Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Simulating Society Requires Simulating Thought

Chance Jiajie Li, Jiayi Wu, Zhenze Mo, Ao Qu, Yuhan Tang, Kaiya Ivy Zhao, Yulu Gan, Jie Fan, Jiangbo Yu, Jinhua Zhao, Paul Liang, Luis Alonso, Kent Larson

TL;DR

The paper argues that simulating complex society with large language models requires cognitively grounded reasoning rather than surface-level plausibility. It introduces GenMinds, a framework for structured belief representations via causal motifs and belief graphs, and RECAP, a benchmark for reasoning fidelity including traceability, demographic grounding, and intervention coherence. By shifting from output-centric prompts to a cognition-centric paradigm grounded in causal, compositional, and revisable reasoning, the work offers a path toward agents that simulate how people think, not just what they say. The proposed framework aims to enable transparent diagnostics, pluralistic modeling of public reasoning, and principled evaluation for high-stakes social simulations, while outlining open challenges and next steps for the field.

Abstract

Simulating society with large language models (LLMs), we argue, requires more than generating plausible behavior; it demands cognitively grounded reasoning that is structured, revisable, and traceable. LLM-based agents are increasingly used to emulate individual and group behavior, primarily through prompting and supervised fine-tuning. Yet current simulations remain grounded in a behaviorist "demographics in, behavior out" paradigm, focusing on surface-level plausibility. As a result, they often lack internal coherence, causal reasoning, and belief traceability, making them unreliable for modeling how people reason, deliberate, and respond to interventions. To address this, we present a conceptual modeling paradigm, Generative Minds (GenMinds), which draws from cognitive science to support structured belief representations in generative agents. To evaluate such agents, we introduce the RECAP (REconstructing CAusal Paths) framework, a benchmark designed to assess reasoning fidelity via causal traceability, demographic grounding, and intervention consistency. These contributions advance a broader shift: from surface-level mimicry to generative agents that simulate thought, not just language, for social simulations.

Simulating Society Requires Simulating Thought

TL;DR

The paper argues that simulating complex society with large language models requires cognitively grounded reasoning rather than surface-level plausibility. It introduces GenMinds, a framework for structured belief representations via causal motifs and belief graphs, and RECAP, a benchmark for reasoning fidelity including traceability, demographic grounding, and intervention coherence. By shifting from output-centric prompts to a cognition-centric paradigm grounded in causal, compositional, and revisable reasoning, the work offers a path toward agents that simulate how people think, not just what they say. The proposed framework aims to enable transparent diagnostics, pluralistic modeling of public reasoning, and principled evaluation for high-stakes social simulations, while outlining open challenges and next steps for the field.

Abstract

Simulating society with large language models (LLMs), we argue, requires more than generating plausible behavior; it demands cognitively grounded reasoning that is structured, revisable, and traceable. LLM-based agents are increasingly used to emulate individual and group behavior, primarily through prompting and supervised fine-tuning. Yet current simulations remain grounded in a behaviorist "demographics in, behavior out" paradigm, focusing on surface-level plausibility. As a result, they often lack internal coherence, causal reasoning, and belief traceability, making them unreliable for modeling how people reason, deliberate, and respond to interventions. To address this, we present a conceptual modeling paradigm, Generative Minds (GenMinds), which draws from cognitive science to support structured belief representations in generative agents. To evaluate such agents, we introduce the RECAP (REconstructing CAusal Paths) framework, a benchmark designed to assess reasoning fidelity via causal traceability, demographic grounding, and intervention consistency. These contributions advance a broader shift: from surface-level mimicry to generative agents that simulate thought, not just language, for social simulations.

Paper Structure

This paper contains 37 sections, 2 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. The Rise of LLMs in Social Simulation.
  3. An Oversimplified Paradigm: Demographics In, Behavior Out.
  4. Structural Failures: Modeling, Evaluation, and Calibration
  5. Toward Mechanistic and Individual-Level Alignment
  6. Position and Vision
  7. Social Simulation: Opportunities and Gaps
  8. Problem Statement: Beyond Behavioral Plausibility in Generative Social Simulation
  9. Fidelity: Coherence, Traceability, and Causal Grounding
  10. Traceability and Interpretability
  11. Alternative View: Post-hoc rationalization may be cognitively authentic and functionally sufficient.
  12. Counterfactual Intervention Sensitivity and Belief Revision
  13. Alternative View: Human cognition is non-monotonic and contextually fluid, thus demanding coherence is unrealistic.
  14. Individuality: Heterogeneity and Positional Reasoning in Social Simulation
  15. Illusion of Consensus in Multi-Agent Systems
  16. ...and 22 more sections

Figures (2)

  • Figure 1: From surface imitation to cognitively grounded social simulation. Current LLM-based simulations (top left) capture only surface opinions, shaped by demographics or language patterns, while the deeper belief formation processes remain unmodeled (bottom left, beneath the waterline). This yields population-level simulations that are flattened and stereotyped, reflecting aggregated personas rather than genuine diversity. In contrast, cognitively grounded reasoning (bottom right) models the latent belief dynamics behind individual decisions, producing collective patterns that are heterogeneous, interpretable, and causally faithful.
  • Figure 2: Motif-based belief graph and intervention. Natural language responses are parsed into motif-level causal links, forming a personalized belief graph. A simulated intervention on Transparency propagates downstream updates, shown as highlighted nodes and edges.