Scaling LLMs Improves Social Simulation Fidelity in Most Cases, But Fails on Cognitive Biases

This research paper investigates whether scaling up Large Language Models (LLMs) improves the fidelity of social simulations. Using scaling laws and a suite of 85 transformer LLMs (Qwen3 architecture) trained on DCLM web text, the authors study three sub-domains: opinion modeling, behavioral simulation, and longitudinal forecasting. They find strong compute scaling in most settings — behavioral and opinion simulation tasks improve rapidly with scale, especially for populations well-represented in English web corpora. However, longitudinal forecasting and underrepresented opinions scale more sl