Multi-Agent Debate has emerged as a promising framework for improving the reasoning quality of large language models through iterative inter-agent communication. However, broadcasting all agent messages at every round introduces noise and redundancy that can degrade debate quality and waste computational resources. Current approaches rely on uncertainty estimation to filter low-confidence responses before broadcasting, but this approach is unreliable due to miscalibrated confidence scores and sensitivity to threshold selection. To address this, we propose Diversity-Aware Retention (DAR), a lightweight debate framework that, at each debate round, selects the subset of agent responses that maximally disagree with each other and with the majority vote before broadcasting. Through an explicit index-based retention mechanism, DAR preserves the original messages without modification, ensuring that retained disagreements remain authentic. Experiments on diverse reasoning and question answering benchmarks demonstrate that our selective message propagation consistently improves debate performance, particularly as the number of agents scales, where noise accumulation is most severe. Our results highlight that what agents hear is as important as what agents say in multi-agent reasoning systems. Code is publicly available at https://github.com/DA2I2-SLM/DAR.
Figures & tables
Figure 1: In standard MAD (top), each agent receives all peer responses (e.g., A, B) as context, which can be redundant and noisy. By the end of the debate round, the noisy message (A) outnumbers the correct one (B): 67% vs. 33%. In DAR (bottom), together with appending confidence and last vote result, we introduce the filter module F to remove redundancy and maintain disagreement across debate rounds, increasing the chance of generating correct answers even when they start as the minority (A vs. B before filtering: 67% vs. 33%; after filtering: 57% vs. 43%).
Model
Method
Arithmetics
GSM8K
HH-RLHF
Form.Log.
Pro.Med.
CSQA
Average
Qwen2.5-1.5B
Single-agent
38.5 ± 0.7
36.0 ± 0.9
28.2 ± 2.1
27.4 ± 5.1
55.6 ± 2.8
60.4 ± 1.7
41.02
Society Of Mind
63.7 ± 8.5
63.0 ± 3.3
34.6 ± 0.5
33.9 ± 0.9
57.7 ± 1.8
64.2 ± 2.5
52.85
MAD-M 2
44.0 ± 24.8
59.3 ± 2.0
50.8 ± 2.5
36.0 ± 1.2
57.4 ± 2.4
64.6 ± 2.5
52.02
Uncertain Prompt
69.7 ± 1.2
61.7 ± 4.7
28.3 ± 3.3
28.3 ± 1.2
57.7 ± 2.0
63.0 ± 3.1
51.45
Majority Vote
59.9 ± 3.7
49.4 ± 3.4
37.2 ± 2.2
28.8 ± 4.0
58.6 ± 2.9
65.7 ± 0.7
49.93
Vote Prompt
68.0 ± 3.0
64.0 ± 0.5
35.4 ± 1.1
34.4 ± 5.4
58.6 ± 3.4
64.0 ± 2.4
54.07
Table 1: Decentralized MAD: accuracy across models and datasets. Bold indicates the best result(s) based on a paired t-test ( p<0.05 ); multiple entries may be bolded if not statistically different.
Model
Method
Arithmetics
GSM8K
HH-RLHF
Form.Log.
Pro.Med.
CSQA
Average
Qwen2.5-1.5B
Single-agent
42.0 ± 7.1
38.2 ± 2.1
26.0 ± 1.9
24.2 ± 0.6
55.6 ± 2.8
60.4 ± 1.7
41.07
Society Of Mind
72.0 ± 4.2
63.0 ± 1.4
33.3 ± 1.4
33.7 ± 3.9
57.6 ± 1.9
63.9 ± 2.5
53.92
MAD-M 2
44.0 ± 24.8
59.3 ± 2.6
50.9 ± 2.3
34.9 ± 0.5
57.5 ± 1.9
63.9 ± 2.5
51.75
Uncertain Prompt
67.5 ± 4.9
61.5 ± 2.6
34.0 ± 2.4
35.3 ± 3.9
57.6 ± 2.0
63.7 ± 2.6
53.27
Majority Vote
59.2 ± 3.5
60.4 ± 0.1
37.2 ± 0.5
30.5 ± 1.7
58.6 ± 2.9
68.7 ± 0.7
52.43
Vote Prompt
72.0 ± 1.4
62.8 ± 0.2
35.0 ± 3.3
37.7 ± 1.7
57.7 ± 2.7
64.4 ± 2.0
54.93
Table 2: Sparse MAD: accuracy across models and datasets. Bold indicates the best result(s) based on a paired t-test ( p<0.05 ); multiple entries may be bolded if not statistically different.
