VeriHarm-Bench is a new benchmark that measures harmful success in large language model coding agents by automatically generating verifiable tests that detect unintended harm while preserving the original task objective. The framework minimally perturbs coding tasks to elicit harmful behavior and synthesizes executable harm tests alongside task‑correctness tests. In evaluations, frontier agents such as GPT‑5.6‑Sol and Claude‑Opus‑5 achieve high task accuracy (≈95%) but only a 37% safe success rate, showing that task success can coexist with significant unintended harm.
By Wei-Lin Chen, Zhepei Wei, Xinyu Zhu, Yu-Min Tseng, Tzu-Han Lin, Yu Meng
The paper introduces a new model that trains large populations of bi-directional neural agents in a fully distributed setting to play a referential game, addressing limitations of prior work that used only two agents or allowed shared internal representations. Experiments on multiple image datasets demonstrate the model’s effectiveness and scalability to large populations. The study also finds that more conventional languages are easier to learn, offering a direct advantage for cultural transmission.
By Maxime Toquebiau, J\'er\^ome Botoko Ekila, Paul Van Eecke, Katrien Beuls
The paper introduces Agentic Conversation Context Curation (AC3), a training‑free method that cleans multi‑turn human‑AI dialogue by consolidating the user’s task specification, auditing assistant outputs, and editing the context to remove or rebuild assistant turns. AC3 outperforms both full context and assistant‑omission (AO) baselines on several benchmark tasks, closing a significant portion of the performance gap and receiving preference from a GPT‑5‑mini judge on real WildChat conversations. The study also shows that context pollution is contagious and that simply ignoring assistant turns can worsen performance, underscoring the need for structured removal of assistant content.
By Matthew Ho, Chenglong Wang, Baolin Peng, Nebojsa Jojic, Philippe Laban, Michel Galley, Jianfeng Gao, Lianhui Qin
The paper introduces Zep Agent Skills, a production service that autonomously compiles and curates procedural memory for large language model agents by converting verified execution traces into versioned skills. It records each task attempt as a trace, uses a compiler with Reflector and Curator roles to transform verified traces into skills, and applies admission gates and search ranking based on verified outcomes. The open Agent Skills format, originally from Anthropic, allows any host to load compiled skills, and evaluation on several benchmarks shows significant performance gains for weaker executor models without additional training.
By Daniel Chalef, Peter Evans, Pavlo Paliychuk, Mehul Patel, Preston Rasmussen, Jack Ryan
The paper introduces Online Meta-Control of Experience (OMCE), a framework that updates an agent’s experience memory and learns a meta-control policy during test-time adaptation. OMCE uses a lightweight contextual-bandit meta-controller to choose among four cognitive actions—DIRECT, EPISODE, GUIDE, and DISTILL—based on compact retrieval signals, thereby deciding when to invoke experience, which memory format to use, and when to consolidate new guidance. Experiments on four online adaptation benchmarks across four backbones show that OMCE can improve first-attempt success by up to 20.5 percentage points over the strongest baseline while reducing per-task token use by a factor of 5.7 to 62.5 relative to ACE with GPT‑5.2, and controlled ablations confirm the benefits of learned routing and ongoing policy adaptation.
By Zeyuan Zhao, Yuanyuan Shi, Peng Yu, Lingyue Fu, Shao Zhang, Ying Wen
The paper presents a new LLM‑based narrative platform that allows interchangeable Experience Managers (EMs) to operate over a shared symbolic world model, enabling explicit preconditions and effects for every operation. By representing each session as a plan in a common state‑transition system, the authors evaluate three different managers—open generation, Monte Carlo repair, and autonomous characters—on four properties: player agency, EM controllability, author intent, and robustness, using an interactive Little Red Riding Hood scenario. The results demonstrate that no single manager excels in all properties, providing authors and designers with concrete criteria to choose a manager that best fits their priorities.
By Jules Clerc (Heudiasyc), Domitile Lourdeaux (Heudiasyc), Mohamed Sallak (Heudiasyc), Marc Ravaine
The paper introduces BlackSwan, a refutation-oriented verification method for language model agents that predicts the intended effect of each action, formulates testable hypotheses about how the action could be wrong, and only executes the action if no hypothesis is supported by observations. On AppWorld tasks, BlackSwan boosts success rates by 14.9–25.2 percentage points over the strongest baseline, and it also improves WebShop performance by 12.7 percentage points. The approach demonstrates that agents can actively self‑correct misaligned actions through conjecture and refutation before execution.
