Can AI decompose social reasoning into distinct cognitive stages?

MetaMind treats social reasoning not as a single-step prediction but as a layered metacognitive process — the same staged interpretation-reflection-adaptation loop that psychology identifies in human social cognition. Three specialized agents each handle a distinct cognitive stage:

1. Theory-of-Mind Agent generates multiple hypotheses about the user's mental state (intent, emotion, belief) from contextual and social cues. When a user says "work has been exhausting lately," the system produces competing hypotheses — burnout, frustration, need for empathy — rather than committing prematurely to one interpretation.

2. Moral Agent filters and revises these hypotheses against cultural norms and ethical constraints. If romantic intent is hypothesized in a professional conversation, the Moral Agent reinterprets it as collegial admiration based on workplace norms. This is not censorship but social calibration — the same process humans perform when refining first impressions.

3. Response Agent generates output conditioned on the refined hypothesis and the user's social memory (emotional patterns, prior preferences), then self-validates for coherence and empathy.

The framework achieves 35.7% improvement on real-world social scenarios and 6.2% gain in ToM reasoning, notably enabling LLMs to match average human performance on key ToM tasks for the first time. Ablation studies confirm all three stages are necessary — removing any one degrades performance.

The key design insight: prior approaches treat social reasoning as surface-level statistical alignment (static role-play prompting, preference fine-tuning). MetaMind instead explicitly models the structured, multi-stage cognitive process humans use to reason about unobservable intent. This is the difference between mimicking social behavior and modeling the cognitive architecture that produces it.

Since Why do reasoning models struggle with theory of mind tasks?, MetaMind's success suggests the solution is not more reasoning effort but differently structured reasoning — decomposed into social-cognitive stages rather than extended chain-of-thought.

Since Do large language models genuinely simulate mental states?, MetaMind's multi-hypothesis approach directly addresses the shallow strategy problem: generating competing hypotheses forces the system beyond pattern-matching to genuine mental state simulation.

Source: Role Play Paper: MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems