Bio

B.S. in Computer Science

• Graduated summa cum laude (top 2%); Best Senior Award; Galileo Circle Scholar (top 0.8% in College of Science).
• Relevant courses: I have built a webpage for my coursework.

Secondary Education

• Graduated with distinction; awarded for excellence in science, technology, and the arts.

Research Scientist

• Working on agents and RL environments.

Senior Member of Technical Staff, Agents Lead

• Led a 15-person cross-functional team on an MCP-based agent RL project, coordinating environment development, synthetic/real data generation, expert annotation, and agent evaluation.
• Designed B2B and B2C RL environment simulators and deployed Harbor for systematic agent benchmarking and evaluation.
• Built a synthetic task and rubric generation pipeline producing 3,000+ high-quality, long-horizon tasks with automated evaluation criteria across 10 environments (~800 tools).
• Led Dr. DocBench (EMNLP, in submission), overseeing evaluation design, data selection, and the human annotation pipeline.

Visiting Researcher

• Advised by Dr. Juanzi Li and mentored by Zijun Yao.
• Research in mechanism interpretability and LLM reasoning.

Undergraduate Research Assistant

• Advised by Dr. Liangming Pan, Prof. Mihai Surdeanu.
• Explored the physics of language models, focusing on how LLMs reason under distracting context and identifying universal principles for robust reasoning. (First Author, EMNLP 2025 Main)
• Advanced LLM-guided reinforcement learning, improving data efficiency and generalization, collaborating with Dr. Chicheng Zhang. (Second Author, EXAIT Workshop, ICML2025)
• Speculative decoding with copying mechanism, boosting the speed 3.08x. (Second Author, EMNLP 2025 Main)

Undergraduate Research Assistant

• Advised by Dr. Kobus Barnard, Dr. Adarsh Pyarelal.
• Developing dynamic Bayesian networks and Theory of Mind models for the ToMCAT project, focusing on modeling human coordination and distinguishing genuine interpersonal synchrony.

Undergraduate Research Assistant

• Advised by Dr. Reyan Ahmed, Dr. Stephen Kobourov.
• Conducted deep learning research on graph drawing, focusing on evaluating and interpreting the behavior of Graph Neural Networks.

Machine Learning Engineer Intern (Full Time)

• Built GPT-powered bots for Slack using RAG, integrated Slack-based live chat into the official website.
• Orchestrated AWS website deployment, simulated 500+ concurrent users with Selenium, and received the Best Intern Award.

Team Leader, Mathematical Contest in Modeling (MCM)

• Advised by Prof. Patrick Shipman.
• Led a team for the 2024 MCM, responsible for project coordination, modeling, and paper writing.

EchoRL: Reinforcement Learning via Rollout Echoing

• Revives advantage-degenerated prompts in RLVR post-training via rollout echoing, restoring learning signal that otherwise collapses as training proceeds.

AlignSAE: Concept-Aligned Sparse Autoencoders

• Aligns sparse autoencoder features with a predefined ontology via a "pre-train, then post-train" curriculum, binding concepts to dedicated latent slots.
• Enables precise causal interventions ("concept swaps") and supports multi-hop reasoning plus a mechanistic probe of grokking-like generalization dynamics.

How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled Benchmark

• Ran controlled experiments showing irrelevant extra sentences consistently degrade LLM reasoning in a predictable pattern, finding 6 novel results for next generation LLMs pretraining laws.
• Found harder (deeper, multi-step) problems are more easily distracted, with mistakes from both picking wrong paths and arithmetic calculations.
• Developed a “hard distractor” training regimen that noticeably increases robustness, even on out-of-distribution problems.
• Added a reward-guided stepwise Tree-of-Thought that gives an up to 6.29% accuracy boost on tough, out-of-distribution cases.

CopySpec: Accelerating LLMs with Speculative Copy-and-Paste Without Compromising Quality

• Developed a method enabling LLMs to speculatively copy repeated outputs, reducing unnecessary computation.
• Introduced the MT-Redundant dataset to benchmark LLM performance on follow-up turns with repeated content.
• Achieved up to 3.08x speed-up on conversational benchmarks and a 49% additional speed-up over speculative decoding, with no extra memory requirements.

Improving the Data-efficiency of Reinforcement Learning by Warm-starting with LLM

• Improving RL data-efficiency by leveraging LLMs for warm-starting, achieving 20X in data efficiency.

Dr. DocBench: A Comprehensive Benchmark for Expert-Level and Difficult Document Parsing

Opening the Black Box: A Survey on the Mechanisms of Multi-Step Reasoning in Large Language Models

Knowledge Index of Noah's Ark

Triaging Threats to Specialized Guardrails

Word2VecGD: Neural Graph Drawing with Cosine-Stress Optimization

Peeking inside the Black-Box: Reinforcement Learning for Explainable and Accurate Relation Extraction

VEFX-Bench: A Holistic Benchmark for Generic Video Editing and Visual Effects

Justified or Just Convincing? Error Verifiability as a Dimension of LLM Quality

• Best Senior Award, University of Arizona
• Graduated summa cum laude, University of Arizona

• Undergraduate Research Travel Grant (10 awardees across the university), University of Arizona

• Galileo Circle Scholar, College of Science (0.8%), University of Arizona

• Mathematical Contest in Modeling (MCM) Award, Consortium for Mathematics and Its Applications (COMAP)

• Academic Highest Distinction (Dean's List), University of Arizona
• Global Wildcat Scholarship, University of Arizona

• Shanghai Youth Science and Technology Innovation Competition (Third Prize), Shanghai Nanyang Model High School

• China National Youth Arts Competition (Group First Prize), Shanghai Nanyang Model High School

President

• Organized workshops on AI agents and LLM applications, including a Hack Arizona event with 100+ participants.
• Led weekly lectures and invited speaker sessions to foster AI learning across disciplines.
• Organized a reading group that mentors students interested in AI research.
• Raised over $14,000 in sponsorships to support club activities and student research projects.

Instructor, Deep Learning and Natural Language Processing

Instructor, Math for AI Workshop Series

TA, CSC-244: Discrete Mathematics II

TA, CSC-144: Discrete Mathematics I

Reviewer

Reviewer

Reviewer

• Served as a reviewer for the ICML 2024 AI4MATH Workshop, evaluating submissions on reinforcement learning for LLM post-training and decoding mechanisms.