About

I am a Ph.D. student in the Artificial Intelligence Thrust, Information Hub, at The Hong Kong University of Science and Technology (Guangzhou), advised by Prof. Yao Shu. Previously, I received my master’s and bachelor’s degrees in Physics from Sun Yat-sen University, where I was advised by Prof. Haiping Huang.

My research focuses on developing theoretically grounded and practically efficient algorithms for modern AI systems. I study variance-reduced zeroth-order optimization for black-box objectives and on-device LLM fine-tuning; efficient LLM post-training in online and offline settings; and post-training quantization and edge deployment, balancing performance, efficiency, and privacy.

Selected Publications

Revisiting Zeroth-Order Hessian Approximation: A Single-Step Policy Optimization Lens

Junbin Qiu, Zhaowei Hong, Renzhe Xu, Yao Shu

ICML 2026

A policy-optimization view of zeroth-order curvature estimation, with variance reduction through baseline selection and query reuse.

2025

Zeroth-Order Optimization is Secretly Single-Step Policy Optimization

Junbin Qiu, Zhengpeng Xie, Xiangda Yan, Yongjie Yang, Yao Shu

ICML 2025 Workshop

Connects finite-difference zeroth-order optimization with single-step policy optimization, motivating an averaged baseline and query reuse for variance reduction.

Education

The Hong Kong University of Science and Technology (Guangzhou)

2025 – present

Ph.D. student in Artificial Intelligence

Artificial Intelligence Thrust, Information Hub

Sun Yat-sen University

2021 – 2024

Master’s degree in Statistical Physics

Sun Yat-sen University

2017 – 2021

Bachelor’s degree in Physics

Publications

8 matching publications

2026

APEX: Accuracy Projection from Verifier-Labeled Experience for Offline RLVR

Junbin Qiu, Yao Shu

Submitted to NeurIPS 2026

Revisiting Zeroth-Order Hessian Approximation: A Single-Step Policy Optimization Lens

Junbin Qiu, Zhaowei Hong, Renzhe Xu, Yao Shu

ICML 2026
2025

Zeroth-Order Optimization is Secretly Single-Step Policy Optimization

Junbin Qiu, Zhengpeng Xie, Xiangda Yan, Yongjie Yang, Yao Shu

ICML 2025 Workshop

An optimization-based equilibrium measure describing fixed points of non-equilibrium dynamics: application to the edge of chaos

Junbin Qiu, Haiping Huang

Communications in Theoretical Physics
2024

Meta predictive learning model of languages in neural circuits

Chan Li, Junbin Qiu, Haiping Huang

Physical Review E
2022

Equivalence between belief propagation instability and transition to replica symmetry breaking in perceptron learning systems

Yang Zhao, Junbin Qiu, Mingshan Xie, Haiping Huang

Physical Review Research

Activities

Contact

I welcome research discussions and collaborations around theoretical machine learning, zeroth-order optimization, efficient LLM post-training, and model deployment.