Search papers, labs, and topics across Lattice.
The paper introduces a novel framework to quantify exploration and exploitation errors in Language Model (LM) agents operating in partially observable 2D grid environments with unknown task DAGs. They design environments with adjustable difficulty to emphasize exploration or exploitation challenges and develop a policy-agnostic metric to evaluate agent actions. Experiments with various LM agents reveal performance limitations and distinct failure modes, highlighting the importance of reasoning and harness engineering for improved exploration and exploitation.
Even state-of-the-art language models struggle with exploration and exploitation in complex decision-making tasks, but minimal harness engineering can significantly improve performance.
Language Model (LM) agents are increasingly used in complex open-ended decision-making tasks, from AI coding to physical AI. A core requirement in these settings is the ability to both explore the problem space and exploit acquired knowledge effectively. However, systematically distinguishing and quantifying exploration and exploitation from observed actions without access to the agent's internal policy remains challenging. To address this, we design controllable environments inspired by practical embodied AI scenarios. Each environment consists of a partially observable 2D grid map and an unknown task Directed Acyclic Graph (DAG). The map generation can be programmatically adjusted to emphasize exploration or exploitation difficulty. To enable policy-agnostic evaluation, we design a metric to quantify exploration and exploitation errors from agent's actions. We evaluate a variety of frontier LM agents and find that even state-of-the-art models struggle on our task, with different models exhibiting distinct failure modes. We further observe that reasoning models solve the task more effectively and show both exploration and exploitation can be significantly improved through minimal harness engineering. We release our code \href{https://github.com/jjj-madison/measurable-explore-exploit}{here}.
