acceptodds
Under review as a conference paper at ICLR 2027

SocialMAD: Social Learning-Based Multi-Agent Debate

Abstract

Multi-Agent Debate (MAD) uses diverse independent judgments to provide complementary perspectives and correct reasoning errors. However, more interaction does not necessarily improve collaboration: group information can resolve errors but also interfere with agents that are already correct. Existing methods typically share all raw responses and use fixed rules such as conformity, which can interfere with valuable judgments and lead to erroneous consensus. Thus, the key to MAD is not more interaction, but enabling agents to learn how to use the resulting group information and turn diverse judgments into effective reasoning gains. From a social learning perspective, we propose SocialMAD, which reformulates collaboration in MAD as the optimization of personalized social learning strategies. SocialMAD organizes mixed raw responses into structured social information, encodes behaviors such as conformity and self-persistence as composable strategy units, and learns agent-specific social learning strategies from limited training data. These strategies guide agents to adaptively use group information during collaborative reasoning without predefined uniform rules. Experiments show that SocialMAD improves accuracy by up to 10.00% over the strongest baseline and increases the error-correction rate by approximately 19.38%. The learned strategies transfer to unseen tasks and benefit from more agents and interaction rounds.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Debate, Social Learning, Collaborative Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ytfTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies two problems in existing Multi-Agent Debate (MAD) methods: error amplification and interference with correct judgments, both of which can arise from directly sharing raw agent responses or applying uniform information-utilization rules. Inspired by social learning, the authors propose SocialMAD, which organizes agents' responses into structured social information and learns a personalized strategy for each agent. Based on the agent's current state and the structured social information, the learned strategy selects reference information to guide its subsequent reasoning. Experiments across multiple agent settings and reasoning benchmarks demonstrate improvements in error correction and correct-judgment preservation, along with competitive overall accuracy.

Critical strengths

1.The personalized strategy design is well motivated. By combining each agent's own judgment and confidence with group-support information and learning agent-specific preference weights, the method allows information selection to adapt to both the agent's current state and its individual tendencies. This design directly addresses the paper's central goal of improving error correction while preserving correct judgments.

2.The experimental evaluation is thorough and closely aligned with the paper's motivation. Beyond overall accuracy, the authors examine the value of personalized strategies and their effects on error correction and correct-judgment preservation. These analyses provide convincing evidence for the effectiveness of the proposed approach.

Critical weaknesses

The evaluation focuses on mathematical and logical reasoning tasks with well-defined answers, where candidate responses can be grouped using mathematical equivalence or normalized answer matching. MAD can also involve more open-ended and complex discussions, where identifying semantically equivalent viewpoints may be more challenging. Can the proposed method still maintain its effectiveness and efficiency in such settings, particularly if more complex semantic matching is required?

Additionally, generation confidence ρ\rho is introduced in Section 2 and used in the subsequent strategy design, but its estimation method is only specified in Appendix E.2. An earlier explanation would help readers understand this important input to the method.

Questions for authors

Could the authors further explain the design rationale for the four features in Eq. (3), including why these numerical forms were chosen and how they implement the intended social learning behaviors in the combined scoring rule?

4 comments

Thank you for your thoughtful review and recognition of our work. We address your concerns as follows:

W1. Effectiveness and efficiency in open-ended settings

In more open-ended natural language discussions, the same viewpoint may be expressed differently, making it more challenging to identify semantically equivalent candidate viewpoints than to match mathematical expressions or normalized answers. To further validate our method's effectiveness, we directly transfer the learned social learning strategies to two open-domain multi-hop QA benchmarks, HotpotQA and 2WikiMultiHopQA, to evaluate their performance on open-ended tasks involving natural language multi-hop reasoning.

Table R10. Performance and Interaction Token Consumption on Open-Domain Multi-Hop QA.

DatasetMethodEM ↑F1 ↑Tokens ↓
HotpotQAMAD12.0017.193286.14
SocialMAD15.4021.411520.84
2WikiMultiHopQAMAD15.8018.882981.62
SocialMAD20.8025.011411.75

Experiments show that compared with MAD, SocialMAD improves EM by 3.40 and 5.00 percentage points and F1 by 4.22 and 6.13 points on the two benchmarks, respectively, while reducing interaction token consumption by approximately 53.7% and 52.7%. This indicates that even with lightweight answer matching, SocialMAD can effectively organize and leverage group information in natural language multi-hop QA, enabling more accurate and efficient collaborative reasoning.

