Ph.D. student in Artificial Intelligence, Shanghai Jiao Tong University
LLM post-training, work agents, and code agents.
I am a Ph.D. student at Shanghai Jiao Tong University, advised by Prof. Zhipeng Zhang. Previously, I received my master's degree in Computer Science from SJTU, advised by Prof. Li Niu and Prof. Linfeng Zhang, and my bachelor's degree in Instrument Science and Control Technology from Southeast University.
My research interests lie in LLM post-training, work agents, and code agents. My previous work spans collaborative work agents, model merging, and the generation and evaluation of visual content.
I'm always happy to discuss LLM post-training, work agents, and code agents. Let's get in touch.
From diffusion models to LLM agents, I explore how AI can support human creativity and productivity.
Human–agent collaboration for research, writing, and scientific workflows.
Adapting and merging language models for better capability, safety, and utility.
Diffusion models, efficient sampling, and image and video generation.
Benchmarks for models and agents, and surveys of emerging research directions.
Agents that perceive, reason, and act in physical environments.
Publications coming soon.
Scroll within the list to explore more publications ↓
A multi-agent framework for coherent posters, slides, videos, and project pages, with cross-artifact verification and the OmniPreBench evaluation suite.
An exploration potential score identifies less informative training prompts and guides teacher-generated scaffolds for adaptive multimodal RL post-training.
A diagnostic benchmark that separates language understanding from motor execution to test whether embodied agents truly follow instructions.
DLR compresses rendered reasoning traces into interpretable discrete latent tokens, supporting autoregressive training, supervised finetuning, and reinforcement learning.
CTRL-S combines chain-of-thought SVG generation with multi-task, multi-reward reinforcement learning for structured and visually faithful vector graphics.
Structural, causal, and geometric probes of latent reasoning guide training-free interventions that improve reasoning accuracy without parameter updates.
Consistent edits across related images by combining geometric correspondences with video generation priors, evaluated on the GroupEditBench benchmark.
A large-scale suite for video reasoning, pairing diverse training tasks with verifiable evaluation and a systematic study of reasoning generalization.
A transparent, human-in-the-loop multi-agent workflow for responding to peer reviews.
Turning research papers into project pages with human–agent collaboration, paired with PageBench for evaluation.
Selecting compact, high-quality reasoning data by measuring its influence on the attention heads that form a model’s mathematical reasoning circuits.
Decouple diffusion finetuning into denoising tasks, then merge them into a stronger model.
Efficient diffusion modeling for low-light image enhancement.
Mitigating safety–utility conflicts when merging large language models.
A comprehensive benchmark suite for evaluating video generative models.
Accelerating Stable Diffusion with token pruning and feature caching.
Cross-stage prompt optimization for text-to-video generation through data alignment and test-time scaling.
A unified view of conventional, vision-language, and hybrid approaches to end-to-end autonomous driving.
Ph.D. in Artificial Intelligence
School of Artificial Intelligence
Advisor: Prof.
Zhipeng Zhang
M.Sc. in Computer Science
Department of Computer Science and Engineering
Advisors: Prof.
Li Niu and Prof.
Linfeng Zhang
B.Eng. in Instrument Science and Control Technology
School of Instrument Science and Engineering
Machine Learning Engineering Intern · Paddle Team
Conference reviewer
ICML, ICLR, ICCV, CVPR, ECCV, COLM, ARR, NeurIPS,
ACM MM, ACM WWW, AAAI
AI2805 · Digital Signal and Image Processing
Teaching Assistant
· SJTU · Spring 2026
AI7002 · AI Ethics
Teaching Assistant · SJTU · Autumn 2025