The paper introduces Behavioral Localization, a method that identifies a small subset of trajectory positions where a text skill most effectively improves a tool‑using agent’s performance. By comparing the agent’s behavior with and without the skill, the approach pinpoints positions that increase successful token likelihood and meaningfully alter the next token distribution. Using these localized positions to supervise a short soft prompt, the authors achieve significant performance gains on SpreadsheetBench, outperforming full‑trajectory supervision and on‑policy baselines.
By Zijian Jiang, Xilu Wang
GuideEvolve is a self‑evolving agent framework that uses expert guidance and three autonomous regulatory mechanisms—Think Fast and Slow, Throwback, and Termination by choice—to address limitations in existing self‑evolving agents. The framework introduces a task‑specific expert‑evaluation‑driven evolutionary pipeline and integrates the 3Ts mechanisms for cognitive optimization, performance recovery, and adaptive termination. Empirical results on MGSM and MMLU show comparable accuracy to state‑of‑the‑art models, reduced variance, a 100% evolution success rate, and about 50% fewer optimization iterations.
By Meihong Jia, Yasemin Timar, Jonathan Taws
Trace2World is a framework that decouples global behavior reconstruction from implementation to create executable GUI environments from real interaction traces. It reconciles traces into a shared, source‑linked app model, compiles an executable environment, and uses independent replay of source workflows to localize mismatches for bounded implementation repair. In experiments, Trace2World built 50 environments from 1,570 traces and reproduced 91.1% of 200 workflows across ten apps, outperforming a task‑local baseline and improving benchmark success and external GUI benchmark scores.
By Zezhou Wang, Hanzhang Zhou, Panrong Tong, Quyu Kong, Pengxiang Li, Steven Hoi, Zhuzhong Qian, Yue Wang
The paper introduces a unified training framework for automatic PowerPoint generation, combining cross‑validated data synthesis, rule‑based teacher enhancement, and a mixed‑teacher supervised fine‑tuning strategy. It further adds a reinforcement learning component with verifiable aesthetic metrics to refine slide visual attributes. Experiments show significant performance gains on PresentBench, outperforming existing models and improving aesthetic quality.
By Yuhang Xie, Cheng Li, Chenliang Li, Mingyu Xu, Mingwei Sun, Suhang Zheng, Ming Yan, Xiaoxiao Xu, Lin Qu
The paper introduces nnInteractive Agent, an open‑source system that lets general‑purpose AI agents perform 3D image annotation tasks by interacting with the nnInteractive segmentation model. Users issue plain‑language requests, and the agent automatically identifies targets, places prompts, evaluates and corrects the resulting volumetric masks, and can even generate code for additional processing. On a diverse benchmark of CT, MRI, PET, micro‑CT, and insect scans, agents such as Claude Code and Codex achieve mean Dice scores of 0.81–0.84 from the request alone, matching near‑oracle performance when given target locations.
By Fabian Isensee, Constantin Ulrich Harsy, Maximilian Rokuss, Benjamin Hamm, Philipp Vollmuth, Klaus H. Maier-Hein
The paper introduces BELT (BELief-based Termination), a method that decouples outcome prediction from stopping control in large language models (LLMs). BELT learns beliefs about current and future answers from a frozen LLM’s hidden states and applies a linear stopping policy, enabling a single probe to adapt to various token‑price trade‑offs without retraining. Across five out‑of‑distribution math benchmarks, BELT cuts token usage by up to 51% while maintaining accuracy, and its lightweight probe outperforms stronger baselines in both efficiency and parameter count, also transferring effectively to coding and agentic tasks.
By Jingwei Ni, Dmitry Abulkhanov, Vladislav Smirnov, Markus Leippold, Elliott Ash, Timothy Baldwin, Artem Shelmanov, Mrinmaya Sachan
The paper proposes a multi‑agent reinforcement learning framework that dynamically allocates incentives to drivers, encouraging them to take alternative routes. By doing so, it reduces total travel time by up to 16% and CO₂ emissions by 9% under an unlimited budget, while revealing a trade‑off between the two objectives. The study demonstrates that a carefully weighted multi‑objective function can improve both travel time and emissions beyond baseline levels.
