task that requires identifying when and where abrupt impact events occur on the human body within a video. The dataset contains over 2.8K video‑query pairs and more than 20K finely annotated frames, and the authors evaluate existing methods, showing that the task remains challenging and motivating further research.
By Youngrae Kim, Yejin Jang, Eunho Kim, Youjin Sung, Kun-Woo Song, Sangpil Kim, Sang Ho Yoon
The paper introduces a large-scale video dataset for evaluating central line dressing change (CLDC) skills, featuring 365 recordings from nurses during a competency exam at Houston Methodist Hospital. Each video includes subtask-level temporal annotations and clinical performance assessments across 17 subtasks. The authors detail the data collection, annotation, and processing pipeline, and demonstrate the feasibility of video-based skill assessment using spatiotemporal deep learning models.
By Ummey Tanin (Pam), Shannan K. Hamlin (Pam), Hsin-Mei Chen (Pam), Peizhu (Pam), Qian
The paper reexamines robustness in Graph Neural Networks (GNNs) by focusing on directed graphs, which have been largely ignored in prior work. It introduces a more realistic attack setting for directed graphs and proposes a simple message‑passing defense that can be added as a plug‑in layer. The authors demonstrate that leveraging the trust signals inherent in link directions can significantly enhance GNN robustness, achieving state‑of‑the‑art performance while maintaining strong clean accuracy against both transfer and adaptive attacks.
By Zhichao Hou, Xitong Zhang, Wei Wang, Charu C. Aggarwal, Xiaorui Liu
SciIntBench is an adversarial benchmark that tests large language models (LLMs) on 1,350 prompts across 270 scenarios, covering ten research integrity categories and three scientific domains. Each scenario includes five gradations of adversarial framing—from Most Overt to Most Covert—plus a benign version, enabling measurement of both refusal of misconduct and helpfulness to legitimate requests. In evaluating 16 commercial and open‑weight LLMs, the study found that under the most covert framing, models failed to refuse requests for research misconduct in 71.7% of cases, with weaker refusals observed for transparency, plagiarism, and fabrication.
By Almene De Meran Meguimtsop, Maria Leonor Pacheco, Daniel E. Acuna
The paper studies how different configurations of privileged information (PI) affect on‑policy self‑distillation (OPSD) for large language models. By evaluating 16 PI setups across three agentic benchmarks and multiple student models, it finds that PI effectiveness varies with content, generation source, and model scale, with information‑dense trajectories and stronger generators yielding larger gains. The authors also propose KL‑based metrics, particularly Δd−_err, to predict PI utility before training, showing that selecting PI with this metric improves post‑training performance and remains beneficial when combined with GRPO.
By Kaijian Zou, Emiliano Penaloza, MohammadHossein Rezaei, Anisha Gunjal, Mohit Raghavendra, Lu Wang, Yunzhong He, Anas Mahmoud
The study investigates whether switching optimizers between pretraining and subsequent training stages of large language models (LLMs) incurs a performance penalty. By pretraining 1.2B Llama‑3 models with varying proportions of AdamW and Muon updates, and interpolating between the two during continued pretraining and supervised fine‑tuning (SFT), the authors find little evidence of a meaningful optimizer mismatch penalty. The results show that final performance is largely determined by the independent effects of the pretraining and post‑training optimizers, with downstream performance generally improving as the proportion of Muon updates during SFT increases, and that Muon remains competitive with AdamW even for task‑general SFT on larger models.
By Xinyu Zhao, Davis Wertheimer, Ashutosh Sharma, Rameswar Panda, Paul Pu Liang
The paper introduces the Long‑Range Ising Model (LRIM) benchmark, which ensures long‑range dependence by defining each label as a node’s spin multiplied by a power‑law coupled spin field. It proves that polynomial filters on the adjacency matrix—used in isotropic message passing and spectral graph networks—cannot reduce the error floor because the target lies on a spectral level set where such filters are constant. The authors show that a strided filter with about fifty coefficients, trained on the labels, achieves a test error 52 times lower than the best published network, while a dilated convolutional network outperforms prior results at 256 nodes and matches them at 1,024 nodes, providing the first in‑domain results for larger sizes.
