Semantic ScholarSafetyMethodImportance 8Annual Meeting of the Association for Computational Linguistics3mo ago

MHSafeEval: Role-Aware Interaction-Level Evaluation of Mental Health Safety in Large Language Models

Suhyun Lee, P. Achananuparp, Neemesh Yadav, Ee-Peng Lim, Yang Deng

2 CitationsPopularity 4.7

AI analysis

TL;DR

A new evaluation framework is proposed to diagnose the safety of LLMs in mental health counseling at the interaction level, incorporating the AI counselor's role.

Problem

Existing evaluation methods focus on single responses or static datasets, limiting their ability to diagnose the accumulation and context of harm that emerges gradually over multi-turn interactions typical of counseling.

Approach

R-MHSafe (Role-Aware Safety Taxonomy): A taxonomy that combines the roles an AI counselor adopts (e.g., perpetrator, instigator, facilitator, enabler) with clinically grounded harm categories.

MHSafeEval (Agent-Based Evaluation Framework): A closed-loop framework that formulates safety assessment as trajectory-level harm discovery through adversarial multi-turn interactions, guided by role-aware modeling.

Results & Contribution

Large-scale evaluation of state-of-the-art LLMs revealed substantial role-dependent and cumulative safety failures systematically missed by existing static benchmarks. The framework significantly improves failure-mode coverage and diagnostic granularity.