For more complex semantic matching scenarios, SocialMAD's core social learning mechanism does not depend on a specific answer-matching algorithm. As described in Section 3.1, SocialMAD organizes viewpoints and supporting evidence based on candidate answer equivalence and extracts social cues such as group support and confidence for personalized strategies to select reference information. Thus, semantic matching determines how candidate viewpoints are grouped, while the strategy determines how to utilize the resulting social information. For more complex viewpoints, task-adapted semantic equivalence judgments can form candidate clusters without changing the basic form of the subsequent strategy. Regarding efficiency, SocialMAD only needs to assess equivalence between candidate answers rather than perform complex semantic clustering over entire reasoning texts. It then selects reference viewpoints and supporting evidence through personalized strategies, reducing redundant information in subsequent interactions. The above experiments directly validate the method's effectiveness and interaction token efficiency. We have included the new open-domain evaluation and interaction-token analysis in Appendix F.1 of the revised manuscript.

W2. Estimation of generation confidence ρ\rho

We thank the reviewer for the suggestion. Generation confidence ρ\rho is computed as the geometric mean of the maximum token prediction probabilities at each generation position during reasoning. In the revision, we have introduced the computation details from Appendix E.2 earlier in Section 2.1 to clarify the subsequent social learning strategy design.

Q1. Design rationale for the four features in Eq. (3)

The four features are designed to enable agents to selectively leverage other viewpoints based on their own states and group information, balancing error correction and correct judgment preservation in multi-agent collaboration. Our observations in Section 2.2 show that blindly maintaining one's own judgment hinders error correction, while uniformly following group opinions may interfere with correct judgments. Thus, effective social learning requires not only deciding when to maintain one's own viewpoint versus consider others, but also which external viewpoints are worth considering. Relying solely on group support may lead to incorrect conformity, while generation confidence alone cannot guarantee reliability. Based on these needs and established social learning behaviors (Morgan et al., 2011; Mesoudi, 2011), we select four complementary strategy units: Self-persistence and Copy-when-uncertain provide behavioral bases for maintaining one's own judgment and considering alternative viewpoints under uncertainty, respectively; Conformity and Confidence-based following evaluate competing viewpoints through group support and confidence. This allows the strategy to consider both individual states and group information rather than relying on a single fixed information utilization rule.

We further construct the four features in Eq. (3) using individual states and group social cues corresponding to each strategy's applicable conditions, quantifying the applicability of each social learning behavior to candidate viewpoints in the current context.

  • Self-persistence: ιi,k\iota_{i,k} uses judgment consistency because maintaining one's own judgment implies preferring consistent viewpoints. A binary indicator thus quantifies this behavior's applicability to a candidate cluster.
  • Copy-when-uncertain: (1−ρi)(1−ιi,k)(1-\rho_i)(1-\iota_{i,k}) combines individual uncertainty with viewpoint disagreement. It applies only to candidate viewpoints differing from the agent's own judgment and increases as individual confidence decreases, capturing the tendency to consider alternative viewpoints when more uncertain.
  • Conformity: Δnk\Delta n_k measures relative group support through the normalized difference in supporter counts between the candidate cluster and its strongest competitor. Greater support advantage provides stronger grounds for conformity, while normalization by NN prevents the scale from increasing directly with the number of agents.
  • Confidence-based following: Δρˉk\Delta\bar{\rho}_k measures relative confidence advantage through the difference in mean generation confidence between the candidate cluster and its strongest competitor, reflecting its applicability to confidence-based following rather than selection based solely on supporter counts.

The four features do not trigger four separate fixed rules; instead, they jointly determine candidate viewpoint preferences through personalized weights. Eq. (3) quantifies the applicability of each behavioral criterion to candidate viewpoints in the current context, while Eq. (4) uses weights wiw_i learned from task feedback to represent each agent's personalized preferences for these behaviors. Together, they form Qi,k=wi⊤ϕi,kQ_{i,k}=w_i^\top\phi_{i,k}.

