SettleAgent: Planning-Enhanced Multi-Agent Legal Negotiation

ACL ARR 2026 January Submission508 Authors

Keywords: Multi-Agent, Legal Negotiation, Tree Search, RoT

Abstract: Current research on LLM-based legal agents has largely focused on verdict-oriented tasks, while lacking systematic modeling of pre-trial negotiation and settlement processes in civil disputes. We propose {SettleAgent}, a multi-agent framework for legal negotiation and settlement, and incorporate a red-team adversarial mechanism to systematically evaluate robustness under extreme bargaining tactics. We model bargaining as a sequential game: at each round, counsel agents first generate candidate proposals, then perform forward-looking strategic exploration via tree search, and finally apply RoT (Reflection on Search Trees) to distill success/failure patterns from branching outcomes to update their memory. We also introduce {SettleBench}, a benchmark for legal negotiation and settlement built from publicly available civil case documents. Experiments show that {SettleAgent} significantly outperforms a range of LLM and agent baselines in both settlement success and outcome quality, while remaining more stable under red-team stress testing.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: NLP Applications

Contribution Types: Model analysis & interpretability

Languages Studied: english chinese

Submission Number: 508