M.S

Master of Science

University of California San Diego

Graduate Student Researcher

Rose Spatiotemporal ML (STL) Lab, UC San Diego

Developing Zephyrus-R1, a two-stage agentic RL framework for weather reporting. Extended ZEPHYRUS for visual tool use and built RL training with rubric- and entailment-based rewards.

• Advised by Rose Yu, Yian Ma, and Taylor Berg-Kirkpatrick.

Graduate Research Assistant

Rose Spatiotemporal ML (STL) Lab, UC San Diego

Developed CSR, a semantic-space RL framework that improves LLM calibration through correctness and semantic-agreement rewards.

• Advised by Rose Yu and Yian Ma.

Graduate Student Researcher (GSR)

Dept. of Computer Science and Engineering, UCSD

Summer research support in RoseSTL Lab.

Zephyrus-R1: Agentic Reinforcement Learning for Weather Reporting

arXiv preprint arXiv:2610.25625

Developed a two-stage RL curriculum that transfers verifiable weather-analysis skills to open-ended report generation, with rubric- and claim-entailment-based rewards. Improved the six-score mean across three LLM judges by 26% (4B) and 13% (27B) over the corresponding base models.

Calibrating LLMs with Semantic-level Reward

NeurIPS 2026

This paper introduces Calibration with Semantic Reward (CSR), a semantic-space RL framework that combines correctness and semantic calibration rewards without relying on verbalized confidence. We designed the reward formulation and training pipeline and evaluated CSR across in-distribution and out-of-distribution question-answering benchmarks.