NeurIPS 2026
Dake Bu
卜 大可 ・ ウラナイ タイカPh.D. Candidate, City University of Hong Kong
Research Intern, CFAR, A*STAR
dakebu2-c [at] my.cityu [dot] edu [dot] hk
About
Hi! I am a Ph.D. candidate in Computer Science at City University of Hong Kong, advised by Prof. Hau-San Wong and Prof. Qingfu Zhang. I also work closely with Prof. Wei Huang and Prof. Andi Han.
My research focuses on the theory and algorithms of large-scale machine learning systems, spanning emergent in-context learning, agentic AI, LLM and diffusion post-training, reinforcement learning, non-log-concave sampling, AI4Math, and quantum computing. I am particularly interested in theory-grounded algorithm design together with reproducible experiments.
I am currently a one-year research intern at CFAR, A*STAR, supervised by Prof. Atsushi Nitanda. Before that, I spent 2024–2025 with the Deep Learning Theory Team at RIKEN AIP, working under Prof. Taiji Suzuki.
I received my B.S. in Mathematics from Xi'an Jiaotong University in 2023, advised by Prof. Hui Li. Earlier, from January to May 2023, I was a research assistant in the LOGO Lab at The Chinese University of Hong Kong, Shenzhen, advised by Prof. Tianshu Yu. In 2026, I also visited the School of Mathematics and Statistics at the University of Sydney, working with Prof. Andi Han.
Affiliations
Recent News
- Serving as Web Chair for the Conference on Parsimony and Learning (CPAL 2027), to be held in Tokyo, March 23–26, 2027.
- Our work An Automated Theorem Proving System and Visualized Lean Library for Sampling Theory, Optimisation and Geometry, together with the public Lean library Samplinglib (GitHub), was accepted to the NeurIPS 2026 Workshop MATH-AI.
- Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation (Code) was accepted to NeurIPS 2026.
- Our GitHub repositories ABRL and the automated theorem-proving system for sampling, optimisation, and geometry are available, together with the public Lean libraries BanditRLlib and Samplinglib.
- Attended the Annual Summer School on Mathematical Aspects of Data Science at NUS.
- Released ASPBE: Automatic State Preparation and Block Encoding for Quantum Computing, with the public Lean library QuantumComputinglib.
- DPRM was selected as an Oral, and Distributional Biases in Post-Training as a Spotlight, at ICML 2026 FoGen.
- Visited the School of Mathematics and Statistics at the University of Sydney and gave a talk on Langevin dynamics with partial structure.
- Received the ICML 2026 Gold Reviewer Award.
- Provable Sample Efficiency of Curriculum Post-Training for Transformer Reasoning was accepted to ICML 2026.
- Slowly Annealed Langevin Dynamics and DPRM became available online.
- Two papers on in-context vector arithmetic and lexicographic multi-objective reinforcement learning were accepted to ICML 2025.
- Provably Transformers Harness Multi-Concept Word Semantics for Efficient In-Context Learning was accepted to NeurIPS 2024.
- Provably Neural Active Learning Succeeds via Prioritizing Perplexing Samples was accepted to ICML 2024.
Scroll for earlier news.
Selected Research
Visitors
ICML 2026
ICML 2025
ICML 2025
NeurIPS 2024