Research map / Multi-agent systems and games / LLM agents
Simulating Dispute Mediation with LLM-Based Agents for Legal Research
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces AgentMediation, an LLM-agent framework that simulates civil dispute mediation grounded in 330 real Chinese court disputes, and uses it to study how disputant TKI strategies, dispute causes, and mediator legal-knowledge access affect outcomes, reporting alignment with sociological theories.
Problem
Empirical study of legal dispute mediation is limited by confidential mediation data and difficulty isolating interacting variables such as disputant strategy, dispute cause, and mediator expertise.
Approach
A five-stage, HDR-inspired LLM multi-agent mediation simulator with TKI-based disputant personas, controllable mediator access to legal retrieval, a dispute-cause taxonomy, and a dual evaluator (outcome metrics via LLM-as-judge plus solution-quality metrics against expert annotations) that enables controlled experiments at scale.
Key results
- DeepSeek-V3-0324 and GPT-4o reach BERTScore above 0.71 and LLMScore about 0.79 against human expert utterances, close to the human-human upper bound of 0.69 / 0.76.
- Competing disputant strategy in NUM=ALL setting drops success rate from 82% to 11% and raises litigation risk from 24.00 to 64.00.
- Behavioral Misconduct dispute cause yields the lowest success rate (36%) and highest litigation risk (53.00) among causes tested.
Surprising findings
- Increasing bargaining turns yields only marginal gains in outcomes.
- Dynamic strategy adaptation raised Avoiding success rate from 36% to 85% in DeepSeek-V3 runs.
Limitations
- Satisfaction metric shows low agreement with human raters (Kappa about 0.36), indicating subjectivity.
- Data drawn from a single Chinese court database, limiting cross-jurisdiction generalization.
- Simulation relies on LLM agents rather than real mediation transcripts for validation beyond expert pairwise preferences.
Experimental setup
Datasets: Dispute Resolution Case Database (330 civil disputes) China Judgment Documents (1,080 cases for taxonomy)
Benchmarks: AgentMediation simulated mediation (330 disputes, 14,000+ processes)
Metrics: BERTScore (F1) LLMScore Success Rate Satisfaction Consensus Litigation Risk ROUGE-L (F1) Recall (legal bases) Cohen's Kappa chi-squared test
Baselines: GENTEEL NEGOTIATOR LLM-Stakeholders WarAgent AgentsCourt
Builds on: AgentsCourt Thomas-Kilmann Conflict Mode Instrument Harvard HDR framework chain-of-thought prompting
Closest papers by meaning
- CyberJurors: A Multi-Agent Simulation Task for E-Commerce Disputes Verdict (ICML 2026 · similarity 0.84)
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.82)
- PROBE: Learning to Audit Policy Compliance in Tool-Using LLM Agents (NeurIPS 2026 · similarity 0.81)
- Can Large Language Model Agents Simulate Human Trust Behavior? (NeurIPS 2024 · similarity 0.81)
- EAI: Emotional Decision-Making of LLMs in Strategic Games and Ethical Dilemmas (NeurIPS 2024 · similarity 0.81)
- The Value of Variance: Mitigating Debate Collapse in Multi-Agent Systems via Uncertainty-Driven Policy Optimization (ICML 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.