Junyi Wu

Junyi Wu

PhD Student · Mathematics · Purdue University

Junyi Wu

About

Hi there! I'm a PhD student at Purdue University , Department of Mathematics, working in Prof. Guang Lin's research group. My research focuses on post-training of large generative models, including LLMs and diffusion models, through principled reinforcement learning and preference alignment. I received my B.S. in Mathematics and Applied Mathematics from the University of Chinese Academy of Sciences .

Research Interests

  • LLM Post-Training

    Reinforcement Learning, RLHF, Preference Optimization, and Reasoning.

  • Generative Model Alignment

    KL-Regularized Reward Fine-Tuning, One-Step Diffusion Generators, and VLA Policies.

Open to collaborations! Feel free to reach out if our research interests align.

News

  1. Oct 2026

    Two papers were accepted to NeurIPS 2026.

  2. Jun 2026

    Junyi serves as a co-mentor for Purdue SURF 2026.

  3. May 2026

    Our paper on one-step generator RL (DIDR) was released, with models and a demo on Hugging Face.

  4. 2026

    Junyi serves as reviewer for NeurIPS 2026.

Publications * equal contribution

4 papers

DIDR one-step text-to-image samples
NeurIPS 2026

Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL

Junyi Wu, Weijian Luo, Haoyang Zheng, Ruizhe Zhang, Guang Lin

A data-free, trajectory-level RLHF objective for one-step text-to-image generators that shares the minimizer of clean-image KL-regularized RLHF.

Paper Code Model Demo
Safe quantum bandit models
NeurIPS 2026

Quantum Safe Stochastic Linear Bandits

Ruizhe Zhang, Junyi Wu, Guang Lin

Constrained Thompson sampling with high-probability action feasibility and polylogarithmic regret under coherent quantum feedback.

COCOA action correction overview
Preprint 2026

COCOA: Action Correction for VLA Policies via Supervised Fine-Tuning and Preference Optimization

Yi Wu, Manav Kulshrestha, Junyi Wu, Kai Cheng, Nan Jiang, Aniket Bera, Lin Tan

An action-conditioned correction policy that repairs a frozen VLA's proposed actions, trained with SFT followed by Flow-DPO-style preference optimization.

PO-CKAN framework overview
Preprint 2025

PO-CKAN: Physics Informed Deep Operator Kolmogorov Arnold Networks with Chunk Rational Structure

Junyi Wu, Guang Lin

A physics-informed DeepONet built from chunk-wise rational KAN layers, enabling scalable KAN-based operator learning with consistent gains over PI-DeepONet.

Paper Code

GitHub

Loading repositories…

Education

Ph.D. in Mathematics, Purdue University

Advisor: Prof. Guang Lin

B.S. in Mathematics and Applied Mathematics, University of Chinese Academy of Sciences

Academic Services

Conference Reviewer
NeurIPS 2026
Mentoring
Purdue SURF 2026