Exploration and Exploitation Errors Are Measurable for Language Model Agents

This research introduces a method to systematically quantify exploration and exploitation errors in Language Model (LM) agents, addressing the challenge of evaluation without access to internal policies. It proposes controllable environments and a policy-agnostic metric to measure these errors, revealing flaws even in state-of-the-art LMs.