中文

Jiangcheng Song (宋蒋成)

I am an incoming Ph.D. student (starting Fall 2026) at the College of artificial intelligence, Xi'an Jiaotong University, advised by Prof. Nanning Zheng.

My research interests include Knowledge Distillation, Large Language Models, Efficient Inference, and Dataset Condensation.

Email  /  Google Scholar  /  GitHub

profile photo

Publications

I am interested in efficient deep learning, knowledge distillation, and large language model optimization. Representative papers are highlighted.

DR-DPO: Dual-Regularized DPO for Efficient Dataset Condensation
Haiduo Huang*, Jiangcheng Song*, Yadong Zhang*, Guansu Wang, Pengju Ren
* Three authors contributed equally
CVPR, 2026 Finding

A dual-regularized framework that recasts dataset condensation as preference optimization. An inter-class DPO loss maximizes class separability while an intra-class Jensen-Shannon term preserves within-class statistics.

SelecTKD: Selective Token-Weighted Knowledge Distillation for LLMs
Haiduo Huang*, Jiangcheng Song*, Yadong Zhang*, Pengju Ren
* Three authors contributed equally
CVPR, 2026
arXiv

A selective token-weighted knowledge distillation method for large language models that adaptively weights tokens during distillation for improved efficiency and performance.

FastEagle: Cascaded Drafting for Accelerating Speculative Decoding
Haiduo Huang*, Jiangcheng Song*, Wenzhe Zhao, Pengju Ren
* Two authors contributed equally
ICASSP, 2026
arXiv

A cascaded drafting approach for accelerating speculative decoding in large language models, enabling faster inference without sacrificing quality.

PPDD: A Unified Push-Pull Adversarial Objective in Feature and Logit Spaces for Dataset Distillation
Haiduo Huang*, Yadong Zhang*, Jiangcheng Song*, Wenzhe Zhao, Pengju Ren
* Three authors contributed equally
ICASSP, 2026

A unified push-pull adversarial objective that operates in both feature and logit spaces for effective dataset distillation.

DeepKD: A Deeply Decoupled and Denoised Knowledge Distillation Trainer
Haiduo Huang*, Jiangcheng Song*, Yadong Zhang*, Pengju Ren
* Three authors contributed equally
NeurIPS, 2025
arXiv / code

A knowledge distillation framework that deeply decouples and denoises the distillation process for improved training stability and performance.

Collaboration Publications

"The Whole Is Greater Than the Sum of Its Parts": A Compatibility-Aware Multi-Teacher CoT Distillation Framework
Jin Cui, Jiaqi Guo, Jiepeng Zhou, Ruixuan Yang, Jiayi Lu, Jiajun Xu, Jiangcheng Song, Boran Zhao, Pengju Ren
IJCAI, 2026
arXiv

A compatibility-aware multi-teacher chain-of-thought distillation framework that leverages the complementary strengths of multiple teacher models.

MIND: From Passive Mimicry to Active Reasoning through Capability-Aware Multi-Perspective CoT Distillation
Jin Cui, Jiaqi Guo, Jiepeng Zhou, Ruixuan Yang, Jiayi Lu, Jiajun Xu, Jiangcheng Song, Boran Zhao, Pengju Ren
ACL, 2026
arXiv

A capability-aware multi-perspective chain-of-thought distillation method that transforms passive mimicry into active reasoning.

Education

Xi'an Jiaotong University
B.Eng. degree
2022 - 2026
Xi'an Jiaotong University
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence
Institute of Artificial Intelligence and Robotics
Research Intern
advisor: Professor Pengju Ren

9/2024 - 2/2026
Xi'an Jiaotong University
College of artificial intelligence
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence
Institute of Artificial Intelligence and Robotics
PhD Student
advisor: Professor Nanning Zheng
9/2026 -