LLM reasoning models fail behavioral simulation in multi-agent negotiation

A new research paper explores the mismatch between reasoning capabilities and behavioral simulation in large language models used for multi-agent negotiation. The study found that models like DeepSeek and OpenAI's GPT-5.2, when used for their reasoning abilities, often defaulted to authority-driven outcomes rather than negotiated ones. The paper suggests that evaluating models based on their intended behavioral role, rather than just strategic capability, is crucial for accurate institutional simulations. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Highlights the need to evaluate LLMs for specific behavioral roles in simulations, not just raw strategic capability.

RANK_REASON The cluster contains an arXiv paper detailing research findings on LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• DeepSeek
• GPT-5.2
• OpenAI
• Sandro Andric

• paper
• other