MADS Pipeline for Simulating Adversarial Multi-Agent Dialogues

MADS is a synthetic data pipeline that generates diverse, non-cooperative dialogue trajectories to improve agent robustness. It includes MADS-Bench, a benchmark with 1,095 conversations where 66% feature failure-prone scenarios to train models via SFT and RLVR.