L-MAD: Multi-agent debate boosts legal AI by 8% but too many rounds backfires
Adding more agents reduces errors, but longer debates cause AI to reinforce mistakes.
The L-MAD framework systematically evaluates how different multi-agent debate structures affect legal textual entailment. By assigning distinct expert personas (e.g., judge, prosecutor, defense) to each agent, the system achieves up to an 8% improvement over strong single-agent baselines. This suggests that role specialization helps agents cover more nuanced legal reasoning and cross-check conclusions effectively.
However, scaling up debates reveals a critical trade-off. Increasing the number of agents reduces output inconsistency and boosts accuracy. But extending the number of discussion rounds triggers a detrimental 'over-deliberation drift' — agents start reinforcing each other's mistakes instead of correcting them. The paper, awarded Outstanding Paper at ICML 2026's AI4Law Workshop, provides practical boundaries for safely deploying collaborative multi-agent systems in high-stakes legal reasoning environments.
- L-MAD boosts legal reasoning accuracy by up to 8% over single-agent baselines by assigning expert personas to each agent.
- Increasing the number of agents reduces inconsistency and improves accuracy in legal text entailment.
- Extending discussion rounds causes over-deliberation drift, where agents reinforce each other's errors.
Why It Matters
Highlights critical design trade-offs for deploying collaborative multi-agent AI in high-stakes legal settings.