News

1. Oct 2026
   KV-cache compression is not a free lunch: when n tokens are compressed into one cache entry, models spontaneously learn to keep only a single content-agnostic slot—inducing what we call phase sensitivity. We study this theoretically and empirically.
2. Jan 2026
   Paper “ERC loss for MoEs”, an auxiliary loss for MoE autonomy, accepted to ICLR 2026 Oral
3. Oct 2025
   Awarded National Scholarship for Doctoral Students Ranked 1st in GSAI
4. Sept 2025
   Paper “PolarQuant”, effective post-RoPE KV-cache quantization, accepted to NeurIPS 2025.
5. July 2025
   Joined ByteDance Seed Top Seed Intern
6. July 2025
   Paper “HoPE”, on why partial RoPE works, received the ACL 2025 SAC Highlights Award
7. May 2025
   Paper “Autonomy-of-Experts Models”, self-selecting MoE experts, accepted to ICML 2025.

Honors and Awards

1. 2025
   National Scholarship for Doctoral Students1st-Ranked in GSAI
2. 2025
   ByteDance Top Seed Intern
3. 2025
   ACL 2025 SAC Highlights Paper AwardTop 1.5%
4. 2025
   CIE-Tencent Doctoral Student Research Incentive ProgramHunYuan Large Language Model Special Project · 1 of 17 selected individuals nationwide
5. 2024
   CCF-Tencent Rhino-Bird Elite Talent Program1 of 50 selected individuals nationwide
6. 2023 – 2025
   Outstanding Innovative Talents Cultivation Funded ProgramsRenmin University of China

Academic Services

1. Area Chair
   EMNLP, ACLACL Rolling Review (ARR)
2. Reviewer
   ICML Gold, ICLR, NeurIPS

Internships

1. 2025.07 – Now
   ByteDance Seed LLM.
2. 2024.05 – 2025.07
   Tencent Hunyuan, mentored by Ruobing Xie. We conducted a series of work on autonomous MoE experts.
3. 2023.09 – 2024.05
   Alibaba, Tongyi Lab.
4. 2023.03 – 2023.09
   Microsoft Research, Machine Learning Area, mentored by Xu Tan. I am deeply grateful to Xu Tan for his patient and rigorous mentorship, which laid the foundation for my growth as a researcher. Our collaborative efforts on the Muzic project boast 5k stars on GitHub.

Recent Publications