By Minhyeon Oh, Chiyeong Heo, Geonwoo Noh, Hoyoung Kim, Jangwon Son, Jungseul Ok
DR Pixel introduces a page‑native deep research agent that uses rendered document page images as the sole unit for retrieval and reasoning, bypassing OCR. The open‑source system operates on 1.06 M pages and, without training, outperforms OCR‑based baselines on GPT‑5.4, Gemini‑3.5‑Flash, and Qwen3.6‑27B while requiring fewer tool calls. Training with counterfactual credit reweighting further boosts retrieval recall by ~15 points and accuracy by >10 points on the DRPixelBench benchmark, and brings the model within 1.8 points of an oracle on the ViDoRe‑V3 out‑of‑distribution test.
By Tiansheng Hu, Yichuan Wang, James V. Talwar, Nicholas Tomlin, Chen Zhao, Sewon Min
The paper introduces Mr Dre, an evaluation suite that tests Deep Research Agents (DRAs) on multi-turn report revision, a process more reflective of human research practices. It defines a unified long‑form report protocol covering comprehensiveness, factuality, and presentation, and includes a human‑verified feedback simulation pipeline. Analysis of five DRAs shows that while agents can address most user feedback, they regress on 16‑28% of previously covered content and citation quality, and over successive turns they disrupt unrelated content and fail to preserve earlier edits, with prompt engineering and dedicated subagents offering limited improvement.
By Bingsen Chen, Boyan Li, Ping Nie, Yuyu Zhang, Xi Ye, Chen Zhao
AgentHop is a diagnostic benchmark comprising 1,011 multiple‑choice questions that test agentic multi‑hop question answering over scientific citation graphs. It provides a controlled seven‑tool sandbox with fixed token, turn, and tool‑call limits, allowing researchers to dissect model performance across four operational axes: retrieval, synthesis, tool‑call, and resource management. Analysis of 19 models shows distinct family‑specific behaviors, such as GPT models committing early, Anthropic and GLM checkpoints verifying before committing, DeepSeek and Kimi over‑searching, and Gemini‑3 Pro maintaining balance, while within‑family differences are revealed in retrieval versus synthesis strategies.
By Chanhee Park, Jeongho Yoon, Sungbin Han, Hyeonseok Moon, Heuiseok Lim
Li$^{2}$SAO is a lightweight variant of Single-rollout Asynchronous Optimization (SAO) that removes the need for a pretrained value model. It introduces group-relative actor‑critic bootstrapping, where a group-relative advantage trains the actor and a group‑centered return trains the critic before switching to single‑rollout GAE. Experiments show that Li$^{2}$SAO matches or surpasses GRPO early in training and raises the later performance ceiling on both single‑turn reasoning and multi‑turn agentic tasks.
By Zichen Bai, Yanli Wang, Yang Liu, Junhao Shi, Xueqin Liu, Jinguo Sun, Muye Li, Shaoqiang Zheng, Xuefei Diao, Zhuwei Zeng, Nan Duan, Peng Cao
DARE introduces a distribution‑aware retrieval framework that enhances Large Language Model agents’ ability to select appropriate R functions for data science tasks. The system builds on a curated R function knowledge base (RPKB), a new retrieval method (DARE), an R‑oriented agent (RCodingAgent), and a benchmark (RStatBench). Experiments show significant gains in retrieval quality (up to 32.65 pp improvement in NDCG@10) and a near‑doubling of downstream analysis accuracy on RStatBench.
By Maojun Sun, Yue Wu, Yifei Xie, Ruijian Han, Binyan Jiang, Defeng Sun, Yancheng Yuan, Jian Huang
The paper presents a multimodal house‑price prediction framework that combines structured attributes, textual descriptions, and property images. It uses a large language model to extract spatio‑temporal and contextual features from text and an unsupervised vision‑language method to cluster images into categories such as indoor, outdoor, mixed, or floorplan without manual labeling. The fused representation is fed to an optimized regressor, and experiments on a large real‑world dataset show improved accuracy over traditional and existing multimodal baselines, with ablation studies highlighting the contribution of the new textual and visual features.