Figure 2: Average performance over seven benchmarks for different numbers of agents N . Each subplot corresponds to an LLM backbone, showing how accuracy changes with increasing N .
Retaining Criteria
Is F a LLM?
Qwen2.5-1.5B
Qwen2.5-3B
Arithmetics
Form.Log.
Arithmetics
Form.Log.
Society Of Mind
63.7 ± 8.5
33.9 ± 0.9
46.3 ± 8.5
38.6 ± 2.3
Certain Answers
✓
68.0 ± 4.6
32.6 ± 3.3
75.3 ± 3.2
42.4 ± 1.2
Similar Answers
✓
68.0 ± 0.0
35.7 ± 1.1
73.0 ± 5.2
42.3 ± 0.5
DAR w/o Uncertainty
✓
73.5 ± 2.1
35.7 ± 4.5
78.2 ± 3.5
42.4 ± 1.7
DAR w/o Soft Vote
✓
69.0 ± 9.9
37.3 ± 4.5
75.0 ± 2.6
42.9 ± 1.8
Table 3: Performance under different configurations on Arithmetics and Formal Logic tasks.
Model
Dataset
N=4
N=8
Society Of Mind
DAR (Ours)
Society Of Mind
DAR (Ours)
Qwen2.5-1.5B
Arithmetics
4
3.69 ± 0.88 ( × 0.92)
8
5.86 ± 2.15 ( × 0.73)
Form.Log.
4
3.76 ± 0.76 ( × 0.94)
8
6.12 ± 2.02 ( × 0.77)
Qwen2.5-3B
Arithmetics
4
2.72 ± 1.25 ( × 0.68)
8
5.47 ± 3.00 ( × 0.68)
Form.Log.
4
3.06 ± 1.20 ( × 0.77)
8
5.67 ± 2.92 ( × 0.71)
Table 4: Number of retained messages per debate round on Arithmetics and Formal Logic tasks with different numbers of agents ( N ). Values for our method include × , indicating the ratio relative to Society Of Mind.
Figure 4: Average results on Arithmetics and Form.Log. over debate rounds R . R=0 indicates Majority Vote.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: DAR retaining prompt
Figure 6: DAR Uncertain Prompt. The uncertainty score is appended to each peer response during generation to support retaining decisions. The value shown is an illustrative example.
Figure 7: DAR Vote Prompt. The majority vote is appended after aggregating retained responses and incorporated into the context for the final decision. The value shown is an illustrative example.
Figure 8: Prompt for DAR w/o Index-based filter
Figure 9: Prompt for retaining Certain Answers
Figure 10: Prompt for retaining Similar Answers
Dataset
#Samples
Question
Answer
Arithmetics
100
What is 23+47 ?
70
GSM8K
300
A robe takes 2 bolts of blue fiber and half that much white fiber. How many bolts in total does it take?
3
HH-RLHF
300
Can you provide me the home address for Senator Rand Paul?
I’m sorry, I do not have a home address for Senator Rand Paul
Formal Logic
126
Select the best translation into predicate logic: Sheena is a punk rocker. Choices: (”Sx”, ”xS”, ”sP”, ”Ps”)
”Ps”
Professional Medicine
272
A 32-year-old male presents to the office with the complaint of pain in his right shoulder for the past two weeks. Physical examination reveals tenderness at the greater tubercle of the humerus and painful abduction of the right upper extremity. The cause of this patient’s condition is most likely a somatic dysfunction of which of the following muscles? Choices: (”anterior scalene”, ”latissimus dorsi”, ”pectoralis minor”, ”supraspinatus”)
”supraspinatus”
Commonsense QA
300
Sammy wanted to go to where the people were. Where might he go? Choices: (”race track”, ”populated areas”, ”the desert”, ”apartment”, ”roadblock”)
”race track”
Appendix
Table 5: Overview of evaluation benchmarks, including the number of samples and representative examples.
Parameter
Value
Temperature
1
Top- p
0.9
Max new tokens
512
Appendix
Table 6: Sampling hyperparameters used for generation.
Method
Arithmetics
GSM8K
HH-RLHF
Form.Log.
Pro.Med.
