COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Shen, Wenkai, Zhou, Pengyang, Xu, Jiahe, Qian, Jiaming, He, Haozhe, Huang, Zhihao, Chen, Chaochao, Zheng, Xiaolin
Format:	Preprint
Published:	2026
Subjects:	Artificial Intelligence
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

_version_	1866910271649349632
author	Shen, Wenkai Zhou, Pengyang Xu, Jiahe Qian, Jiaming He, Haozhe Huang, Zhihao Chen, Chaochao Zheng, Xiaolin
author_facet	Shen, Wenkai Zhou, Pengyang Xu, Jiahe Qian, Jiaming He, Haozhe Huang, Zhihao Chen, Chaochao Zheng, Xiaolin
contents	LLM-powered search agents enable multi-step reasoning and tool use. However, these capabilities introduce retrieval-induced safety degradation, as harmful intents may decompose into seemingly innocuous sub-queries that lead to unsafe outcomes. Existing alignment methods struggle to capture sparse safety signals and fail to supervise diverse violations across multi-step interactions. We propose COMPASS, a Cognitive MCTS-Guided Process Alignment framework designed to achieve robust safety alignment throughout the agent workflow while preserving general utility. COMPASS integrates cognitive tree exploration (CTE) to efficiently synthesize stealthy attack trajectories, and introspective step-wise alignment (ISA) to isolate risky intermediate actions for fine-grained process supervision. Empirical results show that COMPASS achieves a favorable safety-utility trade-off while requiring substantially less training data.
format	Preprint
id	arxiv_https___arxiv_org_abs_2605_30838
institution	arXiv
publishDate	2026
record_format	arxiv
spellingShingle	COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents Shen, Wenkai Zhou, Pengyang Xu, Jiahe Qian, Jiaming He, Haozhe Huang, Zhihao Chen, Chaochao Zheng, Xiaolin Artificial Intelligence LLM-powered search agents enable multi-step reasoning and tool use. However, these capabilities introduce retrieval-induced safety degradation, as harmful intents may decompose into seemingly innocuous sub-queries that lead to unsafe outcomes. Existing alignment methods struggle to capture sparse safety signals and fail to supervise diverse violations across multi-step interactions. We propose COMPASS, a Cognitive MCTS-Guided Process Alignment framework designed to achieve robust safety alignment throughout the agent workflow while preserving general utility. COMPASS integrates cognitive tree exploration (CTE) to efficiently synthesize stealthy attack trajectories, and introspective step-wise alignment (ISA) to isolate risky intermediate actions for fine-grained process supervision. Empirical results show that COMPASS achieves a favorable safety-utility trade-off while requiring substantially less training data.
title	COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents
topic	Artificial Intelligence
url	https://arxiv.org/abs/2605.30838