Thomas Bush

Hi, I’m Tom! My goal is to ensure advanced AI lives up to its potential to benefit humanity. My current research revolves around the unifying theme of safe and reliable AI. Specifically, I am interested in answering the following two questions:

Under what conditions do models learn to reason in a safe and reliable fashion?
How can we verify whether a model is reasoning in a safe and reliable manner?

To this end, my primary research interests are (i) mechanistic interpretability, (ii) reinforcement learning and (iii) world models. I am especially excited about research at the intersection of these topics.

I am currently a MATS research scholar supervised by Adria Garriga-Alonso. Until recently, I was also research assistant at Krueger AI Safety Lab supervised by Prof David Krueger, and working with Usman Anwar and Stephen Chung. Previously, I was a MPhil student in Machine Learning and Machine Intelligence at the University of Cambridge, and a BSc student in Philosophy, Politics and Economics at LSE. I am currently actively looking for research opportunities and PhD positions so please reach out!

selected publications

Interpreting Emergent Planning in Model-Free Reinforcement Learning

Thomas Bush, Stephen Chung, Usman Anwar, and 2 more authors

2024

under review

Abs Code Website

We present the first mechanistic evidence that model-free reinforcement learning agents can learn to plan. This is achieved by applying a methodology based on concept-based interpretability to a model-free agent in Sokoban – a commonly used benchmark for studying planning. Specifically, we demonstrate that DRC, a generic model-free agent introduced by Guez et al. (2019), uses learned concept representations to internally formulate plans that both predict the long-term effects of actions on the environment and influence action selection. Our methodology involves: (1) probing for planning-relevant concepts, (2) investigating plan formation within the agent’s representations, and (3) verifying that discovered plans (in agent’s representations) have causal effect on agent’s behavior through interventions. We also show that the emergence of these plans coincides with the emergence of a planning-like property: the ability to benefit from additional test-time compute. Finally, we perform a qualitative analysis of the planning algorithm learned by the agent and discover a strong resemblance to parallelized bidirectional search. Our findings advance understanding of the internal mechanisms underlying planning behavior in agents, enabling improved diagnosis, interpretation, and control of agent planning processes.