The paper introduces SEAC, a method that augments a fixed fact verifier with a lightweight evidence‑acquisition policy to decide when to stop or request more evidence. SEAC evaluates each claim by retrieving evidence, running the verifier, and estimating a clarification‑utility score that balances the benefit of additional evidence against its cost. Experiments on SciFact, LIAR, and FEVER show that SEAC improves mean accuracy and macro‑F1 compared to existing fact‑verification and active‑retrieval baselines.
By Yuhang Dong, Jinghong Li, Xinyue Wang, Yuanning Cui
RSIV-Harness is a framework that keeps language model weights fixed while recursively improving the surrounding harness based on failure traces, using a four-layer evaluation to ensure stability. It addresses challenges in open-domain conversational text-to-visualization, such as ambiguous user requests, data source discovery, and iterative chart refinement. The introduced ODAC-Vis benchmark evaluates these capabilities across 640 tasks, and experiments show significant gains in executability and chart structure after a single RSIV-Harness iteration.
By Jiawei Lu, Chen Zhang, Jinwei Lu, Yuanwei Song, Yuanfeng Song, Raymond Chi-Wing Wong
CLOSURE introduces a two‑task framework for multivariate time‑series anomaly detection that separately evaluates recoverability and cross‑channel consistency. The recoverability task uses a diffusion reconstructor conditioned on a discrete memory of normal temporal bases, while the consistency task employs a relation‑closure operator that predicts each channel from its peers using sparse signed gates. Experiments demonstrate that combining these two scores outperforms recent reconstruction, frequency, and topology‑based detectors on public benchmarks, highlighting the importance of assessing consistency alongside recoverability.
By Fan Xu, Hao Wu, Zhi Song, Hanbo Huang, Wenzhuo Ma, Nan Wang, Wei Gong, Xibin Zhao
EvoBench is a diagnostic benchmark framework designed to reveal where and when machine‑learning models for environmental variables fail, rather than relying on a single average accuracy score. It tracks model performance over time and across spatial extrapolation, highlighting issues such as non‑stationarity, poor performance at unmonitored sites, and under‑prediction of rare events like floods. In a case study on daily streamflow prediction across 9,000 U.S. basins, EvoBench shows that no model dominates all scenarios and that targeted fixes can mitigate identified weaknesses.
By Yiming Sun, Yuan Zhang, Qi Cheng, Yingda Fan, Runlong Yu, Yiqun Xie, Xiaowei Jia
The paper introduces the Continual Learning Exam (CLE), a benchmark designed to evaluate how large language model (LLM) agents improve over time by learning from experience. CLE consists of 120 human-authored tasks organized into five-level ladders, totaling 600 levels across 89 professional roles in 12 domains, where agents must pass each level to unlock the next. Experiments with 17 frontier LLM configurations show that providing diagnostic feedback and allowing one revision improves median completion by 5.3 percentage points, yet the best system only clears 47.7% of levels and completes all five levels in 23.3% of tasks, underscoring the challenge of sustained learning from experience.
By Haoyu Zhao, Sitian Tu, Yuxuan Huang, Huichi Zhou, Zhangyue Yin, Jiahao Huang, Jingyuan Xu, Tianmiao Shi, Jin Shi, Nianxi Zhang, Yangmiao Zhang, Zhuolang Chan, Huiyan Hu, Chenchao Sheng, Yaqing Luo, Yue Chen, Huaxiu Yao, Meng Fang, Jun Wang
The paper introduces LARVA, a framework that distills multi‑round video agents by jointly learning latent reasoning and visual observations. Instead of using textual captions, an observer sends compact continuous representations to a reasoner, which then selects the next action or produces an answer. The approach incorporates action and answer supervision through this interface and uses observation reconstruction and reasoning‑distribution alignment as auxiliary signals, achieving strong results on five video‑understanding benchmarks with 4B and 7B models.
By Junzhang Liu, Hao Yuan, Qi Zhang, Yiming Wu, Long Chen, Zhecan Wang
The paper introduces ACSAN, an All‑Convolutional Self‑Attention Network that replaces the fully‑connected matrix multiplications of Vision Transformers with convolution operations. Its core module, Dynamic Convolutional Self‑Attention (DCSA), uses multi‑scale dilated convolutions to produce Query, Key, and Value tensors, and combines spatial and channel self‑attention into a single output. Experiments on CIFAR‑10, CIFAR‑100, and ImageNet‑1k demonstrate that ACSAN attains competitive or superior classification accuracy to ViT while using comparable or fewer parameters, and ablation studies confirm the positive contribution of each component.
