PhD Candidate in Computer Science · University of Utah
I am a PhD candidate at the University of Utah, advised by Prof. Daniel S. Brown. I expect to graduate in May 2027. Previously, I earned an M.S. in Electrical and Computer Engineering from Carnegie Mellon University, where I worked with Prof. Carlee Joe-Wong.
I am currently seeking postdoctoral research opportunities starting in Summer/Fall 2027 at the intersection of Human–AI alignment, Reinforcement Learning from Human Feedback(RLHF), Imitation learning and robot learning, with particular interests in Pluralistic Alignment, AI safety and Robustness, and Interpretability.
My research aims to develop AI agents that can reliably learn and adapt to human objectives from preferences and demonstrations. While learning from human feedback offers a promising alternative to manually specifying objectives and behaviors, it does not guarantee alignment with human intent. An agent may optimize a misaligned reward without exhibiting obvious failures, a policy learned from demonstrations may be vulnerable to adversarial manipulation, and feedback aggregated across users may obscure legitimate differences in individual objectives.
My PhD tackles these challenges along three complementary directions:
My long-term vision is to bridge the gap between learning from human feedback and reliably acting on human intent, enabling intelligent AI systems and robots that serve human objectives while simultaneously adapting to the complexity, uncertainty, and diversity of the real world.
Earlier version: “P3L: Learning Pluralistically Aligned Policies from Heterogeneous Crowd Feedback”, Reinforcement Learning Beyond Rewards (RLBrew) Workshop, RLC 2026.
Under Submission
We learn structured representations of heterogeneous preferences to support personalized reward functions and policies, including few-query adaptation when annotator identities are unavailable during training.
IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2026
TL;DR: We present the first systematic study of adversarial attacks, across a range of both classic and recently proposed imitation learning algorithms, including Vanilla Behavior Cloning (Vanilla BC), LSTM-GMM, Implicit Behavior Cloning (IBC), Diffusion Policy (DP), and Vector-Quantized Behavior Transformer (VQ-BET) to white-box, grey-box, and black-box transfer adversarial perturbation attacks.
* Equal contribution.
EAI SmartSP 2025 · Springer LNICST, vol. 699 (proceedings dated 2027)
TL;DR: We conduct the first study and discussion of clean-label backdoor poisoning o in imitation learning and introduce a novel entropy-based test-time trigger attack that substantially degrades policy performance by identifying critical states where test-time triggering of the backdoor is expected to be most effective at degrading policy performance.
Reinforcement Learning Journal (RLJ) / Reinforcement Learning Conference (RLC), 2024
TL;DR: We introduce a novel reward learning framework to learn an interpretable reward function from human preferences without requiring any hand-crafting of the input feature space, demonstrating how interpretable reward functions can help identify silent misalignment that in-distribution downstream RL policy performance alone may conceal.
Earlier version: “Interpretable Reward Learning via Differentiable Decision Trees,” ML Safety Workshop, NeurIPS 2022 (Oral).
IEEE/ACM Transactions on Networking, 2022
TL;DR: We quantify the performance trade-offs between error, cost, and training time for volatile cloud instances and develop cost-aware training strategies for ML models using preemptible cloud instances.