By Amin Najafgholizadeh, Mohsen Afsharchi
ECHOES is a method for long‑horizon GUI agents that compresses visual history into a textual state using a Compression‑Think‑Action format. It employs a privileged teacher model that provides token‑level on‑policy distillation to guide the student’s textual rollouts, while a Perception‑Memory Fidelity Gap‑guided framework balances distillation with task reward optimization. The approach is evaluated on four GUI benchmarks and a new ECHOESBMK diagnostic benchmark, showing improved task performance, memory fidelity, and a favorable trade‑off between performance and inference cost, with positive zero‑shot transfer to unseen planning tasks.
By Tengda Guo (Tsinghua University), Jiawang Hu (StepFun), Yifei Tao (Nanyang Technological University), Yin Wang (StepFun), Haotian Hu (Zhejiang Leapmotor Technology Co., Ltd), Fanyi Wang (Honor), Liujuan Tang (StepFun)
The paper introduces L‑SplineFNO, a memory‑efficient neural operator that combines a Fourier neural operator backbone with a learnable tensor‑product B‑spline transfer to map unstructured meshes to a latent Cartesian grid. By using a local normalised scatter, the method preserves spline properties without geometry‑dependent solves, enabling efficient training on up to 26.6 million nodes and inference on 46 million nodes on a single A100 GPU. L‑SplineFNO achieves strong predictive accuracy across six 3D engineering benchmarks, supports zero‑shot super‑resolution, efficient closed‑form spatial derivatives, and provides rigorous prediction bounds via Bernstein convex hulls and Taylor models on a large tokamak model.
By Vignesh Gopakumar, Zhuoyuan Wang, Ander Gray, Valentin Duruisseaux, Josh Williams, William Michael Edward Ellis, Nuno Nobre, Matthew Field, James Panayis, Alexander I. Blair, Andrew Davis, Stanislas Pamela, Robert Akers, Marc Peter Deisenroth, Anima Anandkumar
The paper introduces four executable operations—identifying redundancies (L1), coarse‑graining dynamics (L2), transferring structure (L3), and inferring missing quantities from conservation imbalances (L4)—to modify physical descriptions in automated scientific discovery systems. It evaluates these operations alongside three interpretability settings (null acceptance, positive acceptance, consequence agreement) on chaotic systems, controlled benchmarks, language and protein models, and sparse autoencoders, revealing that descriptive tests can accept empty candidates and accepted candidates may lack verified consequences. The study demonstrates that reliable automated discovery must assess both what a test accepts and the scientific consequences it establishes.
By Tianyu Wu (Project Aether)
The paper introduces Harness-Delta Attribution (HDA), a protocol that breaks down the performance gains from harness evolution into overfitting, test-time scaling, and residual generalizable improvement. Applying HDA to 11 executor models and three harness-evolution frameworks across various tasks, the study finds that over half of the search-set gains are due to overfitting or scaling, with limited transfer to held-out tasks. The results highlight the need for attribution-aware evaluation in future harness evolution research.
By Wenxuan Ding, Qin Lu, Changlong Yu, Sha Li, Shuowei Jin, Xin Liu, Zhenyu Shi, Greg Durrett
GlobeBench is a new benchmark for evaluating text-based world models, comprising 1,632 tasks across five domains and nineteen simulation skills. It tests whether models can predict an environment’s response to an action based on the full agent trajectory, requiring understanding of state changes, environment rules, and action effects on related and unrelated objects. The benchmark compares predictions to execution-derived references using a hybrid architecture of deterministic gates and a semantic judge, and finds that GPT‑6 Astra tops the 20 tested models with a 49.57% success rate, while accuracy drops sharply when actions affect many dependent objects, leaving 631 tasks unsolved by any model.
