[ad_1]
They’re good, however not perfect cooperators it appears, at the very least not with out the right prompts:
Massive Language Fashions (LLMs) are reworking society and permeating into numerous functions. In consequence, LLMs will regularly work together with us and different brokers. It’s, due to this fact, of nice societal worth to know how LLMs behave in interactive social settings. Right here, we suggest to make use of behavioral sport concept to review LLM’s cooperation and coordination conduct. To take action, we let completely different LLMs (GPT-3, GPT-3.5, and GPT-4) play finitely repeated video games with one another and with different, human-like methods. Our outcomes present that LLMs typically carry out effectively in such duties and in addition uncover persistent behavioral signatures. In a big set of two players-two methods video games, we discover that LLMs are significantly good at video games the place valuing their very own self-interest pays off, just like the iterated Prisoner’s Dilemma household. Nonetheless, they behave sub-optimally in video games that require coordination. We, due to this fact, additional concentrate on two video games from these distinct households. Within the canonical iterated Prisoner’s Dilemma, we discover that GPT-4 acts significantly unforgivingly, all the time defecting after one other agent has defected solely as soon as. Within the Battle of the Sexes, we discover that GPT-4 can not match the conduct of the straightforward conference to alternate between choices. We confirm that these behavioral signatures are secure throughout robustness checks. Lastly, we present how GPT-4’s conduct might be modified by offering additional details about the opposite participant in addition to by asking it to foretell the opposite participant’s actions earlier than making a selection. These outcomes enrich our understanding of LLM’s social conduct and pave the way in which for a behavioral sport concept for machines.
Right here is the full paper by Elif Akata, et.al.
[ad_2]