Projects
Ongoing
RL for LLMs Risk Assessment Generative Models Safety
- Tech Lead, developing a three-stage optimization framework that reframes generative model jailbreaking as a sequential decision problem. The system structurally decomposes environmental and algorithmic components to isolate the core drivers of adversarial success, scales exploitation pipelines across multi-modal action spaces (text, vision, and audio), and embeds these insights into a unified two-player zero-sum game.
Learning to Be Cautious for Industrial Chemical Control Critical Applications Safety
- Generalized the "learning to be cautious" framework to continuous action spaces, extending its mathematical applicability to high-dimensional, safety-critical settings. Validated on complex chemical process control benchmarks, bridging the gap between theoretical safe RL and real-world industrial deployments.
Plasticity Loss in Monitored MDPs Continual RL
- Investigating structural mechanisms of plasticity loss within Mon-MDPs to preserve agent adaptability during continual learning. Analyzing how the dual-learning architecture compounds long-term performance degradation, and formulating a systematic framework to evaluate intrinsic vulnerabilities while adapting regularization and reset mitigation strategies.
Previous
Learning to Be Cautious Robust Optimization
Agents trained with standard RL can behave catastrophically in novel or out-of-distribution states. This project introduces a minimax regret framework that makes agents learn to be cautious, preferring conservative actions when epistemic uncertainty is high, without human supervision or prior knowledge of the failure modes. Published in TMLR 2025.
code
Agents trained with standard RL can behave catastrophically in novel or out-of-distribution states. This project introduces a minimax regret framework that makes agents learn to be cautious, preferring conservative actions when epistemic uncertainty is high, without human supervision or prior knowledge of the failure modes. Published in TMLR 2025.
code
Wheelchair Robot Controlled by Brain Signal
A wheelchair robot that helps disabled people move using EEG signals from the brain.
code
A wheelchair robot that helps disabled people move using EEG signals from the brain.
code
NeurIPS 2018 Challenge: RL for Prosthetics
RL Prosthetics controller that learns to walk like humans, accelerated with Imitation Learning.
code
RL Prosthetics controller that learns to walk like humans, accelerated with Imitation Learning.
code