Researchers built 'general' LLM agents with theory‑grounded instructions and a small set of human 'seed' games, then tested them across 883,320 novel games. In preregistered tests, these agents predicted human play better than game‑theoretic equilibria, out‑of‑the‑box agents, and even the most relevant published human data for select new games. This shows LLM‑driven simulations can transport behavioral insight to new settings without ad hoc tweaks.
— If AI agents can reliably forecast human choices, social‑science methods, policy testing, and regulation could shift toward simulation‑first evaluation.
Alexander Kruel
2025.09.06
72% relevant
It links to a 'pathbreaking paper' on AI simulations of human behavior (via Marginal Revolution), matching the claim that LLM agents can forecast human choices in novel games better than traditional baselines.
Tyler Cowen
2025.09.03
100% relevant
Tyler Cowen’s summary of the Manning & Horton paper reporting preregistered performance gains over equilibria and human benchmarks on 1,500 sampled games from a 883k‑game population.
← Back to All Ideas