Ph.D. Student in Computer Science
I train and evaluate large language models: post-training for reasoning, robust benchmarks, and foundation models for coding agents.
University of Waterloo · Vector Institute · TIGER-Lab · Waterloo, Canada
I am a Ph.D. student in Computer Science at the University of Waterloo, advised by Prof. Wenhu Chen in TIGER-Lab, and I am affiliated with the Vector Institute. I am currently a Student Researcher at Google Cloud AI Research, working on automated research methods. Before Waterloo, I was an Applied Scientist at Tencent Smart Healthcare (Apr 2021 – Aug 2023), and I have been working with TIGER-Lab since 2023. I received my M.S. in Computer Science from UC Davis and my B.Eng. in Computer Science from Chu Kochen Honors College, Zhejiang University. From June to December 2025 I was a Research Intern at MiniMax, working on pre-training and mid-training for the MiniMax-M2 series.
My research asks how to make LLMs both more capable and more measurable. I work on post-training for reasoning (Critique Fine-Tuning teaches models to critique rather than imitate, and works even with a single training problem), evaluation (MMLU-Pro has become a standard benchmark for frontier models; ClawBench tests agents on live websites), and foundation models for coding agents (function-aware fill-in-the-middle mid-training).
Started as a Student Researcher at Google Cloud AI Research, working on automated research methods.
ClawBench is accepted to Findings of EMNLP 2026.
Released Function-Aware FIM Mid-Training for coding-agent foundation models, with code, data and models.
RewardHarness is accepted to COLM 2026.
The MiniMax-M2 technical report is out; I contributed as a research intern at MiniMax.
Released ClawBench: 153 everyday online tasks on 144 live websites for AI agents.
Critique-Coder is accepted to ICLR 2026.
MMMU-Pro is featured in Google's Gemini 3 launch (Gemini 3 Pro scores 81%).
SuperGPQA is accepted to NeurIPS 2025 (Datasets & Benchmarks).
One-Shot CFT is accepted to EMNLP 2025 (Main).
Critique Fine-Tuning and ScholarCopilot are accepted to COLM 2025.
MMMU-Pro and MAmmoTH-VL are accepted to ACL 2025 (Main).
MEGA-Bench is accepted to ICLR 2025.
MMLU-Pro is accepted to NeurIPS 2024 (Datasets & Benchmarks) as a Spotlight, and LLM-AMT is accepted to Findings of EMNLP 2024.
Released MMLU-Pro; it was adopted as a core benchmark of the Hugging Face Open LLM Leaderboard v2.
Affiliations & experience
15 papers, 8 marked Selected. * equal contribution · † core contributors · Full list on Google Scholar.
Proposes function-aware fill-in-the-middle mid-training on 2.6B tokens of code before agentic post-training, improving SWE-Bench-Verified by +2.8 to +3.2 points on Qwen2.5-Coder-7B/14B and Qwen3-8B.
Introduces MiniMax-M2, open-weight 229.9B-total/9.8B-active MoE models trained with agent-driven data pipelines and the Forge RL system, competitive with closed frontier models on agentic coding and search benchmarks.
A training-free agentic reward model for image editing that evolves a skill-and-tool library from 100 preference demonstrations, beating GPT-5 by 5.3 points on EditReward-Bench and GenAI-Bench.
Evaluates agents on 153 everyday online tasks across 144 live websites, blocking final submissions for safety; the best of eight frontier models (Claude Sonnet 4.6) completes only 33.3%.
Proposes Critique Reinforcement Learning (CRL), rewarding models for correctly judging question–solution pairs; replacing 20% of RL data with CRL pushes Critique-Coder-8B past 60% on LiveCodeBench v5.
Fine-tuning LLMs on teacher critiques of diverse solutions to a single problem improves Qwen-Math-7B by 15% on six math benchmarks, matching RL at 20x less compute.
Fine-tunes Qwen2.5-7B on 500K arXiv papers to emit a [RET] token that retrieves citations mid-generation, reaching 40.1% top-1 citation retrieval accuracy versus 15.0% for E5-Mistral-7B.
Introduces SuperGPQA, a 26,529-question benchmark spanning 285 graduate disciplines built via human-LLM collaborative filtering; the strongest model tested, DeepSeek-R1, reaches 61.82% accuracy.
Trains LLMs to critique noisy responses rather than imitate correct ones, outperforming SFT by 4-10% on six math benchmarks after one hour of training on 50K examples.
Builds a 12M-example multimodal instruction-tuning dataset with chain-of-thought rationales using only open models; training MAmmoTH-VL-8B on it gains +8.1% on MathVerse and +7% on MMMU-Pro.
Builds a benchmark of 505 real-world multimodal tasks (8,000+ samples) with diverse output formats and 40+ metrics, yielding fine-grained capability reports for frontier vision-language models.
Builds a harder version of MMMU by filtering text-only-solvable questions, expanding to ten options and adding vision-only screenshot inputs; model accuracy drops 16.8–26.9 points versus MMMU.
Extends MMLU to 12K harder, reasoning-focused questions with ten options each; model accuracy drops 16-33% and prompt sensitivity falls from 4-5% to 2%.
Fully open bilingual 7B LLM trained from scratch on 4.5T tokens, releasing the Matrix corpus, data pipeline, intermediate checkpoints, and training code, with performance comparable to open-weight peers.
Augments black-box LLMs with retrieval from medical textbooks via a query augmenter, hybrid retriever and knowledge self-refiner, improving accuracy 11.6-16.6% on three medical QA benchmarks.
Advisor: Prof. Wenhu Chen · TIGER-Lab · Vector Institute
