Ph.D. student, Stanford AI
I am a final-year Ph.D. student at Stanford University, advised by Chelsea Finn and Aviral Kumar, and a Research Scientist Intern at Meta MSL. Prior to Stanford, I was at UC Berkeley, where I was advised by Sergey Levine, and I have held internships at Microsoft Research NYC, Google DeepMind Robotics, and Toyota Research Institute. My research is supported by the NSF Graduate Research Fellowship.
My research develops algorithms for self-improving and creative foundation models, with the long-term goal of enabling progress on difficult, open-ended problems such as scientific research. These problems require more than imitating demonstrations or solving well-specified tasks: models must decide what is worth trying, synthesize partial evidence, discover useful abstractions, and refine ideas whose value is revealed only through downstream validation. I study how feedback from a model's own attempts, natural-language critiques, and inference-time computation can be transformed into learning signals that drive exploration and improvement. The aim is to build models that can generate, evaluate, and refine novel insights, learning problem-solving strategies that generalize across domains.
ICLR 2026 · Spotlight, ES-FoMo Workshop @ ICML 2025
A multi-agent RL method (RLAD) that trains LLMs to discover reusable reasoning abstractions, steering exploration toward useful strategies on hard problems.
Under submission, CoLM 2026
An LM trained with RL to anticipate a paper's core insight from its foundational references, generalizing to unseen scientific disciplines.
ICLR 2025 Workshop on Reasoning and Planning for LLMs · Runner-Up Best Paper
Pairs process verifiers with preemptive backtracking to resample only problematic steps, cutting test-time compute.
CoLM 2025 (Oral)
Shows that intrinsic reasoning behaviors — verification, backtracking, subgoal setting, and backward chaining — are what let LLMs self-improve under reinforcement learning.
2025
Extends chain-of-thought by explicitly modeling the latent reasoning behind each CoT, via process supervision, synthetic data, and search.
Preprint 2024
A generative self-evaluation scheme that lets an LLM predict mid-generation whether restarting would yield a better response, cutting sampling cost.
CoLM 2025
Over 250,000 high-quality, verifiable math problems built for reinforcement learning in language models.
ICLR 2026
FSPO, a meta-learning framework that leverages in-context learning to personalize LLM responses to individual users from only a few preferences.
ICML 2024
Shows that on-policy sampling and negative-gradient methods outperform offline maximum-likelihood objectives for preference fine-tuning.
TMLR 2026
HyRe adapts a multi-head model at test time from about five labeled examples, matching state-of-the-art reward models at the cost of fine-tuning one.
CVPR 2025
Personalizes text-to-image diffusion models from a few examples via user preference embeddings and cross-attention during DPO fine-tuning.
RSS 2023
Combines offline-RL pre-training on diverse robot data with rapid fine-tuning, learning new tasks from as few as 10 demonstrations.
NeurIPS 2023
A calibrated, conservative value initialization that enables fast and stable offline-to-online fine-tuning.
NeurIPS 2023
Reworks the CQL distribution constraint into an approximate support constraint, enabling offline RL on realistic, heteroskedastic data.