Research map / Large language models / Agents, code and math

Simulating Society Requires Simulating Thought

NeurIPS 2025 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks · 1 citations (OpenAlex)

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.

Position paper arguing that LLM-based social simulation should model cognitively grounded, causal, compositional, revisable belief structures, proposing the GenMinds conceptual framework and the RECAP evaluation schema.

Problem

Current LLM-based social simulations follow a behaviorist "demographics in, behavior out" paradigm, producing surface-plausible but flattened, stereotyped, causally unfaithful agent behavior that lacks traceable reasoning and belief revision.

Approach

Reframes generative social simulation as cognitive modeling: agents hold structured causal belief graphs built from semi-structured interviews (cognitive motifs), support counterfactual belief updates via a symbolic-neural hybrid, and are evaluated on reasoning fidelity (traceability, counterfactual adaptability, motif compositionality) via the RECAP schema.

Key results

• Illustrative example only: a do(Transparency=high) intervention on a single interview-derived belief graph shifts P(Privacy Concern) from 0.7 to 0.3 and P(Opposition to Surveillance) from 0.7 to 0.2; no empirical benchmark results are reported.

Limitations

• No empirical validation; RECAP and GenMinds are conceptual and the benchmark is described only as a schema.
• Constructing causal belief networks from natural language is ambiguous in concept identification, causal direction, polarity, and granularity.
• Causality alone does not capture associative, analogical, or emotional reasoning.

Experimental setup

Metrics: motif alignment belief coherence counterfactual robustness

Builds on: chain-of-thought persona prompting reinforcement learning from human feedback do-calculus

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 0.88)
• MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
• Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.87)
• Social Choice Foundations for Simulation-Augmented Generation (NeurIPS 2026 · similarity 0.87)
• Language Models Represent Beliefs of Self and Others (ICML 2024 · similarity 0.87)
• AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.86)