Baiyu Su

The University of Texas at Austin · Computer Science

Email LinkedIn CV PDF

Research Interests

Empirical AI safety for generative and reasoning models, especially signed-measure methods for controllable sampling, rejection, and robust post-training.

Education

The University of Texas at Austin

Ph.D. Candidate, Computer Science · Austin, TX

University of Cambridge

M.Eng. in Computer Engineering with Distinction · Cambridge, UK

University of Cambridge

B.A. in Computer Engineering, First Class Honours · Cambridge, UK

Experience

Machine Learning Modelling Engineer Intern

Optiver · Shanghai, China

Student Researcher

Google · Mountain View, CA

Investigated expert-solution injection into on-policy training of LLM reasoning systems for mathematical reasoning.

Graduate Research Assistant

The University of Texas at Austin · Austin, TX

Developed signed-measure sampling methods and studied optimization theory for modern deep learning.

Publications

  1. Su, B., Chen, L., Liu, Q. The Curious Case of AdamW NeurIPS 2026, under review
  2. Su, B.*, Liao, R.*, Chen, L., Liu, Q. Signed Rectified Flow NeurIPS 2026, under review; co-first author
  3. Wang, Q., Chen, L., Liao, R., Su, B., Yang, S., Liu, Q. Neural Numerical Solvers NeurIPS 2026, under review
  4. Liao, R., Yu, J., Su, B., Zhang, C., Chen, L., Liu, Q. Momentum Guidance: Plug-and-Play Guidance for Flow Models ECCV 2026, under review
  5. Peng, B., Chen, L., Su, B., Quesnelle, J., Kingma, D. P., Liu, Q. DeMo: Decoupled Momentum Optimization ICLR 2026
  6. Chen, L., Li, J., Liang, K., Su, B., Xie, C., Pierse, N. W., Liang, C., Lao, N., Liu, Q. Cautious Weight Decay ICLR 2026
  7. Su, B., Liu, Q. Quadratic Quantum Variational Monte Carlo NeurIPS 2024

* Equal contribution.

Awards

Technical Skills

Languages Python, C++

Tools Git, Docker, Slurm

ML / Systems PyTorch, JAX, vLLM, Transformers, Diffusers, Megatron-LM, VeRL, NumPy