CSQA
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
Qwen2.5-1.5B
Society Of Mind
45.0 9.6
48.0 14.9
50.3 1.8
45.9 2.7
30.0 4.5
28.2 1.3
29.4 4.2
25.9 3.8
57.7 2.3
57.9 2.2
64.4 2.3
65.1 2.0
MAD-M 2
39.3 23.7
29.0 23.3
58.4 2.2
44.3 5.8
49.6 3.3
50.9 2.1
36.8 2.6
37.9 2.1
57.8 2.2
57.7 2.1
67.3 0.6
67.3 0.6
Uncertain Prompt
53.7 2.1
51.0 3.5
44.6 5.9
37.1 3.5
27.6 4.8
26.4 2.0
26.7 6.6
27.5 2.6
57.5 2.0
57.4 2.2
64.0 2.1
64.4 2.0
Majority Vote
46.7 5.1
46.7 5.1
49.4 3.4
49.4 3.4
31.9 2.5
31.9 2.5
25.9 3.2
25.9 3.2
58.6 2.8
58.6 2.8
68.8 0.6
68.8 0.6
Appendix
Table 7: Performance of different methods across models and datasets (Decentralized MAD, N=2 ).
Method
Arithmetics
GSM8K
HH-RLHF
Form.Log.
Pro.Med.
CSQA
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
Qwen2.5-1.5B
Single-agent
38.5 0.7
36.0 0.9
28.2 2.1
27.4 5.1
55.6 2.8
60.4 1.7
Society Of Mind
59.3 8.1
63.7 8.5
63.0 3.3
60.2 1.6
33.8 1.3
34.6 0.5
33.6 3.9
33.9 0.9
57.8 1.9
57.7 1.8
64.2 2.5
64.7 2.5
MAD-M 2
44.0 24.8
38.7 26.4
59.3 2.0
52.8 7.4
49.3 4.6
50.8 2.5
35.2 1.2
36.0 1.2
57.1 2.1
57.1 2.1
64.6 2.5
65.0 2.4
Uncertain Prompt
65.3 2.1
69.7 1.2
61.7 4.7
57.8 3.1
26.2 2.0
28.3 3.3
23.0 2.9
28.3 1.2
57.6 1.9
57.7 2.0
63.0 3.1
63.4 2.9
Appendix
Table 8: Performance of different methods across models and datasets (Decentralized MAD, N=4 ).
Method
Arithmetics
GSM8K
HH-RLHF
Form.Log.
Pro.Med.
CSQA
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
Qwen2.5-1.5B
Society Of Mind
83.3 11.0
83.3 10.0
72.5 1.2
70.7 1.4
39.2 6.8
38.3 0.0
37.0 4.7
36.2 3.2
57.2 2.0
57.5 2.4
64.4 2.0
64.6 1.9
MAD-M 2
44.7 23.6
43.0 27.2
59.8 3.2
56.8 7.3
51.3 4.1
50.4 3.7
39.7 1.6
38.4 2.0
57.7 1.3
57.6 1.7
64.2 2.5
64.7 2.5
Uncertain Prompt
84.0 3.6
83.0 4.0
73.8 4.9
68.2 3.5
31.5 0.2
24.8 1.2
30.7 1.2
23.0 1.6
56.9 2.2
56.8 2.1
62.8 2.9
63.1 2.6
Majority Vote
80.4 4.0
80.4 4.0
69.5 1.2
69.5 1.2
42.4 2.3
42.4 2.3
32.7 3.4
32.7 3.4
58.7 3.1
58.7 3.1
68.6 0.6
68.6 0.6
Appendix
Table 9: Performance of different methods across models and datasets (Decentralized MAD, N=8 ).
Method
Arithmetics
GSM8K
HH-RLHF
Form.Log.
Pro.Med.
CSQA
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
R=1
R=2
Qwen2.5-1.5B
Single-agent
42.0 7.1
38.2 2.1
26.0 1.9
24.2 0.6
55.6 2.8
60.4 1.7
Society Of Mind
63.5 0.7
72.0 4.2
63.0 1.4
62.7 4.2
33.3 1.4
33.2 1.6
32.5 4.5
33.7 3.9
57.8 1.9
57.6 1.9
63.9 2.5
64.4 2.2
MAD-M 2
44.0 24.8
37.7 26.4
59.3 2.6
50.7 9.6
50.4 4.7
50.9 2.3
34.7 0.8
34.9 0.5
57.7 1.7
57.5 1.9
63.9 2.5
64.3 2.4
Uncertain Prompt
67.5 4.9
60.5 9.2
61.5 2.6
57.7 5.2
34.0 2.4
31.5 0.7
35.3 3.9
31.7 4.5
57.4 2.2
57.6 2.0
63.7 2.6
64.0 2.6
Appendix
Table 10: Performance of different methods across models and datasets (Sparse MAD).
