Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation
Xuanfa Jin, Zhijian Ma, Yongcheng Zeng, Xinyu Cui, Haifeng Zhang, Jun Wang
cs.CL · cs.AI
Quick verdict
R2-MAD adds experience memory and confidence weighting to multi-agent debate; consistent small gains across 4 benchmarks, but no significance testing.
Summary
The paper targets the shared-misconception failure of multi-agent debate (MAD): when a majority of LLM agents initially converge on a wrong answer, discussion amplifies the error. Building on the latent-concept decomposition of debate dynamics into a biased concept prior and an amplifying peer skew, the authors propose R2-MAD, which gives each agent an experience memory accumulated from past debates. A debate-state-aware retrieval policy (BGE-M3 cosine pre-filtering, then MMR whose relevance/diversity trade-off scales with the current consensus ratio) selects K past cases to correct the prior; the same cases yield per-peer confidence scores (similarity-weighted historical correctness) thresholded into prompt-level confidence annotations that modulate peer influence. On MATH500 (Level-5), MMLU-Pro Economics/Engineering, and TruthfulQA with Qwen2.5-7B, Qwen3-8B, Gemma-3-4B plus Llama-3.3-70B and GPT-4o-mini, R2-MAD attains the best average accuracy (e.g., 0.607 vs 0.593 compute-matched Self-Consistency on Qwen2.5-7B), with larger gains on the shared-misconception subset and stance-transition rates moving as the theory predicts. Main limitations: no significance testing despite small test sets (134-242 items) and temperature-1.0 decoding; one ablation cell where removing memory beats the full model, contradicting the text; the retrieval budget K is never given numerically; and memory construction depends on gold outcome labels.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Shared-misconception failures in MAD stem from two compounding factors — a biased concept prior and an amplifying peer skew — and effective mitigation must intervene on both.
- Retrieved experiences act as a likelihood-ratio correction to the agent's prior (Prop. 4.1), and memory-derived confidence weights slow majority dominance from O(m) to O(αm) (Prop. 4.2), with additive log-odds lifts when combined (Theorem C.3).
- R2-MAD achieves the highest average accuracy across MATH500, MMLU-Pro Economics/Engineering, and TruthfulQA on all five evaluated models, e.g., 0.607 vs 0.593 for compute-matched Self-Consistency on Qwen2.5-7B (Table 1).
- Gains concentrate in the shared-misconception regime (relative improvements on that subset exceed overall gains, Figure 3), and confidence weighting shifts stance-transition rates in the direction Prop. 4.2 predicts (Table 4).
- Both components contribute, and the debate-state-aware retrieval policy beats five alternatives including random and similarity-based retrieval (Table 3: 0.663 vs 0.635/0.647 avg).
Strengths
Dual-factor design is principled, not ad hoc
The two mechanisms map one-to-one onto the factors of the latent-concept decomposition in Eq. (2): memory-corrected prior P(θ|E,ϕ) targets the biased prior, confidence weights w(t)_{j,i} in Eq. (3) target the peer skew — a clean architectural story inherited from Estornell & Liu (2024).
Controls against the 'it's just RAG' confound
ICL-CoT retrieves from the identical memory bank as single-agent in-context examples (Appendix B.5: avg 0.620 vs 0.663 for R2-MAD), and RQ3 shows random retrieval from the same bank scores 0.635 vs 0.663 for the state-aware policy (Table 3), isolating retrieval design rather than mere memory availability.
Compute-fair baselines and cost transparency
Mechanism-level evidence, not just end accuracy
Weaknesses
No statistical inference anywhere
Text–table contradiction in the ablation
Retrieval budget K never specified
Theory is thinner than it appears
Memory construction cost and label dependence untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
R2-MAD (arXiv 2609.03619) targets the 'shared misconception' failure of multi-agent debate (MAD): when a majority of the 3 agents answer wrongly at round 0, standard debate amplifies the error (Figure 1). The paper's claims: (1) following Estornell & Liu (2024)'s latent-concept framework, debate failures decompose into a biased concept prior and an amplifying peer skew, and both must be intervened on; (2) an experience memory of past debates does both — retrieved cases correct the prior (Prop. 4.1, a Bayes-rule identity) and support per-agent reliability estimates whose weights slow majority dominance from O(m) to O(αm) (Prop. 4.2); (3) empirically, R2-MAD attains the best average accuracy on all five models tested, with the largest gains on the shared-misconception subset (RQ4).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the shared misconception problem in multi-agent debate?
In multi-agent debate, several LLM agents discuss a question and iteratively refine their answers. If a majority of them happen to produce the same wrong answer at the start, the discussion tends to amplify that error instead of fixing it — the agents holding the correct answer get persuaded to abandon it. R2-MAD is designed specifically to resist this majority-pressure failure mode.
How does R2-MAD work?
Each agent keeps a memory of past debate episodes: the debate state, everyone's responses, the correct answer, and who was right. During a new debate the agent retrieves similar past cases — favoring diverse, contrastive cases when current consensus is high and successful cases when consensus is low — to adjust its prior beliefs, and uses the same cases to estimate how historically reliable each peer's stance is, annotating peers' responses with high/low confidence to reweight their influence.
Does R2-MAD beat self-consistency and standard debate?
In the paper's experiments on MATH500, MMLU-Pro (Economics and Engineering subsets) and TruthfulQA across five LLMs (Qwen2.5-7B, Qwen3-8B, Gemma-3-4B, Llama-3.3-70B, GPT-4o-mini), R2-MAD achieves the highest average accuracy, e.g., 0.607 vs 0.593 for compute-matched 9-path self-consistency on Qwen2.5-7B. Gains are largest when a majority of agents started wrong, but no statistical significance tests are reported and test sets are small (134-242 items).
What are the main limitations of R2-MAD?
It needs an outcome signal (e.g., gold answers) to build its memory; adds roughly 1.5-2× the wall-clock cost of standard debate; uses a memory that is frozen after offline construction, so it may degrade on out-of-distribution tasks; and could reinforce errors if the memory bank itself is systematically wrong (the authors recommend auditing it). A retrieval parameter K is also not stated numerically in the paper.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 3/4 bar values arrived as flat numeric sequences; per-method mapping was reconstructed only where the prose anchors it (e.g., Economics 29.41% vs 17.65% for MAD, Engineering 27.10% vs 17.42%), so unanchored figure values are not cited in this review. All other content (abstract through Appendix D prompts) appears completely extracted; no truncation detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.