Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MetaMind, a three-agent metacognitive framework (ToM agent, moral agent, response agent) that wraps LLMs to improve theory-of-mind and social reasoning across ToMBench, social cognition tasks, and STSS.
Problem
LLMs struggle to infer unspoken intentions, emotions, and beliefs and to respond appropriately in socially nuanced, ambiguous, and culturally sensitive interactions, treating social reasoning as single-step prediction.
Approach
A staged multi-agent pipeline inspired by metacognition: generate multiple mental-state hypotheses, refine them with cultural/ethical/role norms and select via a composite plausibility/information-gain score, then generate and self-validate a response conditioned on the selected hypothesis and social memory. The framework wraps prompted LLMs without weight updates.
Key results
- On ToMBench, MetaMind raises GPT-4 average accuracy from 74.8% to 81.0%, above all compared prompting baselines.
- On the social cognition suite, MetaMind reaches 80.5% average vs 71.5% for base GPT-4 (about 9% gain).
- On STSS, MetaMind achieves 73.9% average score vs 39.4% for GPT-4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MetaMind | GPT-4 · ToMBench | accuracy (AVG.): 74.8 -> 81.0 | Helped |
| Add MetaMind | GPT-4 · Social cognition tasks | accuracy (AVG.): 71.5 -> 80.5 | Helped |
| Add MetaMind | GPT-4 · STSS | score (AVG.): 39.4 -> 73.9 | Helped |
| Add MetaMind | Claude 3.5 Sonnet · ToMBench | accuracy: 70.7 -> 81.0 | Helped |
| Add MetaMind | DeepSeek-R1 · ToMBench | accuracy: 86.0 -> 88.6 | Helped |
| Add MetaMind | OpenAI o1 · ToMBench | accuracy: 88.6 -> 90.3 | Helped |
| Add MetaMind | OpenAI o3 · ToMBench | accuracy: 90.3 -> 92.2 | Helped |
| Add Stage 1 mental-state reasoning (ablation) | MetaMind · Social cognition tasks | accuracy (AVG.): 77.9 -> 80.5 | Helped |
| Add Stage 2 norm-aware refinement (ablation) | MetaMind · Social cognition tasks | accuracy (AVG.): 76.7 -> 80.5 | Helped |
| Add Stage 3 response generation and validation (ablation) | MetaMind · Social cognition tasks | accuracy (AVG.): 79.1 -> 80.5 | Helped |
| Add social memory (ablation) | MetaMind · Social cognition tasks | accuracy (AVG.): 75.4 -> 80.5 | Helped |
| Add Stage 1 mental-state reasoning (ablation) | MetaMind · STSS | score (Avg.): 68.3 -> 73.9 | Helped |
Surprising findings
- Removing the response validation stage drops STSS average by 16.1 points, the largest ablation effect.
- Base GPT-4 scores 2.3% and 1.2% on two STSS categories, showing very low baseline on some social goal tasks.
Limitations
- Performance depends on backbone LLM capability and on the quality of domain knowledge and memory coverage.
- Benchmarks are constrained textual scenarios and do not capture multimodal cues or long-term relationships.
- Claims of matching human performance rest on selected dimensions and limited human comparison.
Experimental setup
Datasets: ToMBench STSS SocialIQA SOTOPIA
Benchmarks: ToMBench Social cognition tasks (UOT, SIT, PST, FBT, AST, HT, SST, FRT) STSS SocialIQA SOTOPIA
Metrics: accuracy normalized success score overall score
Baselines: Chain-of-Thought Hypothetical Minds (HM) ToM2C Generative Agents SymbolicToM TDP
Builds on: chain-of-thought prompting SymbolicToM Hypothetical Minds Generative Agents
Closest papers by meaning
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.91)
- Language Models Represent Beliefs of Self and Others (ICML 2024 · similarity 0.90)
- MuMA-ToM: Multi-modal Multi-Agent Theory of Mind (AAAI 2025 · similarity 0.90)
- AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.90)
- Thought Communication in Multiagent Collaboration (NeurIPS 2025 · similarity 0.89)
- OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.