AI RESEARCH CV
Kangqiao Liu
Machine Learning Researcher & Theoretical Physicist
kqliu@xhu.edu.cn · GitHub · Google Scholar · Full academic CV · Download PDF
Research profile
I study stochastic learning dynamics and theoretical models of complex systems. My earlier machine-learning work developed finite-learning-rate theories of SGD noise, minibatch fluctuations, and escape dynamics; my recent work has focused on nonequilibrium response, quantum information, and mathematically controlled dynamical problems. I am now bringing these lines back together around reasoning dynamics, scientific AI, and verifiable autonomous discovery.
Research focus
- Stochastic optimization and training dynamics: finite-learning-rate SGD, minibatch noise, escape and rare-event dynamics.
- Reasoning and scientific AI: dynamics of search and reasoning, verifiable scientific agents, research-grade evaluation environments.
- Theoretical modeling: stochastic processes, nonequilibrium response, information-theoretic bounds, quantum and complex dynamics.
Selected machine learning research
Noise and Fluctuation of Finite Learning Rate Stochastic Gradient Descent — ICML 2021
Kangqiao Liu*, Liu Ziyin*, and Masahito Ueda (*equal contribution)
Developed a discrete-time theory of SGD at finite learning rate, deriving analytical noise and parameter-fluctuation formulas and stability boundaries beyond continuous-time Langevin approximations.
[paper] [arXiv]
Strength of Minibatch Noise in SGD — ICLR 2022 Spotlight
Liu Ziyin*, Kangqiao Liu*, Takashi Mori, and Masahito Ueda (*equal contribution)
Analyzed minibatch noise in discrete-time SGD and derived how noise strength and model fluctuations depend on learning rate, batch size, width, and regularization.
[paper] [arXiv]
Power-law Escape Rate of SGD — ICML 2022 Spotlight
Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda
Derived the stationary distribution around local minima and identified power-law escape dynamics for minibatch SGD, linking optimization escape to the loss-dependent structure of stochastic noise.
[paper] [arXiv]
Selected independent research
Dynamical activity universally bounds precision of response in Markovian nonequilibrium systems — Communications Physics (2025)
Kangqiao Liu and Jie Gu
Established a universal kinetic bound connecting static response precision to dynamical activity in nonequilibrium Markov processes.
[journal] [arXiv]
Classical codes violate the conjectured square-root bound for quantum random access codes — 2026 preprint
Kangqiao Liu (sole author)
Constructed a family of counterexamples and characterized the asymptotic region between the conjectured square-root curve and Nayak’s entropy bound.
[arXiv]
Maximal-velocity deficit under a finite-support constraint in a hard-wall half-line continuous-time quantum walk — Physical Review A (2026)
Kangqiao Liu and Deyou Chen
Reduced an optimal transport problem to a principal-eigenvalue problem and derived the exact asymptotic velocity deficit under finite-support preparation.
[journal] [arXiv]
Current direction and research tooling
- Developing a research program that connects stochastic-process methods with modern reasoning and agent systems: trajectory ensembles, search dynamics, verification, failure recovery, and compute allocation.
- Exploring verifiable scientific-agent environments built from genuine mathematical and physical research tasks, with emphasis on mechanisms and evaluation signals that remain meaningful beyond a single model generation.
- Built Scientific Manuscript Audit, an open-source research workflow for Codex and Claude Code with claim-to-evidence tracing, synthetic evaluation cases, automated validation, and reproducible distribution packages. [GitHub]
Experience
2023–present — Lecturer of Physics, School of Science, Xihua University, Chengdu, China
Independent research in nonequilibrium physics, quantum information, stochastic dynamics, and machine learning.
2020–2023 — Ph.D. in Physics, The University of Tokyo, Japan
Advisor: Prof. Masahito Ueda. Thesis: Theoretical Study on Information Engines for Quantum Transport.
2018–2020 — M.Sc. in Physics, The University of Tokyo, Japan
Advisor: Prof. Masahito Ueda. Thesis: Thermodynamic Uncertainty Relations in Markovian Processes.
Selected funding
- National Natural Science Foundation of China, Young Scientists Fund (Category C), Principal Investigator, CNY 300,000, 2027–2029.
- Xihua University Scientific Research Start-up Foundation, Principal Investigator, 2024–2026.
- JSPS DC2 Grant-in-Aid for Research Fellow, Principal Investigator, 2023.
Technical and research skills
- Programming / research computing: Python, PyTorch, NumPy/SciPy, Jupyter, Git, Linux, reproducible numerical workflows.
- Research methods: stochastic-process modeling, optimization theory, analytical derivations, numerical experiments, benchmarking and evaluation design.
- Scientific communication: LaTeX, reproducible research packages, technical writing and peer review across machine learning and physics.
Academic service
Reviewer for NeurIPS, ICLR, ICML; Physical Review Letters, Physical Review Research, Physical Review E, Communications Physics.