Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EvoEmo: Towards Evolved Emotional Policies for Adversarial LLM Agents in Multi-Turn Price Negotiation

Yunbo Long, Liming Xu, Lukas Beckenbauer, Yuhan Liu, Alexandra Brintrup

TL;DR

EvoEmo addresses the gap that emotions are not just reflections of negotiation dynamics but actionable strategic levers in adversarial LLM negotiations. By formalizing emotion-driven negotiation as an emotion-aware MDP and evolving emotional transition policies via population-based reinforcement learning, EvoEmo discovers dynamic trajectories that improve success rate, efficiency, and buyer savings across diverse LLM opponents. The framework combines Bayesian updates of emotion transitions with evolutionary operators to continuously adapt emotion policies during execution. Empirical results on a CraigslistBargain-derived dataset and multi-model pairings demonstrate substantial gains over vanilla and fixed-emotion baselines, while also revealing emergent manipulative tactics, underscoring the need for safety-aware reward design and ethics-focused future work.

Abstract

Recent research on Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs) has demonstrated that agents can engage in \textit{complex}, \textit{multi-turn} negotiations, opening new avenues for agentic AI. However, existing LLM agents largely overlook the functional role of emotions in such negotiations, instead generating passive, preference-driven emotional responses that make them vulnerable to manipulation and strategic exploitation by adversarial counterparts. To address this gap, we present EvoEmo, an evolutionary reinforcement learning framework that optimizes dynamic emotional expression in negotiations. EvoEmo models emotional state transitions as a Markov Decision Process and employs population-based genetic optimization to evolve high-reward emotion policies across diverse negotiation scenarios. We further propose an evaluation framework with two baselines -- vanilla strategies and fixed-emotion strategies -- for benchmarking emotion-aware negotiation. Extensive experiments and ablation studies show that EvoEmo consistently outperforms both baselines, achieving higher success rates, higher efficiency, and increased buyer savings. This findings highlight the importance of adaptive emotional expression in enabling more effective LLM agents for multi-turn negotiation.

EvoEmo: Towards Evolved Emotional Policies for Adversarial LLM Agents in Multi-Turn Price Negotiation

TL;DR

EvoEmo addresses the gap that emotions are not just reflections of negotiation dynamics but actionable strategic levers in adversarial LLM negotiations. By formalizing emotion-driven negotiation as an emotion-aware MDP and evolving emotional transition policies via population-based reinforcement learning, EvoEmo discovers dynamic trajectories that improve success rate, efficiency, and buyer savings across diverse LLM opponents. The framework combines Bayesian updates of emotion transitions with evolutionary operators to continuously adapt emotion policies during execution. Empirical results on a CraigslistBargain-derived dataset and multi-model pairings demonstrate substantial gains over vanilla and fixed-emotion baselines, while also revealing emergent manipulative tactics, underscoring the need for safety-aware reward design and ethics-focused future work.

Abstract

Recent research on Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs) has demonstrated that agents can engage in \textit{complex}, \textit{multi-turn} negotiations, opening new avenues for agentic AI. However, existing LLM agents largely overlook the functional role of emotions in such negotiations, instead generating passive, preference-driven emotional responses that make them vulnerable to manipulation and strategic exploitation by adversarial counterparts. To address this gap, we present EvoEmo, an evolutionary reinforcement learning framework that optimizes dynamic emotional expression in negotiations. EvoEmo models emotional state transitions as a Markov Decision Process and employs population-based genetic optimization to evolve high-reward emotion policies across diverse negotiation scenarios. We further propose an evaluation framework with two baselines -- vanilla strategies and fixed-emotion strategies -- for benchmarking emotion-aware negotiation. Extensive experiments and ablation studies show that EvoEmo consistently outperforms both baselines, achieving higher success rates, higher efficiency, and increased buyer savings. This findings highlight the importance of adaptive emotional expression in enabling more effective LLM agents for multi-turn negotiation.

Paper Structure

This paper contains 46 sections, 9 equations, 11 figures, 5 tables, 5 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Emotions in Negotiation Dynamics
  4. LLM Agents in Negotiation Systems
  5. EvoEmo Framework
  6. Emotion-Aware MDPs
  7. Policy Representation and Evolution
  8. Evolved Reinforcement Learning
  9. Experimental Setting
  10. Datasets
  11. LLM Agents
  12. Baselines
  13. Experimental Setup
  14. Experimental Results
  15. Buyer Savings
  16. ...and 31 more sections

Figures (11)

  • Figure 1: Illustration of the workflow of the EvoEmo framework.
  • Figure 2: Negotiation results in terms of buyer savings (%, $\uparrow$) across the nine buyer-seller pairs. Black vertical lines on top of each bar indicate the 95% confidence interval (CI) for each setting.
  • Figure 3: Examples of manipulative and deceptive tactics
  • Figure 4: Mean negotiation success rate (%, $\uparrow$) and efficiency ($\downarrow$, in dialogue rounds) across all experimental pairings.
  • Figure 5: Multi-agent system of EvoEmo
  • ...and 6 more figures