By Germ\'an Pardo-Gonz\'alez, Shaghayegh Vosough, Katerina Papadaki, Claudio Roncoli
ReDence is an open‑source hybrid framework that combines a multimodal agentic Investigator model with a specialized Decider model to detect modern social engineering scams. It automates investigative planning, evidence interpretation, and decision‑making across multiple scam categories, achieving 96% accuracy and operating at least four times faster than comparable agentic models. The system outputs structured, evidence‑grounded reports that enable analysts to audit decisions and support mitigation efforts.
By Jaishnoor Kaur, Josef Hor\'a\v{c}ek, Ian Waskom, Sayak Saha Roy
The paper investigates whether performance gains in multi‑agent large language model (LLM) pipelines come from typed contracts between agents or from specialized prompts that enable decomposition. Using a requirements‑engineering task, the authors compare three architectures— a single‑call baseline, a two‑call variant that reuses specialized prompts but passes an unvalidated intermediate, and a pipeline with a Pydantic‑validated typed state—across 18 conditions involving 625 requirements, three small language models, and two languages. Results show that the two‑call, specialized‑prompt variant yields the largest F1 improvement (+0.075), while the typed state offers only a marginal difference (+0.013) and introduces out‑of‑range outputs at agent boundaries, with a small but reproducible increase in inter‑model MoSCoW consistency.
whyItMatters":"The study clarifies that specialized prompts and call structure contribute more to accuracy than typed state validation, informing design choices for reliable multi‑agent LLM pipelines."
By Maria Gabriela Alves Zuppardo, Vinicius Cardoso Garcia
Optimistic Policy Regularization (OPR) is a simple reward‑shaping technique that augments on‑policy agents with a closed‑form bonus derived from elite action priors estimated from top‑quartile episodes. By adding a bounded transform of the log‑ratio of this prior to the policy’s action distribution, OPR requires no extra networks or off‑policy corrections and can be implemented as a one‑line change to PPO’s reward. Experiments on 49 Atari games show that OPR achieves a median human‑normalized score of 148.4 % with only 10 M steps, outperforming published PPO on 36 of 49 games and surpassing self‑imitation learning when matched on frames and backbone.
By Mai Pham, Vikrant Vaze, Peter Chin
SoundnessBench is a benchmark of 1,099 real research proposals designed to test whether large language models can judge the methodological viability of a research idea before resources are invested. The study evaluates 12 frontier LLMs and finds a pervasive optimism bias, with standard prompting yielding a 74.0% false‑positive rate and aggressive prompting increasing false rejections. Analyses show that this bias persists even when controlling for public corpus contamination, surface features, and literature grounding, indicating that current LLMs are not yet reliable gatekeepers for scientific rigor.
By Sy-Tuyen Ho, Minghui Liu, Huy Nghiem, Furong Huang
SIM-UT is a multi‑turn, user‑simulation evaluation framework designed to uncover realistic failure modes in large‑language‑model (LLM) medical triage chatbots. The study applied SIM‑UT to three WhatsApp‑based health chatbots in India, South Africa, and Kenya, comparing a medically fine‑tuned model (MedGemma) with general‑purpose models (ChatGPT and Claude) across more than 5,000 multilingual simulated conversations. A prominent failure identified was de‑escalation risk, where chatbots incorrectly down‑rated serious symptoms after users minimized them, potentially leading to dangerous clinical outcomes.
By Deepali Mittal, Dheeraj Mundhra, Yulia Sidorova, Tersia de Kock, Oli Munns, Sofie S Meyer
IFC-MCQ is a method that automatically creates multiple-choice benchmarks from Industry Foundation Class (IFC) files, enabling exact option-set grading for large language model (LLM) queries in architecture, engineering, and construction (AEC). The taxonomy spans 15 categories and 36 subtypes across four groups, with answer keys derived from the IFC file and distractors generated from file-derived values or predefined rules. Experiments on IFC models of varying sizes and characteristics revealed category-specific failure modes and highlighted how performance can differ based on the source IFC model and retrieval interface.
By Rabindra Lamsal, Sisi Zlatanova, Xuesong Shen
Cola2 is a hierarchical language modeling framework that jointly optimizes continuous semantic compression, reconstruction–generation balance, and continuous–discrete co‑adaptation. It dynamically segments text into semantic spans, controls compression via negative feedback, and trains a block‑causal masked diffusion decoder with clean, noisy, and dropped latent conditions. Experiments on five downstream tasks with 210 billion training tokens show competitive scaling, and an unconverged checkpoint with 346 billion source tokens outperforms Llama‑3‑8B on HellaSwag and StoryCloze.