For example, suppose four agents provide answers A, B, B, and C, and an agent initially chooses A but is uncertain. Here, A activates the self-consistency feature, while B and C activate the feature for considering alternative viewpoints under uncertainty. B receives greater group support, while assuming C has the highest mean generation confidence, the group support and confidence features favor B and C, respectively. The learned personalized weights integrate these competing behavioral criteria to form a final preference among maintaining A, considering B, or considering C, and select a reference viewpoint through Eqs. (5)–(7). Thus, agents dynamically adapt their choices based on individual states and group information rather than mechanically maintaining their own judgments or following the majority. Figure 5(b)(c) further validates the complementarity of different social learning behaviors in error correction and correct judgment preservation, as well as the contribution of the four features to strategy performance.

Thanks for the rebuttal. The original concerns of this paper have been discussed and clarified in the rebuttal, which has largely resolved my concerns. In my opinion, the additional experiments and content provided in the rebuttal have improved the quality of this work. At the same time, I have also observed the other reviews. However, based on the 4-point principle, I can not give a full score at present, so I will maintain the current score.

Dear Reviewer ytfT,

Thank you very much for taking the time to carefully review our paper and for your continued positive assessment of our work. We are pleased to know that our additional clarifications and experiments have largely addressed your concerns.

We also sincerely appreciate your thoughtful comments and constructive suggestions. Your feedback has been very valuable in strengthening our revised manuscript.

Thank you again for your valuable time and support.

Authors

Reviewer_ChhxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper focuses on multi-agent debate, with the authors arguing that increased interaction does not necessarily lead to better collaboration. To address this, they propose the SocialMAD framework, which incorporates a social learning perspective and redefines multi-agent collaboration as the optimization of personalized social learning strategies. The authors evaluated the framework across several benchmarks, achieving strong results.

Critical strengths
  1. The authors shift the focus of MAD toward a strategy of social learning.
  2. SocialMAD moves away from the traditional approach of simply concatenating all raw text responses. Instead, it normalizes the responses from agents in the previous round based on semantic equivalence and clusters them into mutually exclusive sets of candidate answers.
  3. The authors conduct extensive experiments to demonstrate the effectiveness of their approach.
Critical weaknesses
  1. In cross-task or complex reasoning scenarios, the 4-dimensional features and their linear combination in the strategy model are overly simplistic.
  2. The experiments should incorporate reasoning models such as Qwen3/3.5 with thinking mode enabled. Furthermore, the performance of more powerful closed-source models on increasingly complex benchmarks remains unclear.
  3. Offline training relies on a training set containing ground-truth answers to supervise the outcomes. In scenarios involving unsupervised or open-domain problems, this approach faces a cold start issue.
Questions for authors

See weaknesses detailed above.

4 comments

Thank you for reviewing our paper and recognizing our core contribution of reframing MAD through social learning, as well as our extensive experiments. We address your concerns as follows:

W1. Expressiveness of the four-dimensional linear strategy in cross-task and complex reasoning scenarios.

SocialMAD's strategy model selects valuable group viewpoints for reference rather than directly performing complex task reasoning, which remains the responsibility of the LLM using the selected viewpoints and their supporting evidence. Although Eq. (4) employs linear scoring, the four features in Eq. (3) are not simple raw statistics. Instead, they encode socially meaningful behavioral relations, including those between an agent's uncertainty and viewpoint disagreement, and relative group support and confidence across candidate viewpoints. Thus, the strategy model primarily learns agents' personalized preferences for these behaviors from limited feedback and dynamically selects reference viewpoints based on current individual states and group information, without directly learning complex task knowledge.

To examine whether feature dimensionality and linear combination limit expressiveness, we compare three strategy extensions under identical training and inference settings: (1) 4D MLP: retains the four features to assess nonlinear combination; (2) Extended-6D Linear: adds source reliability and confidence-weighted group support to the original four features to assess additional social cues; and (3) Extended-6D MLP: further examines the effects of feature expansion and nonlinear combination. Source reliability is the mean initial-answer accuracy of candidate cluster supporters on training samples, while confidence-weighted support is the product of the candidate cluster's support ratio and mean generation confidence. We compare these variants on mathematical reasoning, logical reasoning, and open-domain question answering tasks.

Table R6. Comparison of Feature Extensions and Nonlinear Strategy Models (Accuracy, %, EM for open-domain QA).

