Should You Use Your LLM to Explore or Exploit?

Keegan Harris, Aleksandrs Slivkins

Keywords: bandits, exploration, exploitation

Abstract: We evaluate the ability of the current generation of large language models (LLMs) to help a decision-making agent facing an exploration-exploitation tradeoff. We use LLMs to explore and exploit in silos in various (contextual) bandit tasks. We find that while the current LLMs often struggle to exploit, in-context mitigations may be used to substantially improve performance for small-scale tasks. However even then, LLMs perform worse than a simple linear regression. On the other hand, we find that LLMs do help at exploring large action spaces with inherent semantics, by suggesting suitable candidates to explore.

Primary Area: reinforcement learning

Submission Number: 3203