Original title: Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation
Xuanfa Jin, Zhijian Ma, Yongcheng Zeng, Xinyu Cui +2 · September 2026
When multiple artificial intelligence agents debate to solve problems, they can make each other worse. If most agents start with the same wrong idea, called a shared misconceptionA failure mode where most agents initially converge on an incorrect answer and amplify the error through discussion., their discussion actually amplifies the mistake instead of fixing it. To stop this, researchers built R^2-MAD, or Remember and Reweight for Multi-Agent DebateA process where multiple AI models iteratively discuss a question to refine their responses.. This system gives agents a memory of past debates. It checks how much the agents currently agree and retrieves relevant past evidence to fix their biased initial concepts. It also uses those past experiences to measure how reliable each agent is, turning down the influence of less trustworthy peers. In tests across various benchmarks, R^2-MAD consistently beat standard single-agent systems and existing multi-agent debate methods.
Imagine a group of friends voting on where to have a picnic. If most initially suggest a spot that's actually prone to bad weather, they might talk each other into thinking it's fine. In this paper, AI agents face a similar groupthink problem where early wrong ideas can dominate. The solution is to remember past picnics to better judge who usually gives good advice and adjust their input accordingly.
This matters because AI systems used in decision-making could confidently give wrong answers if errors spread during discussion. By fixing this, AI outputs become more trustworthy for people relying on them in fields like healthcare or finance.
The abstract doesn't specify the exact benchmarks used or how generalizable the method is across different tasks. It focuses only on mitigating shared misconception, leaving other potential vulnerabilities in MAD unaddressed. Computational costs or real-world deployment challenges are not discussed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation — Xuanfa Jin, Zhijian Ma, Yongcheng Zeng et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.