acceptodds
Withdrawn from ICLR 2027

SocialAgent: From Collaborative Reasoning to Self-Improving Social Intelligence

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Artificial Social Intelligence (ASI) refers to the ability of machines to understand and respond to human social behavior, constituting an important dimension of general artificial intelligence. Such ability depends on understanding how different participants perceive the same situation and drawing on experience from previous interactions. However, current multimodal large language models (MLLMs) largely reason about each interaction in isolation, without explicitly reconciling participants' perspectives or accumulating reusable social knowledge. We introduce SocialAgent, a framework that develops social reasoning through collaborative experience acquisition and its subsequent internalization into a single model. SocialAgent-Scaffold combines Role-based Collaboration (RoC) with an Evolving Experience Pool (EEP) without parameter updates. RoC reasons from individual participants' perspectives and reconciles their complementary evidence, producing explicit demonstrations of perspective taking. The EEP abstracts successful reasoning processes into reusable role and reasoning templates, turning individual solutions into strategies that can guide new interactions. Although this collaboration provides structured reasoning demonstrations, repeatedly simulating multiple roles incurs substantial inference cost. SocialAgent-Evolve therefore internalizes this experience into a single-agent MLLM through Strategy-Mediated Self-Refinement (SMSR). Its Strategy-guided Learning stage trains the student on scaffold-generated trajectories paired with their guiding EEP strategies. The trajectories demonstrate how to solve concrete cases, while the strategies make generalizable reasoning principles explicit. This pairing enables the student to learn perspective taking without replaying multi-role collaboration at inference time. Since imitation alone leaves gaps beyond the demonstrated solutions, Multi-strategy Policy Optimization then targets the student's remaining failures, using alternative EEP strategies to guide exploration of diverse reasoning paths. Together, these stages connect learning from collaborative demonstrations with failure-driven self-refinement, without relying on a stronger external teacher. Extensive experiments on social reasoning benchmarks demonstrate that SocialAgent achieves state-of-the-art performance on both open-ended generation and multiple-choice question answering.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Social Intelligence, Multi-modal Large Language Models, Self-Improvement

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_u9YROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SocialAgent connects collaborative social reasoning with single-model improvement. SocialAgent-Scaffold uses role analysis, parallel role-specific reasoning, and arbitration to answer video questions without parameter updates. A curator abstracts correctly answered training examples into an Evolving Experience Pool (EEP) of reusable strategies. SocialAgent-Evolve then applies Strategy-guided Learning (SGL) to successful scaffold outputs paired with strategies, followed by Multi-strategy Policy Optimization (MSPO) on student failures. MSPO samples under alternative strategies, normalizes rewards across each question's rollouts, and updates a shared policy. At inference, the student retrieves one strategy and generates one response. The authors report improvements on Social-IQ 2.0 and scaffold-only MuMA-ToM evaluation, together with reduced inference latency after internalization.

Critical strengths
  1. S1: A coherent connection between collaboration, memory, and learning. Sections 3.2–3.3 (pp. 4–7) explain how the same experience pool guides both collaborative reasoning and subsequent student refinement. Using the same backbone for scaffold agents and the student makes the approach relevant to settings without a stronger external teacher.

  2. S2: Encouraging reported performance and inference savings. Table 1 (p. 8) reports Social-IQ 2.0 multiple-choice accuracy increasing from 63.49% for the backbone to 71.04% for Scaffold and 83.49% for Evolve. Evolve also exceeds the same-backbone vanilla SFT+GRPO baseline by 3.13 points. The reported latency reduction from 72.29 s to 15.35 s illustrates a practical benefit of replacing repeated collaboration with single-agent inference.

  3. S3: A useful controlled exploration comparison. Table 4 (pp. 8–9) holds the total at eight rollouts per question while varying strategy allocation; two strategies with four rollouts improve accuracy by 2.02 points over one strategy with eight. Appendix C (pp. 14–15) also specifies a frozen training-derived EEP and excludes same-video/question entries during MSPO retrieval. These are meaningful controls, although they do not resolve the separate issues below.

Critical weaknesses
  1. W1: MuMA-ToM overall scores cannot be reconciled with the stated evaluation. Table 1 (p. 8) reports Scaffold category accuracies of 81.68%, 54.95%, and 46.57%, whose mean is 61.07%, versus an Overall of 56.33%. The backbone similarly gives 54.27% by averaging its categories, versus 49.78% reported. The original MuMA-ToM benchmark contains 300 questions per category, so full-set overall accuracy should equal that mean. Section 4.1 and Appendix C do not specify alternative subsets or denominators. This is a substantial, potentially repairable reporting/protocol issue; it does not erase the reported improvements within every category. Provide evaluated counts, aggregation rules, and corrected values if needed. The broad state-of-the-art wording (pp. 1, 3) also needs qualification: the benchmark paper reports LIMP at 76.6% overall. Its different backbone and inference procedure preclude a simple controlled comparison, but Table 1 establishes superiority only over the selected baselines.

  2. W2: The contribution of pairing strategies with demonstrations is not isolated. Section 3.3.1 (p. 6, Eqs. 6–7) attributes value to combining scaffold trajectories with their strategies. However, every row in Table 3 (p. 8) already starts with SGL. Table 1 includes vanilla SFT, but neither Section 4.1 nor Appendix C defines whether that baseline uses identical successful examples, scaffold reasoning targets, optimization settings, and test-time retrieval. Consequently, the SGL improvement cannot be assigned specifically to strategy pairing rather than demonstration quality, filtering, or retrieval differences. This is a moderate, repairable attribution limitation, not evidence that the complete system is ineffective. Clarify the baseline recipe; if it is not matched, compare trajectory-only learning with strategy-paired learning under otherwise equivalent conditions, or narrow the mechanism claim.