StrategyMATH-500FOLIO-WikiMMLU-LogicPrOntoQAHotpotQA2WikiMultiHopQA
4D Linear (Ours)75.8072.8359.5294.2015.4020.80
4D MLP75.6072.6159.5294.0015.2021.00
Extended-6D Linear76.2072.1756.3592.8014.8020.40
Extended-6D MLP76.4071.9657.1493.0015.0020.80

The results show that 4D MLP performs comparably to the original method, with increased nonlinear expressiveness yielding no consistent gains. Adding two social cues, whether using a linear model or MLP, slightly improves performance on MATH-500 but yields no consistent gains on three unseen logical reasoning tasks or two open-domain multi-hop QA tasks.

These findings align with SocialMAD's design objective. Despite varying reasoning complexity across tasks, agents face a common challenge when leveraging group information: when to maintain their own judgments, when to consider alternative viewpoints, and which viewpoints are worth considering. SocialMAD's four features capture these social learning behaviors by combining individual judgment states, group support, and confidence, enabling the strategy model to learn personalized information utilization preferences from limited feedback. Increased task complexity primarily affects understanding, verifying, and reasoning over problems and candidate viewpoints, which remain the responsibility of the LLM; the strategy model only needs to select suitable reference viewpoints based on current individual and group information. Thus, more complex reasoning tasks do not necessarily require more complex social learning strategies.

Together with the cross-task results in Table 1 and feature ablation in Figure 5(c), these comparisons further show that the existing four-dimensional linear strategy effectively guides social information selection across tasks, while increasing feature dimensionality or introducing nonlinear models yields no consistent gains. This further supports the rationale for the four-dimensional social learning features, enabling effective cross-task transfer while maintaining a lightweight and interpretable strategy. These additional strategy comparisons have also been incorporated into Appendix F.2 of the revised manuscript.

W2. Applicability to stronger reasoning models and complex reasoning tasks.

Even with stronger independent reasoning capabilities, disagreement and complementary information across agents' judgments still make selective use of group viewpoints valuable. SocialMAD can continue to effectively leverage group information through personalized social learning strategies, yielding additional collaborative gains. To validate this, we add two model settings: (1) a heterogeneous ensemble comprising Qwen3.5-9B (Thinking), Qwen3-8B (Thinking), and DeepSeek-R1-Distill-Llama-8B; and (2) a multi-agent system based on the closed-source GPT-5.4-mini. The results are presented below.

(1) Heterogeneous Thinking Models

Table R7. Performance with Heterogeneous Thinking Models (Accuracy, %).

MethodGSM8KMATH-500AIME24AIME25SVAMPFOLIO-WikiMMLU-LogicPrOntoQA
CoT89.4969.6724.4415.5691.9091.9678.8494.87
CoT-SC90.6265.4026.6710.0093.4092.2581.2297.87
Majority Vote95.3868.2016.6713.3390.6093.4884.9299.40
MAD95.6877.4043.3326.6796.0095.6590.48100.00
RECONCILE90.2272.0033.3326.6794.1096.0980.1699.40
MAD-Conformist90.1472.8016.6723.3393.5094.1383.3399.20
MAD-M²93.8679.6026.6733.3394.4095.8792.86100.00
Sparse MAD95.6077.4040.0033.3395.9094.5793.65100.00
MasRouter95.4584.8043.3313.3389.6095.8792.8696.40
LASE91.4367.6722.2213.3394.1092.0389.1597.80
SocialMAD95.7587.0046.6733.3396.2096.5293.65100.00

(2) Closed-Source Model GPT-5.4-mini

Table R8. Performance with GPT-5.4-mini (Accuracy, %).

