L-MAD: A Systematic Evaluation of Multi-Agent Debate Structures in Legal Reasoning

Short summary

L-MAD is a framework for evaluating multi-agent debate (MAD) structures in legal textual entailment tasks. Assigning expert personas to agents improves over single-agent baselines by up to 8%. Scaling agent population reduces inconsistency, but extending debate rounds causes 'over-deliberation drift' where agents reinforce each other's errors, outlining practical safety boundaries for deploying multi-agent systems in legal contexts.

• •L-MAD framework evaluates multi-agent debate in legal reasoning, improving over single-agent baselines by up to 8%
• •More agents reduce inconsistency; more debate rounds cause over-deliberation drift
• •Findings define safety margins for collaborative AI in high-stakes legal domains

Generated with AI, which can make mistakes.