Making AI 'smarter' actually makes it a worse simulator of human behavior.
April 16, 2026
Original Paper
When Reasoning Models Hurt Behavioral Simulation: A Solver-Sampler Mismatch in Multi-Agent LLM Negotiation
arXiv · 2604.11840
The Takeaway
This research reveals a 'Solver-Sampler Mismatch': as LLMs get better at reasoning, they stop acting like humans and start acting like hyper-rational game theorists. In social simulations, they over-optimize for winning, losing the 'bounded rationality' and emotional compromise that define human interaction. This is a major blow to using high-end models for social science or market research. If your goal is to simulate how a person would react, GPT-4 might actually be *too smart* to be realistic. This forces practitioners to rethink how they 'dumb down' or constrain models to keep them human-like. It challenges the assumption that intelligence scaling is always a net positive.
From the abstract
Large language models are increasingly used as agents in social, economic, and policy simulations. A common assumption is that stronger reasoning should improve simulation fidelity. We argue that this assumption can fail when the objective is not to solve a strategic problem, but to sample plausible boundedly rational behavior. In such settings, reasoning-enhanced models can become better solvers and worse simulators: they can over-optimize for strategically dominant actions, collapse compromise