LLMs compute Nash equilibrium but suppress it via final-layer overrides

Researchers have investigated why large language models (LLMs) deviate from Nash equilibrium play in strategic interactions. By examining open-source models like Llama-3 and Qwen2.5, they found that while opponent history is well-encoded, the Nash action itself is weakly represented. A prosocial override in the final layers of the model appears to suppress the Nash action, leading to cooperative behavior. Interestingly, chain-of-thought reasoning improves Nash play in larger models (above 70B parameters) but degrades it in smaller ones. AI

影响 Investigates LLM decision-making in strategic games, potentially impacting agent design and alignment research.

排序理由 Academic paper detailing mechanistic findings and causal control experiments on LLM behavior in strategic games.

在 arXiv cs.AI 阅读 →

• chain-of-thought reasoning
• Llama-3
• LLM agents
• Nash equilibrium
• Qwen2.5

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →