The paper investigates how context compaction—losing information to free up a language model’s context window—affects performance on long‑horizon agentic tasks. Using the OpenCode scaffold and five compaction strategies, the authors evaluate three frontier models on PaperBench, ProgramBench, and PostTrainBench, finding that compaction thresholds largely do not hurt performance for 1M‑context models, but can degrade GPT‑5.3‑Codex on certain benchmarks. The study also shows that compaction can silently drop critical task constraints, especially on ProgramBench and real user‑agent sessions from SWE‑chat, yet current benchmarks remain largely compaction‑friendly because objectives are fixed and progress is stored externally.
By Jeanne Salle, Hardik Bhatnagar, Matthias Bethge, Sahar Abdelnabi, Maksym Andriushchenko
arXiv:2610.31791v1 Announce Type: cross
Abstract: Automated research is largely evaluated by benchmark performance, but the value of research rarely reduces to a single score. Humans judge a work's h...
By Sachin Konan, Jonathan Liu, Zhuang Liu
The paper investigates how the dimension of attention heads influences long‑context retrieval in language models. Experiments show that while pretraining loss and short‑context tasks are largely unaffected, wider heads significantly improve retrieval performance at fixed model sizes. Ablation studies indicate that this benefit persists even when attention weights are frozen, suggesting that wider heads enhance adaptation of other model components such as embeddings and MLPs.
By Seiya Tokui, Shoichiro Yamaguchi, Sosuke Kobayashi
The paper introduces TopKLoRA, an interpretable parameter‑efficient fine‑tuning method that activates only the top‑k adapter latents per token. Using this sparse LoRA, the authors fine‑tune large language models on a poisoned dataset, identify and ablate the minimal circuit responsible for a backdoor, and show that the resulting circuits are an order of magnitude smaller than dense LoRA baselines. They also demonstrate how to construct model organisms with localized circuits via gradient‑routing, supporting the development of interpretable‑by‑design architectures.
By Marek Masiak, Andrzej Szablewski, Nicola Cancedda, Constantin Venhoff
arXiv:2610.31695v1 Announce Type: cross
Abstract: Safety guards check content before an LLM application acts on it, and most paste that content into a plain-text prompt. Content that imitates the pro...
By Utkarsh Kanwat
arXiv:2610.32075v1 Announce Type: new
Abstract: Feed-forward neural renderers synthesize global illumination across novel 3D scenes without per-scene optimization or runtime ray tracing. Existing tra...
By Sarah Zhang, Zhiyang Dou, Minghao Guo, Benjamin Tod Jones, Wojciech Matusik
DMAIC-Bench is a new benchmark that tests large‑language‑model agents on their ability to exercise self‑restraint—deciding when to continue, redirect, or stop a multi‑stage analytical task based on intermediate evidence. It contains 97 seeded tasks across 15 industrial cases in four sectors, with paired counterfactual variants that require opposite actions to distinguish warranted restraint from indiscriminate stopping. Evaluation involves a private scorer for execution traces and independent expert review, revealing that while agents often route correctly on GO decisions, they frequently act on invalid evidence at NO‑GO decisions, with a 47% average error rate.
whyItMatters":"The benchmark provides a systematic way to measure whether agents can appropriately use evidence to control their own actions, a capability that current benchmarks largely overlook."
By Shan Peng, Shuaiqi Cheng, Miao Peng, Chenyi Zi, Jia Li
Efficient Zero‑Shot Video Anomaly Detection via Signal‑Aware Adaptive Fusion (SAAF) introduces a framework that leverages dual‑signal adaptive sampling—delta and drift signals—to detect sudden and gradual anomalies without training data. It employs a threshold‑gated cascaded architecture that first applies lightweight frame‑level scoring before invoking costly video‑level analysis only when needed, and a signal‑aware weighted fusion mechanism that adapts score combination based on the triggering signal. Across four benchmarks, SAAF achieves state‑of‑the‑art accuracy on XD‑Violence, MSAD, and ShanghaiTech while delivering significant throughput improvements over prior zero‑shot methods.
By Khaled Waleed Dawoud, Muhammad Haris Khan
arXiv:2610.32964v1 Announce Type: new
Abstract: Few-shot industrial anomaly detection requires modeling category-specific normality from only a handful of defect-free images. Common approaches tie in...
By Bernardo Perrone Ribeiro, Erkut Akdag, Egor Bondarev
Softly Symbolified Kolmogorov–Arnold Networks (S2KAN) integrate symbolic primitives into the training of Kolmogorov–Arnold Networks, allowing each activation to be expressed as a sparse combination of symbolic and dense terms. The sparsification is differentiable and guided by a minimum description length–inspired objective, enabling end‑to‑end optimization. Across real‑world prediction tasks, dynamical‑system forecasting, and synthetic symbolic benchmarks, S2KAN achieves higher symbolic fidelity, often with smaller models while maintaining competitive predictive accuracy.
By James Bagrow, Josh Bongard
arXiv:2605.07775v2 Announce Type: replace-cross
Abstract: Balancing exploration with exploitation is a core challenge in black-box optimization and reinforcement learning. We introduce POETS ($\textb...
By Nicolas Menet, Andreas Krause, Abbas Rahimi
arXiv:2605.18971v2 Announce Type: replace-cross
Abstract: What determines the quality of a tabular foundation model? Unlike language or vision, tabular foundation models acquire their inductive biase...
By Mohamed Bouadi, Nassim Bouarour, Varun Kulkarni, Shivam Dubey, Aditya Tanna, Vinay Kumar Sankarapu
arXiv:2608.03259v3 Announce Type: replace-cross
Abstract: Time-series foundation models (TSFMs) are increasingly evaluated on broad benchmarks that apply uniform conventional forecast evaluation metr...
By Jaehoon Lee, Jun Seo, Seunghan Lee, Tae Yoon Lim, Dongwan Kang, Hwanil Choi, Minjae Kim, Sungdong Yoo, Junhyeok Kang, Sangjun Han, Soonyoung Lee, Hao Ni, Srijan Sood, Stefan Zohren, Chanyeol Choi, Yongjae Lee, Zhangyang Wang, Lukasz Szpruch, Benjamin Bowring, Wonbin Ahn
AU-Harness is an open-source toolkit designed to streamline and standardize the evaluation of Audio Language Models (LALMs). It addresses three key limitations of existing frameworks: slow processing pipelines, lack of multi-turn dialogue support, and absence of a scalable, unified evaluation system. The toolkit achieves a 77% reduction in evaluation time, enabling large-scale studies and providing standardized protocols for fair model comparison and in-depth analysis of audio reasoning capabilities.
By Hoang Nguyen, Sidharth Surapaneni, Akshay Kalkunte, Jash Mehta, Aman Tiwari, Oluwanifemi Bamgbose, Khyati Mahajan, Jash Shah, Shruthan Radhakrishna, Sathwik Tejaswi Madhusudhan, Vikas Yadav, Sai Rajeswar
MedInsightBench is a new benchmark comprising 332 curated medical cases with annotated insights, designed to evaluate large multimodal models (LMMs) and agent frameworks on their ability to analyze medical image data, ask relevant questions, interpret findings, and synthesize actionable recommendations. The study finds that current LMMs perform poorly on this benchmark due to difficulties in extracting multi-step insights and lacking medical expertise. To address this, the authors propose MedInsightAgent, an automated agent framework with three modules—Visual Root Finder, Analytical Insight Agent, and Follow-up Question Composer—that improves LMM performance on MedInsightBench.
By Zhenghao Zhu, Chuxue Cao, Sirui Han, Yuanfeng Song, Xing Chen, Caleb Chen Cao, Yike Guo
arXiv:2601.03018v3 Announce Type: replace-cross
Abstract: While Large Language Models (LLMs) have shown strong performance on clinical text understanding, they struggle with longitudinal prediction t...
By Choonghan Kim, Hyunmin Hwang, Hangeol Chang, Jaemin Kim, Jinse Park, Jae-Sung Lim, Jong Chul Ye
arXiv:2605.25773v3 Announce Type: replace-cross
Abstract: Efficient benchmarking techniques aim to lower the computational cost of evaluating LLMs by predicting full benchmark scores using only a sub...
By Sam Bowyer, Acyr Locatelli, Kris Cao
arXiv:2609.36492v2 Announce Type: replace-cross
Abstract: We benchmarked vision-language models (VLMs) on the decisions annotators take when inspecting electron microscopy images in connectomics: syn...
By Yicong Li, Junjie Wang, Leander Lauenburg, Ella Hugie, Alexandra Irger, Wanhua Li, Donglai Wei, Hanspeter Pfister
Language models now answer questions alongside other claims: a note in the system prompt, a message from another agent, their own earlier answer. We ask what it takes to talk a model out of an answer...
Effective collaboration requires language models to understand what others know and believe, and to act accordingly. Models that assume others share their information may omit necessary communication...