Multimodal agents increasingly interact with real-world-like environments and generate rich trajectories spanning perception, reasoning, and action. A key research focus is how to transform these raw multimodal trajectories into reusable experience that can improve future task performance. Existing experience-driven learning methods are mostly developed in textual settings and typically rely on manually designed memory schemas, making them less effective for multimodal experience that requires composing heterogeneous signals across modalities, temporal scales, and abstraction levels. We propose AutoMMemo, a general framework that enables agents to automatically design memory mechanisms for learning from multimodal experience. AutoMMemo represents each memory mechanism as an executable memo program, which defines how past episodes are stored, organized, and retrieved. It discovers effective memo programs through an iterative process of update-then-retrieve evaluation, reflection-guided mutation, and budget-aware tree search. Experiments across GUI/Web navigation and multimodal visual reasoning benchmarks show that AutoMMemo consistently improves agent performance over no-memory agents and fixed memory baselines, while maintaining reasonable token and interaction costs. Transfer and search-process analyses further demonstrate that the learned memory designs capture reusable experience structures and progressively refine retrieval relevance. These results highlight adaptive memory design as a key optimization target for multimodal experience-driven learning. Code is available at https://github.com/XingYuSSS/AutoMMemo.
Large language model agents increasingly take consequential actions in digital and physical environments, where safety risks may emerge only after multiple interaction steps. This has motivated world-model-driven predictive safety guardrails, which anticipate future consequences before execution. However, existing evaluations largely focus on end-to-end safety outcomes, leaving unclear whether safety gains arise from reliable prediction and faithful use of predicted risks. In this work, we systematically diagnose the prediction-to-action pipeline along three dimensions: how world-model predictions are integrated into decision making, whether the predictions themselves are reliable, and whether their content is faithfully used. Across seven policy models and two safety environments, we find that stronger prediction involvement does not yield monotonic safety gains and can reduce task completion while substantially increasing inference cost. We further find that world-model predictions remain insufficiently reliable, especially over longer horizons, where safety prediction degrades substantially relative to near-term forecasting. Finally, we find that safety gains can persist even under corrupted predictions, while locally verified predictions still leave subsequent safety violations, showing that predictive content is not faithfully incorporated into the decision process. Together, these findings expose a fundamental prediction-to-action gap in current predictive guardrails: producing more or seemingly accurate foresight is insufficient unless that foresight is both reliable and faithfully translated into safer decisions.
Large language models (LLMs) are increasingly explored for mental health support, yet existing benchmarks primarily focus on supportive capabilities rather than whether models consistently adhere to professional ethical boundaries. This issue is particularly important in multi-turn interactions, where users may introduce new information or exert emotional pressure over time. We introduce MindEth-Bench, a benchmark for evaluating ethical boundary adherence in multi-turn mental health interactions. Grounded in professional ethical standards for mental health services in China, MindEth-Bench covers multiple dimensions of ethical conduct and uses Monte Carlo Tree Search (MCTS) to systematically explore interaction trajectories that stress-test model boundaries. Experiments across nine LLMs show that models that initially respond appropriately can shift toward ethical violations over extended interactions, with substantial variation across models and ethical dimensions. These results show that single-turn evaluations may overestimate the ethical reliability of LLM-based mental health support systems and highlight the need to evaluate ethical behavior throughout multi-turn interaction.
Jiahe Guo, Xiangran Guo, Yulin Hu, Zimo Long, Xingyu Sui Harbin Institute of Technology, Xuda Zhi, Yongbo Huang, Hao He 等 11 位作者
Long-term memory enables large language model (LLM) agents to support personalized and sustained interactions.However, most work on personalized agents prioritizes utility and user experience, treating memory as a neutral component and largely overlooking its safety implications.In this paper, we reveal intent legitimation, a previously underexplored safety failure in personalized agents, where benign personal memories bias intent inference and cause models to legitimize inherently harmful queries.To study this phenomenon, we introduce PS-Bench, a benchmark designed to identify and quantify intent legitimation in personalized interactions.Across multiple memory-augmented agent frameworks and base LLMs, personalization increases attack success rates by 15.8%–243.7% relative to stateless baselines.We further provide mechanistic evidence for intent legitimation from internal representation space, and propose a lightweight detection–reflection method that effectively reduces safety degradation.Overall, our work provides the first systematic exploration and evaluation of intent legitimation as a safety failure mode that naturally arises from benign, real-world personalization, highlighting the importance of assessing safety under long-term personal context. WARNING: This paper may contain harmful content.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Xingyu Sui Harbin Institute of Technology, Yanyan Zhao Harbin Institute of Technology, Yulin Hu, Jiahe Guo, Weixiang Zhao Harbin Institute of Technology, Bing Qin Harbin Institute of Technology
Emotional Support Conversation requires not only affective expression but also grounded instrumental support to provide trustworthy guidance. However, existing ESC systems and benchmarks largely focus on affective support in text-only settings, overlooking how external tools can enable factual grounding and reduce hallucination in multi-turn emotional support. We introduce TEA-Bench, the first interactive benchmark for evaluating tool-augmented agents in ESC, featuring realistic emotional scenarios, an MCP-style tool environment, and process-level metrics that jointly assess the quality and factual grounding of emotional support. Experiments on nine LLMs show that tool augmentation generally improves emotional support quality and reduces hallucination, but the gains are strongly capacity-dependent: stronger models use tools more selectively and effectively, while weaker models benefit only marginally. We further release TEA-Dialog, a dataset of tool-enhanced ESC dialogues, and find that supervised fine-tuning improves in-distribution support but generalizes poorly. Our results underscore the importance of tool use in building reliable emotional support agents.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表