Jiancheng Dong

I am now a second-year PhD candidate at AML Lab, City University of Hong Kong, supervised by Prof. Xiangyu Zhao. I received my bachelor's degree from the School of Artificial Intelligence at Nanjing University.

My research focuses on LLM context compression, with the long-term goal of building AI systems that proactively learn and evolve through deployment (see Ilya Sutskever's discussion of continual learning).

I am currently an intern with ByteDance's Content Consumption team.

News

2026.08: Our paper DiVA-Former was accepted to the EMNLP'26 Main Conference.
2026.08: Our paper Behavior-Equivalent Token was accepted to CIKM'26 as an oral presentation.
2025.09: Start my PhD study at CityU.

Selected Publications

For the full publication list, please refer to my Google Scholar.

LayerExit encoder-layer diagnostics comparing next-token cross-entropy and question-answering accuracy

LayerExit: Adaptive Intermediate-Layer Exiting for Context Compression

Jiancheng Dong, Pengyue Jia, Yingyi Zhang, Haotian Nan, Chao Zhang, Jingyu Peng, Tianbo Huang, Lihua Zhang, Xiaopeng Li, Wanyu Wang, Xiangyu Zhao

In submission, 2026

LayerExit adaptively chooses between an efficient intermediate encoder layer and the final layer, improving question-answering accuracy while reducing compression cost.

How to Utilize Complementary Vision-Text Information for 2D Structure Understanding

Jiancheng Dong, Pengyue Jia, Derong Xu, Jiawei Cheng, Jingyu Peng, Chao Zhang, Bowen Liu, Xin Sun, Lixin Su, Shuaiqiang Wang, Dawei Yin, Xiangyu Zhao

EMNLP'26 Main Paper Code

DiVA-Former uses visual tokens to distill long serialized tables into compact representations that preserve both structure and fine-grained text.

Behavior-Equivalent Token method overview

Behavior-Equivalent Token: Single-Token Replacement for Long Prompts in LLMs

Jiancheng Dong, Pengyue Jia, Jingyu Peng, Maolin Wang, Yuhao Wang, Lixin Su, Xin Sun, Shuaiqiang Wang, Dawei Yin, Xiangyu Zhao

CIKM'26 Oral Paper Code

A single learned Behavior-Equivalent Token can replace a long system prompt while preserving its downstream behavior and sharply reducing prompt overhead.

Threshold Filtering Packing method overview

Threshold Filtering Packing for Supervised Fine-Tuning: Training Related Samples within Packs

Jiancheng Dong, Lei Jiang, Wei Jin, Lu Cheng

NAACL'25 Main Paper

Threshold Filtering Packing groups related yet diverse samples into training packs to reduce cross-sample interference during supervised fine-tuning.

Internship

ByteDance, Content Consumption

5/2026 -- Present

Baidu Inc., Search Science

5/2025 -- 4/2026

Thanks to Xin Sun, Lixin Su, Shuaiqiang Wang, and Dawei Yin for their guidance!

University of Illinois Chicago, Responsible and Reliable AI Lab

3/2024 -- 9/2024

Thanks to Prof. Lu Cheng and Wei Jin for their guidance!

Nanjing University, Natural Language Processing Group

4/2023 -- 1/2024

Thanks to Prof. Xinyu Dai and Prof. Zhen Wu for their guidance!

Teaching

2026.09 - 2026.12, Teaching Assistant, DSC6020 Artificial Intelligence for Scientific Knowledge Discovery

Visitors

Visitor statistics by FeedPulse