  3. W3: The open-ended evaluation protocol is insufficiently specified for the headline comparison. Section 4.1 (p. 7) describes GPT-3.5-turbo judging, whereas the supplied Role-Play/Arbiter prompts (Appendix D, Fig. 7–8, p. 16) include answer choices and request an answer index. No separate open-ended generation prompt or exact scoring prompt is provided. It therefore remains unclear whether open-ended responses are generated without choices, conditioned on choices, or obtained by converting selected options into text. These settings test different capabilities and affect interpretation of the 56.10% versus 54.90% VEGAS comparison. Giving options to the judge is consistent with the cited evaluation tradition and is not itself an error. This is a substantial but potentially clarification-resolvable reproducibility issue. Provide the generation/scoring data flow, evaluation split and sample counts, and identify which comparator scores were recomputed on identical examples versus taken from prior reports.

  4. W4: Closely related experience-learning and multi-agent distillation work is missing. Section 2 (pp. 3–4) emphasizes social benchmarks and multimodal models but does not discuss Zhao et al.'s ExpeL: LLM Agents Are Experiential Learners (AAAI 2024), which extracts reusable insights from training experiences without parameter updates, or Luo et al.'s AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent (first posted February 2026), which filters multi-agent reasoning traces and studies supervised and reinforcement-based distillation. These overlap with EEP and the collaboration-to-student motivation, respectively, and are absent from the bibliography (pp. 10–12). Differences remain: ExpeL targets interactive agents, while AgentArk uses different teacher/distillation machinery; SocialAgent adds social role decomposition, explicit strategy conditioning, and failure-focused cross-strategy exploration. This is a moderate positioning gap, not proof of redundancy. Subject to the applicable cutoff, explain those distinctions and identify the methodological advance beyond combining existing components; missing citations alone do not mandate new experiments.

Questions for authors
  1. W1: What category sample counts and aggregation rule produce the MuMA-ToM Overall column? Can the discrepancy with the category means be reconciled from existing evaluation records?
  2. W2: Does vanilla SFT use the same filtered scaffold examples and reasoning targets as SGL? What differs in training configuration and test-time strategy retrieval?
  3. W3: Please specify whether answer choices are visible during open-ended generation, what text is scored, and the exact judge prompt. Which split/examples are evaluated, and which baseline results were reproduced under that setup?
  4. W4: What distinguishes EEP and SMSR from ExpeL-style experience abstraction and AgentArk-style multi-agent distillation beyond the social-video application?
Reviewer_rGv1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework called SocialAgent for improving multimodal social reasoning with collaborative experience acquisition and agent learning. First, multi-agent analysis social interactions from different perspectives and collect successful reasoning strategies into an experience pool. Then, the author uses these successful trajectories to finetune a model and apply GRPO to further improve the model on failed questions.

Critical strengths
  1. high empirical improvement: SocialAgent compares with multiple baseliens and have improvement on them.

  2. clear motivation: the idea of using multiple agents to generate better data and utilize it to train one single model is natural and reasonable.

  3. useful ablation studies: the paper includes ablations on different multi-agent components, training settings and number of strategies, which are very useful.

Critical weaknesses
  1. overclaim about self-improving: The paper claims to achieve self-improving social intelligence. However, there is no iterative process in which the improved model generates better experiences and further improve itself. Therefore, claiming it to be self-improving is overstated.

  2. unclear benefits of the experience pool. The paper shows that adding the experience pool improves accuracy, but it is unclear whether the collected strategies are actually better than simple hand-written or randomly retrieved strategies.

  3. lack comparison with compute-matched data generation pipelines: The main improvement may simply come from spending additional inference compute to generate and filter more training trajectories. However, the paper does not compare against simpler approaches, such as Best-of-N sampling or rejection sampling followed by the same SFT and GRPO pipeline. Therefore, the effectiveness of the main contribution of the paper is not well-studied.

  4. limited evaluation: The framework is only completely evaluated on Social-IQ 2.0. On MuMA-ToM, the authors only evaluate SocialAgent-Scaffold without the experience pool. Therefore, it is unclear whether the proposed training framework generalizes to other tasks.

Questions for authors

What happens if the experience pool is replaced with random strategies or simple hand-written reasoning instructions?

Reviewer_sfAvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SocialAgent, a framework for improving multimodal social reasoning through collaborative reasoning and experience reuse. SocialAgent-Scaffold combines Role-based Collaboration (RoC), where multiple agents reason from different perspectives and an arbiter integrates their findings, with an Evolving Experience Pool (EEP) that stores reusable reasoning strategies. SocialAgent-Evolve introduces Strategy-Mediated Self-Refinement (SMSR), consisting of Strategy-guided Learning (SGL) that fine-tunes a single model on successful reasoning trajectories and their guiding strategies, and Multi-strategy Policy Optimization (MSPO) that uses alternative strategies to explore remaining failure cases through GRPO. Experiments on Social-IQ 2.0 and MuMA-ToM show improvements over several baselines, with reduced inference costs compared to multi-agent collaboration.

Critical strengths
  • Proposed method improves performance on SocialIqa and MuMa-ToM
Critical weaknesses
  • I found the paper quite difficult to follow. It introduces many terms and components, as reflected in the summary above, but the motivation for each component and why it is necessary is not sufficiently explained.

  • The performance improvements seem to come from combining several different mechanisms, rather than necessarily improving the model's social reasoning abilities. It is unclear whether the gains are due to better perspective taking and social understanding or simply the additional reasoning, retrieval, and training strategies introduced by the framework.

  • I'm also not a big fan of the figure styles in this paper. The figures are overly crowded with information, making them difficult to follow.

Questions for authors

N/a

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

Related papers

Discussion (0)

Sign in to comment.