I am an Applied Scientist II at Amazon Store Foundational AI, where I build foundation models and large language models for shopping search, retrieval, and recommendation. I received my Ph.D. from the Graduate Group in Applied Mathematics (GGAM) at UC Davis in March 2022.

My research centers on reinforcement learning and policy optimization for large language models, and on stochastic and zeroth-order optimization.

You can find me on Google Scholar, OpenReview, and LinkedIn.

Research Impact

My research spans the post-training of large language models with reinforcement learning and the optimization theory that underlies modern machine learning.

Large language model post-training and reinforcement learning (RLVR). My recent work develops post-training methods for large language models, spanning supervised fine-tuning and reinforcement learning with verifiable rewards (RLVR). Representative contributions include SFT Doesn’t Always Hurt General Capabilities (ICLR 2025), which characterizes when domain-specific fine-tuning preserves versus erodes a model’s general ability; ST-PPO: Stabilized Off-Policy Proximal Policy Optimization for Multi-Turn Agents Training (COLM 2026), for training multi-turn LLM agents; Reward-Wise Value Estimation for Multi-Reward Optimization in LLMs (ICML 2026); and Reinforcement Learning in Inference Time: A Perspective from Successive Policy Iterations (NeurIPS 2025). In applied settings, these methods support small-language-model pretraining and post-training — on-policy distillation and RLVR — for large-scale shopping generative tasks such as query rewriting and ranking.

My first-authored paper Zeroth-Order Algorithms for Nonconvex-Strongly-Concave Minimax Problems with Improved Complexities (Journal of Global Optimization, 2022; earlier NeurIPS 2020 Workshop and arXiv versions) develops derivative-free (zeroth-order) algorithms with improved iteration and query complexity for nonconvex-strongly-concave minimax problems, where gradients are unavailable and only function evaluations can be queried. Cited across its versions by work at NeurIPS, ICML, ICLR, JMLR, AISTATS, TMLR, and SIAM Journal on Optimization, it has contributed to two further lines of research.

Min-max optimization theory. The paper’s complexity analysis and single-loop zeroth-order scheme are a building block for nonconvex saddle-point theory, extended by later work on single-loop algorithms without strong concavity (AISTATS 2022), duality-gap and Polyak–Łojasiewicz rates (JMLR, 2023), Nash-equilibrium search and derivative-free projection methods (SIAM Journal on Optimization 2021, 2024), and robust zeroth-order minimax methods (NeurIPS 2024).

Zeroth-order LLM training. The paper’s core idea — estimating descent directions from forward evaluations alone, without backpropagation — now underpins memory-efficient LLM fine-tuning, and it is cited by MeZO (NeurIPS 2023), DPZero (ICML 2023), Addax (ICLR 2024), HiZOO (ICLR 2025), MUZO (EMNLP 2025), and PaZO (NeurIPS).

Representative citing work across its NeurIPS 2020 Workshop, arXiv, and journal versions (NeurIPS, ICML, ICLR, JMLR, AISTATS, TMLR, SIAM Journal on Optimization, EMNLP, and IJCAI):

Experience

At Amazon Store Foundational AI, I have worked on:

• 2022–2023 — RoBERTa-based behavior foundation models and graph models for search/ads retrieval and recommendation.
• 2024 — Instruction-following LLM-based embedding models for retrieval and recommendation.
• 2024 — Generative recommender for customer representation (Synergen).
• 2025 — Small language model pretraining, midtraining, and post-training (on-policy distillation, RLVR) for shopping generative tasks such as query rewrite, final-stage ranking, and shopping mission generation.
• 2025 — Foundation model post-training on user behavior data.
• 2026 — Personalized search agent for user-context retrieval, based on SQL and semantic search.

Publications

You can also find my articles on my Google Scholar profile.

Large Language Models and Reinforcement Learning

Recommendation Systems

Optimization