Letting Disagreement Surface Beats Resolving It: A Gated Ablation Study of Provenance-Weighted Memory in Multi-Agent Debate
Surfacing Disagreement Beats Resolving It: A Gated Ablation Study of Provenance-Weighted Memory for Multi-Agent Debate
Researchers conducted a gated ablation of five agent memory mechanisms in the DebateGraph multi-agent debate framework; across 39 runs, only the mechanism that preserves contradictions and hands a ranked conflict set to the debate rather than merging it into a single winner outperformed. This mechanism raised F1 by 0.340 (95% CI [+0.267, +0.414]) at unchanged cost, reduced attack successes on 122 poisoned entries from 22 to 7 (McNemar exact p=0.0003), and the effect replicated on two retrieval backbones and directionally replicated on HotpotQA documents. The other mechanisms did not meet the bar, with two performing worse than not enabling them; once contradictions entered the prompt, the write trust gate and the router provenance feature reduced F1 by 0.075 and 0.096 respectively, and the structural reason the trust gate failed is that poisoned objects were always entities already present in the graph (53% of gold labels), causing groundedness to score lies higher than genuine first mentions.