Machine Learning Lab

Large Language Models (LLMs) have enabled robots to autonomously generate and refine action plans and control programs. However, existing approaches often suffer from unreliable knowledge accumulated through experience, costly trial-and-error, and excessive LLM inference overhead. We aim to develop robotic planning agents that continuously improve from their own experiences while minimizing computational costs. Our research focuses on (1) learning from successes and failures while correcting erroneous knowledge, (2) improving token efficiency in failure analysis and plan code generation, and (3) minimizing runtime latency to enable fast and responsive execution when deployed on real-world robots. Through these efforts, we seek to build robust and efficient robotic agents that autonomously evolve their planning capabilities across diverse tasks and environments. Finding source or binary code vulnerabilities is a long-standing and never-ending problem. The recent advance in LLMs potentially provides clues to further advance the current code vulnerability discovery performance. In this project, we will explore the potentials of LLMs in finding code vulnerabilities. Large language model (LLM) agents, e.g., Claude code and Codex, now resolve real-world software engineering (SWE) issues, but at a high cost: multi-agent SWE systems repeatedly re-explore the same repository, spending a large amount of tokens, latency, and energy on redundant search. In this work, we study how to make SWE agents efficient without sacrificing their issue-solving ability, e.g., by sharing retrieved context across agents and turns, designing persistent sub-agents, and rethinking the agent-computer interface. Reinforcement Learning from Human Feedback (RLHF) allows models to generate outputs aligned