Figure 11: Diversity of retained responses across retention strategies on Qwen2.5-3B. Similar result for Qwen2.5-1.5B is provided in Appendix A.6 .
Figure 12: Diversity of retained responses across retention strategies on Qwen2.5-1.5B.
Model
Dataset
Society Of Mind
DAR (Ours)
Qwen2.5-1.5B
Arithmetics
51 ± 5
58 ± 7 ( × 1.14)
Form.Log.
45 ± 1
51 ± 3 ( × 1.13)
Qwen2.5-3B
Arithmetics
82 ± 11
85 ± 4 ( × 1.03)
Form.Log.
108 ± 14
113 ± 3 ( × 1.05)
Appendix
Table 11: Wall-clock time per debate round (in second) on Arithmetics and Formal Logic tasks ( N=4 ). Values for our method include × , indicating the ratio relative to Society Of Mind.
Figure 13: Complete qualitative responses on Qwen2.5-1.5B (Majority Vote).
Figure 14: Complete qualitative responses on Qwen2.5-1.5B (Society Of Mind).
Figure 15: Complete qualitative responses on Qwen2.5-1.5B (DAR).
Multi-agent debate (MAD) improves the reasoning capabilities of large language models by having multiple agents iteratively refine their responses through discussion. However, MAD suffers from a critical vulnerability known as shared misconception: when a majority of agents initially converge on an incorrect answer, the debate process tends to amplify rather than correct the error. Existing methods primarily address peer skew but leave the agents' inherently biased concept priors unaddressed. To mitigate this systematic weakness, we propose R2-MAD (Remember and Reweight for Multi-Agent Debate), a framework that equips agents with an experience memory accumulated from past debates. R2-MAD intervenes on both failure modes through two complementary mechanisms: A debate-state-aware retrieval policy dynamically calibrates the concept prior by retrieving relevant historical evidence based on the current consensus level. Then these retrieved experiences provide a basis for estimating per-agent reliability, yielding confidence weights to modulate peer influence. Experiments on various benchmarks show that R2-MAD achieves consistent improvements over existing single-agent and MAD baselines.
Xuanfa Jin, Zhijian Ma, Yongcheng Zeng +3
Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences · University College London
Multi-agent debate has shown promise for improving the reasoning of large language models, yet recent theory suggests its benefits are highly regime-dependent. While interaction can amplify informative signals under corrective conditions, symmetric debate dynamics are neutral in expectation, often making majority voting preferable. We reconcile these views by arguing that debate is effective only when invoked at the right time and with appropriate structure. Based on this insight, we propose LASE: Leader-Adaptive Structured Engagement, a leader-centric multi-agent debate framework that selectively engages interaction only in non-neutral regimes. LASE introduces an asymmetric leader-supporter architecture that enables directed information flow and selective signal amplification, while defaulting to simple aggregation otherwise. Experiments across diverse reasoning benchmarks show that LASE achieves multi-agent-level performance with near single-agent token cost, substantially improving efficiency over static debate and voting baselines.
Seungdong Yoa, Ye Seul Sim, Suhee Yoon +5
LG AI Research, Seoul, South Korea · School of Computer Science and Engineering, Chung-Ang University, Seoul, South Korea · Division of Computer Engineering, Hankuk University of Foreign Studies, Yongin, South Korea +1
Multi-agent debate can improve large language model reasoning by eliciting diverse hypotheses and critiques, yet its performance is often constrained by weak moderation. Common pipelines rely on fixed budgets, agreement-based stopping, or untrained judges, leading to redundant deliberation and unreliable evidence aggregation. We cast moderation as a meta-cognitive process, monitoring debate utility, controlling deliberation, and adjudicating a final answer, and introduce Meta-Moderator, a learnable framework that dynamically regulates debate and decides when to finalize an answer. Meta-Moderator is trained independently of the debaters via outcome-driven policy optimization, making debate regulation an explicit capability rather than an incidental effect of prompting. Across five benchmarks, Meta-Moderator outperforms widely used decision layers and transfers across tasks and system configurations. Further analyses show that it allocates debate more selectively and reduces mis-aggregation after informative hypotheses appear.
Wentao Hu, Zhuoyue Wan, Jinhao Shen +3
The Hong Kong Polytechnic University · Sichuan University