R²-MAD gives debating agents a memory of past debates to break shared misconceptions

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.03619 (2026-09-03) targets the failure where a majority of agents converge early on a wrong answer and debate amplifies rather than corrects it. Prior work addresses peer skew but leaves the agents' biased concept priors alone; R²-MAD adds an experience memory accumulated across debates, with a debate-state-aware retrieval policy that calibrates the prior based on the current consensus level, then uses those retrieved experiences to estimate per-agent reliability and reweight peer influence. The paper reports consistent gains over single-agent and multi-agent-debate baselines, though it does not name the benchmarks in the abstract, so verify before adopting.