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

Publications

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

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.

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

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

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

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

Collaboration Publications

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

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

Education