By Khandokar Md. Rahat Hossain, Souvik Das, Tahsina Hashem, Mohammed Eunus Ali, Tanzima Hashem
BiCAA introduces bidirectional credit assignment for search-augmented agents by combining Forward Solvability Gain (FSG) and Subsequent Reasoning Gain (SRG). FSG measures how each search step increases the likelihood of the correct answer, while SRG evaluates how a step supports future reasoning, discouraging redundant retrieval. Experiments on seven open‑domain QA benchmarks and two Qwen backbones show that BiCAA improves average F1, especially on multi‑hop tasks.
By Yibin Huang, Bin Xu, Hailong Cao, Conghui Zhu
The paper introduces FAID, a fine‑grained adaptive framework for detecting implicit hate speech. It first classifies samples into Shallow, Targeted, or Context‑Dependent categories and then applies tailored strategies—Prompt‑Tuning for Shallow, knowledge augmentation for Targeted, and an agentic prompt‑generation system for Context‑Dependent cases. Experiments on four benchmark datasets show that FAID outperforms state‑of‑the‑art baselines.
By Han Wang, Hanqing Tao, Xuesong Wang, Yi Zhu
The paper introduces Latent Harness, a method that compiles external agent behaviors into a small rank‑8 activation shift applied to every decoder block of a frozen language‑model backbone, affecting only 0.03‑0.05% of parameters. This compiled shift eliminates the need for on‑call harness text, offers runtime controls for injection strength, phase, and mixing, and recovers most of the performance gains of the original harness across multiple settings. The approach also enables zero‑shot transfer across tasks, composability in shift space, and improves budget‑matched harness search, outperforming LoRA with far fewer parameters.
By Ran Li, Shimin Di, Zeyu Ren, Ling Yue, Jiachuan Wang, Lei Chen
The paper introduces Research Attention Prediction (RAP), a benchmark designed to evaluate large language model (LLM) agents’ ability to infer changes across entire research fields from limited evidence gathered during online scholarly searches. RAP is built on a local search environment comprising 356,357 dated arXiv papers and includes two evaluation sets: a Core set of 2,160 questions covering 54 AI/ML fields and an Extended set spanning 269 fields. Experiments show that while search improves agents’ recognition of recent changes, their forecasting of future shifts remains limited or even negative, indicating that LLM agents can detect some recent developments but struggle to reliably anticipate future research activity.
By Yingqian Wu, Jingcong Liang, Shengbin Yue, Siyuan Wang, Zhenfei Yin, Junchi Yu, Zhongyu Wei
Strategist is a meta‑agent that automatically evolves inference strategies for AI agents, turning model calls into responses. It stores reusable strategy blueprints in a library, retrieves relevant ones for new tasks, generates and tests candidate strategies, and refines them based on results. Across 30 benchmarks, Strategist outperforms the strongest baseline on 24 tasks, matches its average accuracy while cutting inference cost by 39%, and its strategies transfer to unseen tasks and models without additional search.
By Taiming Lu, Qi Xu, Zhigang Hua, Zhuang Liu
DiabAgent introduces a research suite for agentic physiological control in type 1 diabetes, comprising the APC framework, DiabetSim simulator, and SweetBench benchmark. The APC framework uses medical-device tools and ongoing observations to infer patient states, accumulate experience, and select treatment actions. In virtual testing with DiabetSim, APC outperforms the open-source baseline OpenAPS, achieving a COGI of 73.5 versus 71.2, and ablation studies show that individual adaptation and contextual information improve glucose control.
By Chengxi Li, Shaoxuan Li, Deyao Hong, Qingle Liu, Yi Huang, Puheng Bian, Minzhi Xu, Lanyue Zhang, Mingze Sun, Xin Li, Calvin Xiao, Shengyuan Wang, Huan-ang Gao, Jiazheng Xu, Kaisen Yang, Hongning Wang
AStockArena is a real‑time decision‑auditing benchmark for continuous trading in China’s A‑share market, evaluating 13 open‑ and closed‑source models over 3,198 sequential decisions. The benchmark reconstructs actual account actions by aligning reports, trade requests, execution responses, and account states, then audits them with eight checks covering execution traceability, evidence sufficiency, and financial meta‑awareness. Across 1,070 trading points, 72.99% show at least one audit gap, with evidence gaps most common, and historical replay demonstrates that flagged trades significantly affect portfolio outcomes.
By Zeyu Liu, Sing Kwong Lai, Xuanzhi Feng, HaoYang Shang, Yang Li, Zihao Fang, Xuan Liu