Ph.D. Student · Xidian University
I am a final year Ph.D. student in Information and Telecommunication Engineering at Xidian University, advised by Prof. Bo Chen. I received my bachelor's degree in Electronic Engineering from Xidian University in 2022.
My research focuses on finding and leveraging sparse, dynamic, yet highly informative learning signals in foundation models—from efficient attention and multimodal concept learning to reward signals and credit assignment in post-training. I am currently interested in agentic post-training and Recursive Self-Improvement (RSI).
I also worked as an LLM Research Intern at BandAI, ByteDance, and as a Project Up Intern at the Yuanbao AI-Infra team, Tencent.
† Equal contribution. Highlighted cards are first-author papers. See Google Scholar for the latest list.
Reveals that later rollout tokens provide lower-quality teacher supervision, then derives importance weighting that emphasizes earlier, more reliable tokens for faster and stronger on-policy distillation.
Introduces Sparse Feature Attention and an IO-aware FlashSFA kernel, using sparse query/key features to cut attention cost while preserving dense-model quality on long-context tasks.
Models local image patches and concepts as an optimal-transport problem, producing explicit fine-grained alignments and more reliable concept predictions.
Uses textual memory and retrieve-then-filter concept selection to generate grounded zero-shot captions with fewer hallucinations and stronger generalization.
LLM Research Intern
Built evaluation and data-synthesis pipelines for e-commerce agents, deployed Agentic OPD for post-training, and studied stale-sample mitigation in fully asynchronous agentic reinforcement learning.
Project Up Intern
Developed model-sparsification methods for inference efficiency and deployed Sparse Feature Attention in production workflows.
B.Eng. · Electronic Engineering
Official implementation for correcting position bias in on-policy distillation.
View on GitHub → Efficient AttentionTraining code and kernels for scaling attention through feature-level sparsity.
View on GitHub → Vision-LanguageMemory-augmented zero-shot image captioning with grounded concept retrieval.
View on GitHub →