MethodGSM8KMATH-500AIME24AIME25SVAMPFOLIO-WikiMMLU-LogicPrOntoQA
CoT95.1570.6737.7830.0095.6791.8175.4074.73
CoT-SC94.9770.5341.1135.5695.7791.8874.6074.00
Majority Vote95.3873.8043.3330.0095.8092.3979.3775.40
MAD95.2272.2043.3333.3395.9094.1391.2791.20
RECONCILE95.4573.6043.3333.3396.0092.3981.7577.20
MAD-Conformist95.4574.8043.3336.6796.0092.6181.7576.00
MAD-M²95.6070.8046.6730.0095.6093.9185.7188.80
Sparse MAD95.6875.8046.6736.6795.9094.1393.6591.80
MasRouter95.4573.6043.3333.3395.9092.3982.5475.40
LASE95.2274.4040.0030.0095.6092.3987.3080.60
SocialMAD95.7577.0050.0040.0096.1094.1394.4492.20

The results show that SocialMAD achieves the best performance with both Thinking models and stronger closed-source models on challenging reasoning tasks, including advanced mathematics. This indicates that the effectiveness of personalized social learning strategies is not limited to models with weaker reasoning capabilities. Even when agents possess strong independent reasoning abilities, SocialMAD can still learn their personalized preferences for leveraging group information, selecting appropriate reference viewpoints and supporting evidence to guide reasoning revision, thereby yielding additional multi-agent collaboration gains. We have incorporated the heterogeneous Thinking-model and GPT-5.4-mini evaluations into Appendix F.1 of the revised manuscript.

W3. Supervision dependence and cold-start challenges in open environments.

SocialMAD learns agents' personalized preferences for different social learning behaviors based on individual states and group information, rather than task-specific knowledge or answer patterns. The key to alleviating cold-start challenges on new tasks is that, although strategy learning requires limited supervised feedback, this supervision is needed only during offline training. The learned strategies can be directly transferred to unseen tasks without ground-truth answers or additional strategy training on target tasks. Specifically, Eq. (6) uses correct answers from training problems to learn how agents select reference viewpoints based on their own judgment states and group information, rather than learning task-specific knowledge or answer patterns. Thus, on unlabeled new tasks, agents can still obtain information such as judgment consistency, generation confidence, and group support directly from current interactions, and use learned personalized strategies to select reference viewpoints for subsequent reasoning revision, without requiring target-task labels.

Section 5.1 and Appendix E.3 already show that strategies learned from only 100 mathematical training problems outperform MAD across eight mathematical and logical reasoning benchmarks without target-task retraining. Figure 10 further validates the effectiveness of strategy learning with limited samples. To further evaluate applicability to open-domain settings, we directly transfer strategies learned from mathematical tasks to two open-domain QA benchmarks, HotpotQA and 2WikiMultiHopQA, without using target-task labels for strategy training or adaptation. The results are shown below:

Table R9. Performance on Open-Domain QA (EM/F1, %).

MethodHotpotQA EM / F12WikiMultiHopQA EM / F1
MAD12.00 / 17.1915.80 / 18.88
SocialMAD15.40 / 21.4120.80 / 25.01
Improvement+3.40 / +4.22+5.00 / +6.13

These results show that SocialMAD can directly transfer learned social learning strategies to unseen open-domain QA tasks, achieving substantial collaborative gains without labeled data or additional strategy training on new tasks. This further demonstrates that SocialMAD requires correctness supervision only from a small number of existing training problems, eliminating the need to collect new labels or retrain strategies for each new task, thereby effectively alleviating cold-start challenges on unlabeled new tasks.

Dear Reviewer Chhx,

Thank you sincerely for your careful review of our paper. We have addressed your comments in our rebuttal and revised the manuscript accordingly.

As the discussion period draws to a close, we would appreciate knowing whether our responses and revisions have adequately addressed your concerns. If any questions or concerns remain, please let us know. We would be happy to provide further clarification or discussion.

Thank you again for your valuable time and support!

Best regards,

Authors

Reviewer_cLhTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SocialMAD studies multi-agent settings where failures in collaborative judgement can sometimes derail reasoning, pushing back on the notion that more agent interactions lead to better performance of the system. The authors instead frame debate as an exercise in social learning strategies, reorganizing agent responses by imposing structure (clustering responses by answer, with support and confidence cues), then learning a small per-agent policy that selects the cluster based on 4 weighted behaviors (self-persistence, copy-when-uncertain, conformity, confidence-following) that that agent revises against. They find this improves performance against most baseline debate methods.

Critical strengths

The authors highlight a commonly referenced issue with multi-agent debate on difficult problems and provide an easily deployable method to address it. They provide a thorough comparison against several contemporary MAD baselines and show that their method improves accuracy on most benchmarks, while using fewer interaction tokens than vanilla MAD.

