EvoEmo: Evolutionary Framework Optimizes Emotional Policies for LLM Negotiation Agents
EvoEmo has been developed by researchers as an evolutionary reinforcement learning framework aimed at enhancing dynamic emotional expressions during multi-turn price negotiations that involve Large Language Models (LLMs). This study, presented in arXiv preprint 2509.04310v4, tackles a significant issue where current LLM agents produce passive emotional reactions driven by preferences, leaving them susceptible to manipulation by adversarial agents. EvoEmo conceptualizes emotional state changes as a Markov Decision Process and utilizes population-based genetic optimization to cultivate high-reward emotional policies across various negotiation contexts. The framework also features a baseline for evaluating standard strategies against fixed-emotion approaches. This research paves the way for more resilient and strategically flexible emotional responses in complex negotiations within agentic AI.
Key facts
- EvoEmo is an evolutionary reinforcement learning framework for LLM agents.
- It optimizes dynamic emotional expression in multi-turn price negotiations.
- Existing LLM agents generate passive, preference-driven emotional responses.
- EvoEmo models emotional state transitions as a Markov Decision Process.
- It uses population-based genetic optimization to evolve emotion policies.
- Evaluation includes vanilla and fixed-emotion strategy baselines.
- The research is published on arXiv with ID 2509.04310v4.
- It addresses vulnerability to manipulation by adversarial counterparts.
Entities
Institutions
- arXiv