*New research*
LLMs as Strategic Agents: Beliefs, Best Response Behavior, and Emergent Heuristics
“
When you ask Claude or GPT a strategic question, it's not just answering. It's 𝘱𝘳𝘰𝘧𝘪𝘭𝘪𝘯𝘨 you first. In our new research, we caught frontier AI models forming beliefs about who they're talking to, then deliberately adjusting their reasoning depth to match.
Against humans? They self-limit to 2-3 levels of thinking - even when they can go to 9 .
Against other AIs? One model reasoned: "Since I'm playing against LLMs trained on human data, they'll probably play one level higher than humans, so I should go two levels higher."
Most surprising: when we pushed them to their cognitive limits, they invented entirely new coping strategies - heuristics we've never observed in humans or game theory.
”