By Zimian Wei, Tianqi Zhang, Hengyue Pan
The paper investigates how deleting portions of a software engineering agent’s past trajectory affects performance and cost. Across 9,400 runs with two state‑of‑the‑art agents, it finds that large deletions (even up to 100% of reasoning text) can maintain overall effectiveness if applied with a delay, but immediate deletion harms performance. The study also shows that cost savings depend on the pricing model: a flat model yields savings in 26 of 42 cases, while a caching‑aware model actually increases costs because it breaks a low‑rate cached prefix.
By Fernando Vallecillos-Ruiz, Islem Bouzenia, Michael Pradel, Leon Moonen
OmniShowcase introduces a method for generating omnidirectional videos of a specific subject from a handful of multi‑view images. It combines an automated data‑synthesis pipeline that creates training videos with diverse camera trajectories driven by the subject’s geometry, and a geometry‑conditioning technique that supplies the model with surface‑normal maps reconstructed from the reference views. The authors release Showcase‑4K, a large corpus of 3D‑asset‑based videos with 12 camera motions and 88 trajectory configurations, and ShowcaseBench, a held‑out evaluation set with source meshes and camera parameters.
By Junyeong Ahn, Geonwoo Kim, Kinam Kim, Gihwi Kim, Hoonjin Jung, Hyojin Jang, Jaegul Choo
GPT4o-Receipt is a benchmark dataset of 1,235 receipt images that pairs GPT‑4o‑generated receipts with authentic ones from established datasets. The study evaluates five state‑of‑the‑art multimodal LLMs and a 30‑annotator crowdsourced perceptual study, revealing that humans are better at spotting AI artifacts yet perform worse at binary detection than models like Claude Sonnet 4 and Gemini 2.5 Flash. The paradox is explained by the fact that the dominant forensic signals—arithmetic errors—are invisible to visual inspection but can be quickly verified by LLMs.
By Yan Zhang, Simiao Ren, Ankit Raj, En Wei, Dennis Ng, Alex Shen, Jiayu Xue, Yuxin Zhang, Evelyn Marotta
DiagnosticIQ is a benchmark that tests large language models (LLMs) on translating engineer‑written symbolic rules into maintenance actions for industrial assets. It contains 6,690 multiple‑choice questions derived from 118 expert rule‑action pairs across 16 asset types, and includes five variants that probe different failure modes. The benchmark evaluates 29 LLMs and four embedding baselines, showing that while top models perform near the frontier, they exhibit brittleness under distractor expansion and structural perturbations.
By Devin Yasith De Silva, Dhaval Patel, Christodoulos Constantinides, Shuxin Lin, Nianjun Zhou, Dharmashankar Subramanian, Paul J Adams, Sal Rosato, Nicolas Constantinides, Deborah L. McGuinness, Jayant Kalagnanam
The paper introduces Multivariate Time Series Super-Resolution Forecasting (MTS‑SRF), a method that predicts fine‑grained future values from coarse historical data. It proposes the TimeMagnifier framework, which includes Fourier Anchor to handle data sparsity, De‑noise Only to refine predictions along the resolution axis, and banded DTW Alignment to correct cross‑resolution misalignment. On a new benchmark of ten datasets across five domains, TimeMagnifier outperforms 15 baselines, reducing mean‑squared error by an average of 12.3% and improving existing forecasters by 5.1% to 9.0% when its components are used as plug‑ins.
By Shuhao Li, Weidong Yang, Yue Cui, Fan Zhang, Lu Qin
The paper introduces SMART‑Med, a local multi‑agent system for medical reasoning that learns when to collaborate instead of always engaging all agents. By using an uncertainty‑driven routing mechanism, it selectively recruits expert agents only for uncertain queries, thereby converting dense collaboration into query‑adaptive coordination. The authors also present a new multimodal dataset, MedVR‑3K, and demonstrate that SMART‑Med matches or surpasses existing multi‑agent approaches while cutting inference time by over 3.8× and token usage by 72%, with a 6.7% average accuracy gain across four VQA benchmarks.
By Jiacheng Hou, Xueliang Cui, Dan Lu, Ruxin Wang
EnvLunch proposes a warm‑up training stage that uses the executable environment itself as a source of verifiable tasks, eliminating the need for human‑written or model‑generated task descriptions. By sampling valid action trajectories, executing them, and training agents to recover equivalent trajectories from final observations, EnvLunch achieves over 9‑point performance gains on downstream tasks and reduces the required annotated examples by 8 to 32 times compared to baselines. This approach also improves fine‑tuning on benchmarks beyond data analysis, offering a more efficient starting point when annotated tasks are scarce.
By Wei Liu, Min Gao, Yali Du, Yulan He
R2A‑Sales is a new benchmark that pairs static policy probes with multi‑turn sales scenarios to evaluate whether large‑language‑model agents can translate stated rules into compliant, tool‑grounded actions under commercial pressure. The benchmark includes a ledger‑verified interactive arm, 24 policy atoms, 42 semantic cores, and 2,016 matched pairs across three Chinese‑language sales domains. Results show high static compliance (≈90–99%) but low action compliance (≈3–17%), revealing a substantial rule‑to‑action gap of 80–92 percentage points.
By Yihua Qin (Tsinghua University, Futura Lab, Futura.AI), Jiajie Jerry Zhao (Tsinghua University, Futura Lab, Futura.AI)
SAP-ViT is a Vision Transformer that replaces the traditional global classification token with direct aggregation of last‑layer patch embeddings, enabling class predictions to be decomposed into patch‑level contributions during the forward pass. The model incorporates sparsity regularization to produce concise, readable evidence and introduces an input‑space back‑mapping mechanism to identify which input patches influence the output. Experiments on diverse medical imaging datasets show that SAP‑ViT retains strong classification performance while delivering consistent, interpretable patch‑level explanations, with ablation studies confirming the benefits of both sparsity and back‑mapping.
By Wei Zhang, Wolfgang Nejdl, Ming Tang
POD‑RoPE introduces a rotary position encoding that represents token observation footprints as distributions over camera centers, viewing directions, and sampling times, enabling relative relationships between observations to be captured through expected rotations. The method supports Q/K matching and coherence‑dependent V/O transformations, allowing observation geometry to guide information selection and content aggregation without increasing sequence length. Evaluations across robot policy architectures, simulation, real‑world manipulation, and auxiliary tasks show performance gains and computational efficiency, with significant speedups over existing RoPE variants.
By Xiaodi Huang, Zilin Wang, Jinpeng Liu, Pengchuan Xiao, Zhenlei Shao, Zhang Zhang, Shuxiang Lu
The paper introduces BAC-Bench, a benchmark comprising 46 real-world broken access control vulnerabilities from six popular web applications across four programming languages. It evaluates eight base large language models (LLMs) and two commercial LLM agents, showing that single-turn LLMs repair about 38% of the vulnerabilities, while minimal scaffolds can raise this to 66% and the best LLM achieves 75%. Some LLM-generated patches even outperform official fixes and one patch closed a dormant zero‑day vulnerability.
By Mingrui Liu, Liang He, Xinchu Wang, Yeting Li, Dongxu Han, Yunpeng Li, Yuling Liu
The paper describes the Cosmos LLM Gateway, a policy‑driven, multi‑vendor gateway built on Envoy that manages enterprise LLM traffic at PayPal. It separates latency‑sensitive control decisions into an external service while handling wire‑format translation and token metering in‑proxy via a Rust WebAssembly filter, ensuring minimal overhead. Governance is declarative, managed through Open Policy Agent, and includes features like data‑loss‑prevention, budget scopes, and a control plane for role‑scoped authoring and approvals. The gateway currently serves over 12,000 developers, processes roughly 200 billion tokens per day across 100+ models, and maintains 99.999% availability with only a few milliseconds of added latency.
By Mithun Sasidharan, Srinivasan Manoharan, Rupesh Thakur, Abhinav TK, Ashwin Hegde, Anand Gautam
The paper introduces UTA, a Universal Training Adaptive learning rate decay that treats learning rate decay as a closed‑loop shape‑and‑scale problem rather than a fixed offline trajectory. By formulating the decay in continuous time, UTA adapts to the training state, achieving near‑optimal progression shape and learning‑rate scale. Experiments show consistent improvements on language models (LLaMA 60M‑1B) and vision models (ResNet50, ViT‑B) on ImageNet, with minimal overhead and strong robustness across various hyper‑parameters and fine‑tuning settings.
By Anda Tang, Junjie Wang, Xuanlin Yang, Cong Fang, Zhouchen Lin