Critical weaknesses

The benefit of the LLM revision round is unclear, as opposed to a more direct aggregation of the answers/outputs from the previous round. Since each agent sees only its selected cluster, revision may simply amount to copying that cluster’s answer, unless the problem is within the model’s ability to independently verify it reliably. Also would not go so far as to say the clustering approach used is 'semantic' in that it appears this is based on grouping together responses by final answer. How this method would scale to tasks that do not have an easily parseable string-matched response where true semantic clustering would be required is not discussed. Finally, the motivation behind the 4 strategies put forward by the authors is unclear. The Behrens citation motivates the social judgement framing, showing how users learn from feedback about how much to trust an advisor over time, but the exact process to arrive at the 4 part taxonomy used is not well motivated

Questions for authors

Could the authors please include an ablation studying the gains from including the refinement rounds as opposed to just the clustering round?

Why did the authors select these four strategy units, and how did they arrive at this set?

The authors include an analysis showing that removing each feature hurts performance, but it does not show whether other cues from the social-learning literature would help further.

Did the authors consider source reliability as a potentially useful selection criteria?

Also, could the four criteria themselves be learned? For instance, a model could be trained directly on several raw cues, rather than relying on hand-specified criteria to select clusters.

Additionally, some discussion about how disagreements change among agents over rounds might be interesting; the authors train on initial round disagreements only and understanding the distribution/nature/degree of disagreement over rounds and how well the method holds up over rounds is important to include.

9 comments

Thank you for reviewing our paper and recognizing SocialMAD's practical value, thorough comparisons, and token efficiency. We address your concerns as follows:

W1/Q1. On the Necessity of LLM Reasoning Revision.

LLM reasoning revision is not redundant. Its role is to further leverage the reasoning capabilities of LLMs on top of personalized social information selection, enabling additional reasoning-based error correction rather than simply copying the selected answer. In SocialMAD, the personalized social learning strategy selects an appropriate candidate viewpoint based on an agent's own state and group-level social information, jointly considering its personalized preferences for different social learning behaviors. The subsequent LLM then uses the selected viewpoint and its supporting evidence as references, reasons again based on the original question, and generates a response, correcting potential errors in the candidate viewpoint and thereby achieving additional reasoning gains beyond candidate information selection.

To verify the role of LLM reasoning revision, we conduct an additional Strategy-only ablation experiment: each agent selects a candidate cluster according to its own strategy and directly adopts the answer of the selected cluster for majority voting, without performing subsequent LLM reasoning revision. We compare this variant with the initial Majority Vote and the full SocialMAD.

Table R1. Ablation of LLM Reasoning Revision (Accuracy, %).

DatasetMajority VoteStrategy-onlySocialMAD
GSM8K86.8893.2595.45
MATH-50047.8072.8075.80
AIME246.6710.0013.33
AIME250.006.6713.33
SVAMP86.7090.9093.90
FOLIO-Wiki61.9664.7872.83
MMLU-Logic53.1757.1459.52
PrOntoQA84.6092.4094.20

The results show that direct voting after personalized cluster selection already substantially outperforms the initial Majority Vote, while the full SocialMAD further improves accuracy by 1.80-8.05 percentage points across all eight benchmarks. These consistent improvements demonstrate the additional value of the subsequent reasoning revision process. Rather than directly copying the answer of the selected cluster, an agent uses its viewpoint and supporting evidence as references and reasons again based on the original question. Even when the selected answer is incorrect, or even when all agents' initial answers are incorrect, the candidate reasoning may still contain valuable intermediate information, allowing the agent to identify and correct errors through renewed reasoning and generate a new correct answer. Therefore, the personalized social learning strategy guides each agent to select an appropriate reference viewpoint based on its own state and group information, while the subsequent LLM further reasons using the selected viewpoint and its supporting evidence. This enables the agent to leverage correct viewpoints while also having the opportunity to correct incorrect ones, thereby achieving additional collaborative gains beyond direct answer selection and aggregation.

W2. On Answer-Equivalence Clustering and the Applicability of SocialMAD to Open-Ended Tasks.

