I am a Ph.D. student at the Media Synthesis and Forensics Lab (formerly known as Multimedia Computing Group),Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS), advised by Professor Juan Cao and Associate Professor Qiang Sheng. Previously, I received my B.S. in E-commerce Engineering and Law from Beijing University of Posts and Telecommunications.
My research interests include LLM-generated text detection, LLM safety, efficient long-context inference, and hybrid sparse attention.
Feel free to ask me anything about my works! (Email: yangyehan25z@ict.ac.cn)
🔥 News
- 2026.07: 🏆 One co-author paper Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection received an ACL 2026 Outstanding Paper Award and was also selected as an SAC Highlight.
- 2026.06: 🎉 Our new work, Autonomy-of-Heads: Data-Free Sparse Attention from Frozen Query-Key Geometry, is now available. Developed in collaboration with Baidu’s ERNIE pre-training team, it is open for the community to explore, try, and use.
- 2026.03: Beyond the Final Actor was accepted to ACL 2026 as an oral presentation and nominated for an award.
- 2025.09: One Co-author Paper:From Judgment to Interference was accepted to NeurIPS 2025 as a poster presentation.
- 2025.06: 🎉 PhantomHunter: Detecting Unseen Privately-Tuned LLM-Generated Text via Family-Aware Learning is now available.
📝 Publications
Autonomy-of-Heads: Data-Free Sparse Attention from Frozen Query-Key Geometry
Yehan Yang, Junyuan Shang, Guanqun Zhao, Dianhai Yu, Shuohuan Wang
We introduce a data-free hybrid sparse attention method that identifies retrieval-oriented and streaming attention heads directly from the spectral geometry of frozen Query-Key projection matrices. At 256K context, the method achieves up to 3.24× Prefill acceleration, 9.14× Decode acceleration, and nearly 4× lower KV-cache memory.
Yang Li, Qiang Sheng, Zhengjia Wang, Yehan Yang, Danding Wang, Juan Cao
🏆 ACL 2026 Outstanding Paper Award!
RACE distinguishes human text polished by an LLM from LLM-generated text polished by a human through joint modeling of creator logic and editing style at the EDU level.
From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring
Yang Li, Qiang Sheng, Yehan Yang, Xueyao Zhang, Juan Cao
A streaming content monitoring system that reaches full-text-level safety performance after observing an average of only 18% of output tokens, using the 29K-sample FineHarm dataset.
-
arXiv 2026 PhantomHunter: Detecting Unseen Privately-Tuned LLM-Generated Text via Family-Aware Learning
Yehan Yang*, Yuhui Shi*, Qiang Sheng, Hao Mi, Beizhe Hu, Chaoxi Xu, and Juan Cao -
arXiv 2025 JingFang: A Traditional Chinese Medicine Large Language Model of Expert-Level Medical Diagnosis and Syndrome Differentiation-Based Treatment
Yehan Yang*, Tianhao Ma*, Ruotai Li, Xinhan Zheng, Guodong Shana
* Equal contribution.
🎖 Honors and Awards
- 2026-07, ACL 2026 Outstanding Paper Award, Association for Computational Linguistics.
- 2025-09, HICOOL Youth Entrepreneurship Dream Plan, Dream Supernova Award.
- 9th China International College Students’ “Internet+” Innovation and Entrepreneurship Competition, National Gold Award.
- National Computer Design Competition for College Students, National Bronze Award.
- Capital College Students’ Challenge Cup, Beijing Special Prize.
- Beijing-Taiwan Student Innovation and Entrepreneurship Competition, Outstanding Project at the finals.
📖 Educations
- 2025.09 - present, Ph.D. in Computer Science, Institute of Computing Technology, Chinese Academy of Sciences.
- 2021.09 - 2025.06, B.S. in E-commerce Engineering and Law, Beijing University of Posts and Telecommunications.
📚 Academic Services
- Reviewer
- ACL Rolling Review (Jul. 2026)
- Association for the Advancement of Artificial Intelligence (AAAI) Reviewer (2026)
💼 Experience
- 2026.03 - 07, Baidu ERNIE Pre-training Team — Intern. Working on efficient architectures for next-generation long-context reasoning models.
- 2024.04 - 2024.06, Microsoft STAC Innovation Program — Intern. Explored SLM-for-RAG and graph retrieval pre-classification, and developed the Javis-AI assistant workflow for voice-based task planning and reminder emails.