Yue Wu · 武岳
I am a Ph.D. student in Data Science and Artificial Intelligence at The Hong Kong Polytechnic University, advised by Prof. Jian Huang. I received my B.S. in Statistics from Southern University of Science and Technology (SUSTech), advised by Prof. Bingyi Jing.
My research focuses on LLM post-training, efficient inference, and LLM agent systems. I work on training objectives for supervised fine-tuning and on-policy distillation, decoding for diffusion language models, and agent harnesses and evaluation for data science. My work includes first-author papers accepted to EMNLP 2026, alongside industry experience at SenseTime Research Institute and xDAN AI.
Research Interests
- LLM post-training: adaptive supervised fine-tuning, on-policy distillation, and model self-improvement.
- Efficient inference: diffusion language model decoding, remasking, and stability-aware computation.
- LLM agent systems: data science agents, distribution-aware tool retrieval, and real-world evaluation.
News
- 2026: CoDR and DSAEval accepted to EMNLP 2026.
- 2025: Began my Ph.D. at The Hong Kong Polytechnic University after receiving my B.S. in Statistics from SUSTech.
- May 2025: Completed my internship at SenseTime Research Institute.
Selected Research
† Equal contribution (co-first authors).
CoDR: Training-Free Confidence-Drift Remasking for Diffusion Language Models
EMNLP 2026. Accepted.
Training-free, sampler-agnostic remasking guided by confidence drift improves accuracy across two diffusion LLMs, three samplers, and four benchmarks, with 76% fewer forward evaluations than ReMDM in a matched setting.
DSAEval: Evaluating Data Science Agents on a Wide Range of Real-World Data Science Problems
EMNLP 2026. Accepted.
A real-world data science agent benchmark with 641 tasks across 285 datasets, incorporating multimodal perception, multi-query interactions, and multidimensional evaluation of 13 advanced LLM agents.
Bootstrapped On-Policy Distillation (BOPD)
Under review at ICLR 2027.
A two-stage framework bootstraps a task-adapted teacher through reference-conditioned self-distillation, then trains an independent student via on-policy distillation. On Qwen3-4B, mean mathematical reasoning accuracy improves from 18.06% to 23.98%, surpassing the constructed teacher.
All publications & manuscripts → · Research experience →
Get in Touch
I welcome conversations about LLM training, efficient inference, and agent systems. You can reach me at yue0301.wu@connect.polyu.hk.