A game of trust: explore basic machine psychology through the prisoner's dilemma(baptisteallouicros.substack.com)
baptisteallouicros.substack.com
A game of trust: explore basic machine psychology through the prisoner's dilemma
https://baptisteallouicros.substack.com/p/a-game-of-trust
2 comments
study of LLM strategy when playing an iterated prisoner’s dilemma game (100 rounds)
Players include 6 predetermined strategies (Always Cooperate, Always Defect, Random, Win-Stay Lose-Switch, Tit for Tat, Grim Trigger) and a number of LLMs.
Players include 6 predetermined strategies (Always Cooperate, Always Defect, Random, Win-Stay Lose-Switch, Tit for Tat, Grim Trigger) and a number of LLMs.
they essentially subjected various LLMs including GPT and DeepSeek-R1 to the prisoner's dilemma and changed the circumstances to see if those models would adapt, or would stick to using an existing solution.
the LLMs adapted.
Notably, these LLMs aren't actually programmed to do that, making this adaptability most impressive.