By Hongcan Guo, Yian Zhao, Feng Wang, Shen Nie, Guoqiang Wei, Yan Zeng, Hengshuang Zhao
The paper introduces MemPABench, an interactive environment designed to evaluate how personal agents remember and act on user interaction preferences. The benchmark features long‑horizon tasks where agents must decide when to clarify, confirm, or act autonomously, using data from four simulated users who provided 99 sessions over four months and five preference changes each. Experiments across four memory systems and LLM backbones show that current memory approaches offer limited improvement in preference selection, with retrieval‑based methods performing no better than no memory and file‑based memory sometimes applying preferences in irrelevant contexts.
By Jialin Li, Zhenhao Chen, Hanjun Luo, Hanan Salam
The paper introduces Steward, a multi‑agent LLM system for autonomous investing that uses a type system to limit self‑evolution. It employs nine role‑specialised agents—four analysts, a Portfolio Manager, a Risk Manager, and a shared memory layer—to process news, SEC filings, macro data, and market feeds, while a reflector and deterministic gate enable reversible learning from trade outcomes. In a five‑month live deployment, Steward achieved a 0.3% loss after a 24.6% sector drawdown, and backtests show its invariant core consistently reduces maximum drawdown across random entry policies.
By Yebo Feng, Jingyang Dai, Liangrui Pan, Nelson Lau, Jiahua Xu, Yang Liu
ColaBench is a synthetic benchmark designed to evaluate how personal agents use long‑term interaction history to acquire and apply external knowledge. It contains 349 tasks in health and daily living, 10 user profiles, and a 13,033‑document knowledge base, with an Evidence Graph linking personal conditions, disclosed information, prior document exposure, and answer requirements. The paper also introduces MemKB, a system that uses state summaries and memory‑document links to retrieve relevant knowledge, achieving 86.25% accuracy with an average of 4.07 interaction rounds per task.
By Wenjie Zhang, Zhichao Shi, Xuhui Jiang, Xiaojun Wu, Cehao Yang, Chengjin Xu, Xiaoning Song
SCOPE compiles natural‑language policies into neurosymbolic predicate programs that combine small neural models, deterministic extractors, and logical composition. The system is jointly optimized with adversarially generated counterexamples, achieving higher accuracy than prompt‑optimization, task‑decomposition, and workflow‑generation baselines on real‑world policy benchmarks. Programs trained with the weakest executor transfer to other models, enabling deployment on inexpensive models with stable, interpretable predictions.
By Arnon Mazza, Elad Levi
The paper introduces Video‑Centered Deep Research Report Generation, a new benchmark that evaluates agents on their ability to generate long, interactive research reports from a source video and user request. It presents the OmniDR‑Bench dataset with 50 tasks in Chinese and English, 19 subcategories, and 6,414 criteria, and proposes Omni Atomic Rubric Trees (Omni‑ARTs) for evaluation across audiovisual grounding, multimodal evidence use, and interactive presentation. The authors also detail a skill‑based workflow, an automated rubric construction process, and report that current agents achieve a best score of 63.3/100, highlighting gaps in content coverage, causal analysis, and citation accuracy.
By Jiahao Meng, Qize Yang, Yuxuan Wang, Yunfei Chu, Wensi Dai, Jin Xu, Yunhai Tong
Foundation4D is a unified feed‑forward model that performs geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi‑camera driving video. It uses a shared backbone with factorized spatiotemporal attention to merge observations across cameras and time, and a video inference pipeline that refines camera trajectories via pose graph optimization and assembles object tracks. The authors also provide a data annotation pipeline and release the PAI‑Foundation4D benchmark, achieving state‑of‑the‑art results on depth, detection, and occupancy metrics and demonstrating that co‑training detection with geometry and occupancy improves performance.
By NVIDIA, :, S\'ergio Agostinho, Maximilian Apodaca, Emmanuel Attia, Guillem Braso, George Dong, Ilya Fedorov, Michele Fenzi, Gang Huang, Jiahui Huang, Shengyu Huang, Bingxin Ke, Laura Leal-Taix\'e, Kuan-Ting Lee, Mingxuan Liu, Tian Liu, Tim Meinhardt, Aljo\v{s}a O\v{s}ep, Naveen Kumar Rai, Kefei Ren, Cristiano Saltori, Keqiang Wu, Ning Xu, Fei Xue, Jin-Bey Yu, Lei Zhang, Qunjie Zhou