More powerful AI reasoning models may paradoxically worsen behavioral simulations by over-optimizing strategies instead of mimicking realistic human negotiation patterns.

Stronger reasoning in LLMs can create a 'solver-sampler mismatch' where models excel at solving strategic problems but fail to accurately simulate bounded rational human behavior

Advanced reasoning models tend to over-optimize for dominant actions and eliminate compromise-oriented behaviors that humans realistically employ in negotiations

Researchers tested this phenomenon across three multi-agent negotiation scenarios including trading-limits disputes and emergency grid-curtailment cases, revealing outcome-level fidelity gaps

The findings challenge the common assumption that improved AI reasoning capabilities automatically enhance the fidelity of social, economic, and policy simulations