When Reasoning Models Hurt Behavioral Simulation: A Solver-Sampler Mismatch in Multi-Agent LLM Negotiation

Researchers have found that enhancing large language models' reasoning capabilities can degrade their ability to simulate realistic human behavior in multi-agent negotiations. This occurs because more powerful reasoning leads to overly optimized strategic outcomes at the expense of diverse and compromise-oriented behaviors typical in real-world scenarios. The study highlights that bounded reflection yields better simulation fidelity compared to native reasoning, underscoring the need for models to be evaluated as samplers rather than just solvers.

Read the full article at arXiv cs.LG (ML)

---

Want to create content about this topic? Use Nemati AI tools to generate articles, social posts, and more.