Projects and Papers

Catastrophic Misalignment in Bandit Reward Learning

Authors: David Chen

MS&E 338 • 2025

This project explores the recently developed concept of catastrophic misalignment, in an empirical combinatorial-bandit setting. A line of work explores the idea of catastrophy arising not from malicious intent, but rather from competent execution of tasks, under misalignment to human goals. Prior theoretical work frames this in the context of superintelligent agents, and this work develops experimental and preliminary theoretical results in a simple AI alignment bandit environment. Key results include demonstration that catastrophic performance of Thompson Sampling is associated with greater competence and ability to derive certain outcomes in its environment.

Exploring Bandit Algorithms

Authors: David Chen

CS 221 • 2025

This project evaluates classic and modern stochastic multi-armed bandit algorithms by comparing their theoretical regret bounds against empirical performance and runtime. I developed a simulation framework to test strategies such as ϵ-greedy, Thompson sampling, and information-directed sampling across independent and linear bandit settings. This comprehensive analysis aims to provide both quantitative benchmarks and qualitative insights into how different exploration-exploitation strategies behave in practice.

Evaluating Stitching Capabilities of RvS Transformer Algorithms

Authors: David Chen

CS 234 • 2024

This paper benchmarks the ability of Transformer-based reinforcement learning methods to "stitch" suboptimal trajectories into optimal policies across challenging AntMaze environments. We introduce an enhanced Waypoint Transformer with a refined waypoint selection strategy that improves performance. These contributions provide a comprehensive evaluation of current sequence modeling approaches and suggest new avenues for goal-conditioned behavior cloning.