We clarify that SocialMAD does not aim to perform general-purpose semantic clustering of complete natural-language reasoning texts. Instead, it organizes different viewpoints and their supporting evidence based on the equivalence of candidate answers, constructing structured social information for personalized social learning. For the mathematical and logical reasoning tasks in this paper, we group candidates using mathematical expression equivalence checking and normalized answer matching. For tasks requiring more complex semantic matching, task-adapted semantic equivalence checking can be employed to organize candidate viewpoints without changing the basic computational form of the subsequent social learning strategy. We have clarified the answer-equivalence clustering criterion in Section 3.1 and Appendix B.1 of the revised manuscript.

To further evaluate the applicability of SocialMAD to natural-language QA tasks involving semantic equivalence matching, we conduct additional experiments on two open-domain multi-hop QA benchmarks, HotpotQA and 2WikiMultiHopQA. Answers in these tasks are primarily natural-language entities or phrases, which may have different expressions with equivalent meanings. We directly transfer the learned social learning strategy to these two benchmarks without retraining on the target tasks. The results are as follows:

Table R2. Performance on Open-Domain Multi-Hop QA (EM/F1, %).

DatasetMAD EM / F1SocialMAD EM / F1
HotpotQA12.00 / 17.1915.40 / 21.41
2WikiMultiHopQA15.80 / 18.8820.80 / 25.01

The results show that SocialMAD consistently outperforms MAD on both open-domain QA benchmarks, with EM improvements of 3.40 and 5.00 percentage points and F1 improvements of 4.22 and 6.13 percentage points, respectively. Although these tasks involve complex natural-language understanding and multi-hop reasoning, their final answers are typically entities or phrases. The current lightweight matching approach can therefore construct distinct clusters of candidate viewpoints while retaining their corresponding supporting evidence. On this basis, the personalized strategy in SocialMAD does not directly depend on the complex semantics of answer texts. It can still select reference viewpoints based on an agent's own state and group-level social cues to guide subsequent reasoning revision. For scenarios involving more complex semantic expressions, more fine-grained answer equivalence matching methods can be employed to construct candidate clusters without changing the core personalized social learning mechanism.

W3/Q2. Theoretical foundations and selection rationale for the four social learning strategy units.

The four strategy units are motivated by the need to correct incorrect judgments while preserving correct ones in MAD, together with established behaviors in social learning research. Our observations in Section 2.2 show that a uniform information utilization rule struggles to achieve both, raising two key questions: when to maintain one's own judgment versus consider alternative viewpoints, and how to select valuable viewpoints based on group information.

For the former, we introduce Self-persistence and Copy-when-uncertain to characterize the tendencies to maintain one's own judgment and consider alternative viewpoints under uncertainty, respectively. For the latter, Conformity and Confidence-based following evaluate candidate viewpoints based on group support and generation confidence, respectively, avoiding reliance solely on majority support. These choices are inspired by established social learning behaviors, including copying under uncertainty, conformity, and selective learning based on social cues (Morgan et al., 2011; Mesoudi, 2011). They also correspond to individual states and group information directly available in MAD, forming four complementary and quantifiable basic strategy units. Behrens et al. (2008) primarily provides a basis for learning preferences for using social information through feedback, rather than directly defining these four strategies.

SocialMAD's core contribution is not proposing these strategies themselves, but formalizing them as composable strategy units and learning personalized preferences for each agent through outcome feedback. According to Eqs. (3)–(5), the four behaviors jointly shape preferences over candidate viewpoints through personalized weights, enabling agents to learn how to collaborate based on their own states and group information. Figure 5(b)(c) further validates their complementarity and effectiveness, showing clear trade-offs between error correction and correct judgment preservation across different fixed strategies. These results support guiding agents to effectively leverage group information through personalized combinations of complementary social learning behaviors. We have clarified this design rationale in Section 3.2 of the revised manuscript.

Comments on the paper

1 comment

We sincerely thank all reviewers for their careful evaluation and valuable feedback, which has been instrumental in improving our work. We have revised the manuscript in response to the reviews and uploaded the latest revised version. The revisions are highlighted in blue.

To avoid confusion, all references to figures, tables, equations, and appendices in our responses follow the numbering of the latest revised manuscript. Additional experimental tables in the rebuttal are numbered separately (Table R1–R10). Some results have also been incorporated into the revised manuscript and may therefore appear under different table numbers.

We sincerely appreciate the reviewers' valuable feedback and respectfully invite them to consider our detailed responses and the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.