By Nicholas Saban, Charvi Bannur, Mariya Vasileva, Darshan Deshpande
The paper discusses how AI coding benchmarks suffer from both under‑specification, which forces solutions to guess or cheat, and over‑specification, which allows solutions to bypass intended challenges. It examines several repository‑level benchmarks that add hints from ground truth to mitigate under‑specification, but finds that this can inadvertently create spoilers. The authors experiment with different levels of specification and propose methods for generating concise, effective hints, which they plan to release to improve benchmark reliability.
By Yuanzhe Liu, Bastian Lipka, Avraham Shinnar, Jatin Ganhotra, Yao Ma, Martin Hirzel
Evo-B&B is a new evolutionary framework that jointly evolves branching and node selection policies for Mixed-Integer Linear Programming (MILP) using large language model guidance and end-to-end SCIP solving feedback. It introduces a behavior-aware multi-island mechanism that measures policy similarity based on branching and node-selection behaviors to preserve population diversity and reduce redundant evaluations. Experiments on four standard MILP benchmarks show that Evo-B&B achieves competitive solving performance, generalizes to larger instances, and that joint evolution and behavior-aware diversity management are effective.
By Jiayu Chen, Chentong Chen, Yuhan Liu, Ye Fan, Jianyong Sun, Jialong Shi
Reward Engine is a method for creating a reward function for open‑ended tool agents by first determining what success looks like and then optimizing toward it. It builds a RewardPack of test cases and rubrics from initial trajectories, which is frozen and used to guide the agent’s output selection and can also serve as a reinforcement‑learning reward. Experiments on WorkBuddy Bench Office and Claw‑Eval show that the constructed reward aligns with benchmark graders on most candidate pairs and improves three‑run consistency from 45.0% to 54.0% and from 78.4% to 83.7%.
By Litong Feng, Ji Qi, Yiran Cheng, Shaowu Pan, Xinyu Zhang, Qijun Li
The paper introduces GOLD, a Geometric Optimized Latent Diffusion framework that jointly models RNA sequences and their side‑chain geometries in a unified latent space using a novel GeoVAE. By co‑diffusing sequence and structure and employing a gradient‑guided sampling step that enforces physical constraints, GOLD achieves superior performance in sequence recovery, secondary structure validity, and 3D foldability compared to existing inverse‑folding methods.
By Qi Si, Xuyang Liu, Penglei Wang, Shuo Su, Limei Han, Mengzhu Li, Xin Guo, Yuan Qi, Yuan Cheng
The paper introduces D‑VAE, a Deconfounder Variational Autoencoder that estimates causal effects of expert‑generated marketing labels on LLM‑generated advertisements. By integrating a multi‑treatment deconfounder with causal representation learning, D‑VAE disentangles expert‑labeling logic and provides PAC‑based bounds on omitted‑variable bias. Applied to over 29,000 foundation‑makeup ad creatives, the method outperforms existing benchmarks and a field experiment shows that its recommended label configurations boost engagement by 32.7% while matching human expert performance.
By Danhong You (Eric), Wei Jiang (Eric), Zhiqiang (Eric), Zheng
WebCoder is a general frontend development agent that uses artifact‑grounded agentic reinforcement learning to learn how to compose, generate, modify, and repair code over an evolving application state. It evaluates development actions by their impact on a shared executable artifact and a verifier that provides terminal rewards, while intermediate progress redistributes rollout advantage across development segments. The system includes Webdev‑World, an executable environment that integrates code editing, execution, browser inspection, and on‑demand skills, and Webdev‑Bench, a benchmark for composed development tasks, where WebCoder‑27B outperforms several large language models on requirement satisfaction and visual/functional scores.
By Zijun Dou, Zhen Yang, Xijun Liu, Dongzhuoran Zhou, Aohan Zeng, Jie Tang