Eirene: Addressing End-of-Life Counseling as Constrained Optimization via Multi-Agent Coopetition

ACL ARR 2026 January Submission5786 Authors

Keywords: Multi-Agent System, Safety-Critical Dialogue System, constraint-aware dialogue

Abstract: End-of-life counseling for terminal cancer patients constitutes a psychologically intensive clinical interaction that must simultaneously satisfy complex requirements, including the accuracy, safety, and empathetic delivery of medical and advance care planning information. While large language models (LLM) have recently been explored as counseling support tools, single LLM–based response generation struggles to coordinate these competing demands within a single response. To address this challenge, this study proposes Eirene, a multi-agent control architecture that separates goal-oriented response generation from constraint verification. Ablation study results show that Eirene achieves consistent performance improvements across all evaluation metrics. Moreover, analysis of the feedback loop reveals a trade-off between response quality and system cost, indicating that a limited number of iterations functions as an effective balance point. These findings suggest that a multi-agent control framework, which reframes safety and ethical constraints as post-generation judgment and intervention problems, can ensure both stability and conversational quality in high-risk clinical counseling settings.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: applications,dialogue state tracking

Contribution Types: NLP engineering experiment

Languages Studied: Korean

Submission Number: 5786