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-agent pipeline—Theory-of-Mind Agent, Moral Agent, Response Agent—that decomposes social reasoning into staged hypothesis generation, normative refinement, and self-validated generation. It achieves state-of-the-art results on ToM, social cognition, and social simulation benchmarks, enabling LLMs to match human performance on key ToM tasks for the first time. The architecture is explicitly grounded in psychological metacognition theory rather than monolithic prompting.

Problem

LLMs handle literal semantics well but fail at social reasoning: inferring unspoken intent, emotion, and belief from ambiguous, culturally situated language. Prior approaches (role-play prompting, RLHF fine-tuning, CoT) treat social reasoning as a single-step prediction, missing the layered, self-correcting cognitive process humans use—and fail to generalize across cultural/social contexts.

Method

MetaMind routes each input through three sequentially chained agents:

  1. ToM Agent (Stage 1): Given input \(X = (u_t, C_t, M_t)\) (utterance, conversation context, social memory), generates a set of \(k\) candidate mental-state hypotheses \(H_t = \{h_1, \ldots, h_k\}\) typed over \(T = \{\text{Belief, Desire, Intention, Emotion, Thought}\}\) via four-step mental-state reasoning (commonsense → memory cross-reference → ToM marker identification → hypothesis generation).

  2. Moral Agent (Stage 2): Revises each \(h_i\) against constraint rules \(D\) (cultural norms, ethical constraints, role-based expectations) and selects the best revised hypothesis via: $\(\tilde{h}^* = \arg\max_i \left[\lambda \cdot P(\tilde{h}_i | u_t, C_t, M_t) + (1-\lambda) \cdot \log \frac{P(\tilde{h}_i | u_t, C_t, M_t)}{P(\tilde{h}_i)}\right]\)$ balancing contextual plausibility and information gain.

  3. Response Agent (Stage 3): Generates \(o_t\) conditioned on \(\tilde{h}^*\) and social memory \(M_t\), then self-validates via: $\(U(o_t) = \beta \cdot \text{Empathy}(o_t, u_t, M_t) + (1-\beta) \cdot \text{Coherence}(o_t, C_t, \tilde{h}^*)\)$ triggering regeneration if \(U(o_t)\) falls below threshold.

A Social Memory module (dynamic database of user preferences and emotional markers) is shared across all three stages and updated after each interaction.

Key Contributions

  • MetaMind multi-agent framework with three cognitively specialized, sequentially collaborating agents grounded in metacognition theory.
  • First demonstration of LLMs matching average human performance on key ToM benchmarks via a training-free, model-agnostic wrapper.
  • Comprehensive evaluation across 16 LLMs on three benchmark families (ToMBench, 8-task social cognition suite, STSS social simulation).
  • Ablation isolating the necessity of all three stages and the social memory component.

Results

  • ToMBench (ToM reasoning): MetaMind + GPT-4: 81.0% vs. base GPT-4 74.8% (+6.2%); outperforms CoT, SymbolicToM, Generative Agents, ToM2C across all six ToM categories.
  • Social cognition (8-task suite): MetaMind 80.5% vs. base GPT-4 71.5% (+9.0%); largest gains on AST (+11.3%), SIT (+11.4%), FRT (+10.6%).
  • STSS social simulation: MetaMind 73.9% vs. base GPT-4 39.4% (+34.5% absolute, ~88% relative); strongest gains in Appointment (+63.8 pp) and Inviting Companions (+64.8 pp).
  • Ablation: removing Stage 1 → −2.6% social cognition avg, −5.6% STSS avg; removing Stage 3 → −22.7% STSS avg; removing Social Memory → −5.1% social cognition avg, −10.0% STSS avg.

Limitations

  • Evaluation uses GPT-4 as primary backbone; open-source model results are shown only on ToMBench, not the full benchmark suite.
  • Constraint rules \(D\) in the Moral Agent are hand-crafted; no mechanism for automatic rule induction from new cultural contexts.
  • Multi-agent pipeline incurs 3× inference cost per query at minimum; latency not reported.
  • Social Memory update mechanism is not fully described in the main text; unclear how stale or conflicting memories are handled.
  • No evaluation on adversarial or deceptive social scenarios where ToM reasoning could be exploited.

Relevance to Harnesses / Meta-Harnesses

MetaMind is a textbook example of a domain-specific meta-harness: it wraps arbitrary backbone LLMs in a fixed three-stage orchestration layer without modifying model weights, composing specialized subagents (ToM, Moral, Response) through structured inter-agent message passing and a shared memory store. The staged pipeline with a self-validation loop at Stage 3 (utility score → conditional regeneration) directly instantiates harness-level quality control, analogous to critic/verifier patterns in code-generation harnesses. The Social Memory module acts as a cross-turn state store—functionally equivalent to a working memory managed by the harness rather than any individual agent. This paper is relevant as a case study in harness design for a non-coding domain, demonstrating that harness-enforced hypothesis refinement and normative filtering can substitute for task-specific fine-tuning.