Skip to content

MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MetaMind is a three-stage multi-agent framework that decomposes social reasoning into Theory-of-Mind hypothesis generation, norm-aware moral refinement, and validated response generation. It achieves state-of-the-art on ToM and social cognition benchmarks, enabling LLMs to match human-level performance on key ToM tasks for the first time. The framework is model-agnostic and shows consistent gains across 16+ LLM backbones.

Problem

LLMs default to literal semantic interpretation and fail to infer the unobservable mental states (beliefs, desires, intentions, emotions) underlying indirect or ambiguous human communication. Existing approaches—role-play prompting, RLHF fine-tuning, chain-of-thought—treat social reasoning as a single-step prediction and lack the structured, multi-stage metacognitive process humans use to interpret socially nuanced language.

Method

MetaMind implements a staged metacognitive pipeline via three specialized agents:

  1. ToM Agent (Stage 1): Given user prompt \(u_t\), conversational context \(C_t\), and a dynamic social memory \(M_t\), it generates \(k\) candidate mental-state hypotheses \(H_t = \{h_1, \ldots, h_k\}\) labeled by type \(T = \{\text{Belief, Desire, Intention, Emotion, Thought}\}\) via four steps: commonsense hypothesis generation, cross-reference with social memory, ToM marker identification, and candidate enumeration.

  2. Moral Agent (Stage 2): Takes \(H_t\) plus a constraint set \(D\) (cultural norms, ethical rules, role-based expectations) and produces revised hypotheses \(\tilde{h}_i\). Selects the optimal hypothesis via a composite score balancing contextual plausibility and information gain: \(\tilde{h}^* = \arg\max_i [\lambda \cdot P(\tilde{h}_i | u_t, C_t, M_t) + (1-\lambda) \cdot \log \frac{P(\tilde{h}_i | u_t, C_t, M_t)}{P(\tilde{h}_i)}]\).

  3. Response Agent (Stage 3): Generates response \(o_t\) conditioned on \(\tilde{h}^*\) and \(M_t\), then self-validates with a utility score \(U(o_t) = \beta \cdot \text{Empathy}(o_t, u_t, M_t) + (1-\beta) \cdot \text{Coherence}(o_t, C_t, \tilde{h}^*)\), triggering regeneration if the score is too low. Social memory is updated after each interaction.

Key Contributions

  • MetaMind multi-agent framework grounded in metacognitive theory, integrating mental-state inference, ethical constraint filtering, and adaptive social memory into a unified pipeline.
  • Comprehensive evaluation across 16+ LLMs on ToMBench (ToM reasoning), eight social cognition tasks, and STSS (open-ended social simulation), achieving SOTA on all three.
  • Ablation study isolating each stage's contribution, confirming all three components and social memory are necessary.
  • Demonstration that MetaMind-enhanced LLMs match or approach human performance on ToM dimensions (Belief: 89.3 vs. 88.6 human; NL Communication: 89.0 vs. 88.5; Desire: 78.2 vs. 76.5).

Results

  • ToMBench (ToM reasoning): MetaMind+GPT-4 achieves 81.0% average vs. 74.8% base GPT-4 (+6.2%); outperforms CoT (73.6%), SymbolicToM (75.4%), Generative Agents (75.1%).
  • Social cognition (8 tasks): MetaMind+GPT-4 achieves 80.5% average vs. 71.5% base GPT-4 (+9.0%); largest gains in AST (+11.3%), SIT (+11.4%), FRT (+10.6%).
  • STSS social simulation: MetaMind achieves 73.9% average vs. 39.4% base GPT-4 (+35.7% real-scenario improvement claim in abstract, +34.5% absolute here); leads all baselines (best prior: TDP at 54.4%).
  • Advanced models: MetaMind boosts DeepSeek-R1 from 86.0% to 88.6%, OpenAI o3 from 90.3% to 92.2%, Claude 3.5 Sonnet from 70.7% to 81.0% on ToM accuracy.
  • Ablations: Removing Stage 3 causes the largest single drop (−16.1% on STSS); removing social memory causes −5.1% on STSS.

Limitations

  • Performance is contingent on backbone LLM capability; absolute gaps between small and large models persist after MetaMind enhancement.
  • Social memory and domain-knowledge coverage require adaptation for diverse cultural norms and evolving social expectations at deployment scale.
  • Benchmarks are text-only; real-world social interaction involves multimodal cues (tone, facial expressions), complex group dynamics, and long-term relationship building not captured here.
  • Constraint set \(D\) for the Moral Agent must be manually specified; generalization to novel or culturally unfamiliar norms is not evaluated.

Relevance to Agentic AI / LLM Agents

MetaMind is a direct instance of structured multi-agent reasoning pipelines where agents have distinct, complementary roles—relevant to any agentic system that must model user intent rather than just execute instructions. The metacognitive loop (hypothesize → constrain → validate/regenerate) is a reusable design pattern for agentic systems operating in socially complex or ambiguous environments such as empathetic dialogue, negotiation, or collaborative task planning. The explicit social memory component addresses a key open problem in long-horizon agentic interaction: maintaining coherent, personalized context across turns. The model-agnostic design and demonstrated gains even on frontier reasoning models (o3, DeepSeek-R1) suggest this pipeline is a lightweight, drop-in enhancement for existing agent architectures.