Photo of Tianyu Liu

Tianyu Liu刘天宇

Joint Ph.D. student at USTC and Shanghai AI Laboratory.

About me

Hi there! I am Tianyu Liu (刘天宇), a joint Ph.D. student at USTC and Shanghai AI Laboratory, supervised by Xiao Sun. My research focuses on efficient inference for LLMs, especially speculative decoding. I proposed PEARL (ICLR 2025), the first parallel speculative decoding framework, and D-cut, which now ships in Tencent's open-source AngelSlim and has been followed up by DeepSeek's DSpark.

I expect to graduate in 2027 and am on the job market for roles in efficient LLM inference and AI systems. I am also open to research collaborations on related topics. Feel free to reach out at tianyu_liu@mail.ustc.edu.cn.

News

  • 2026.10Awarded the National Scholarship for Ph.D. students, one of the highest student honors in China!
  • 2026.09TALON has been accepted to NeurIPS 2026!
  • 2026.09HIPPO has been accepted to the EMNLP 2026 main conference!
  • 2026.08KVShot has been accepted as a spotlight at the COLM 2026 Workshop on Efficient Reasoning!
  • 2026.07Released AngelSpec, Tencent Hunyuan's technical report on real-world, high-performance inference with speculative decoding.
Earlier updates
  • 2026.07D-cut is now on arXiv, achieving up to 3.0× speedup over autoregressive decoding for MoE models under high concurrency.
  • 2026.06Introduced D-cut, an adaptive verification-depth pruning method that accelerates speculative decoding at high concurrency.
  • 2026.05Released two new preprints on speculative decoding: KVShot and Graft!
  • 2026.04Double was accepted to the ACL 2026 main conference as an oral presentation, and LogitSpec to ACL 2026 Findings!
  • 2026.01SpecBranch was accepted to ICLR 2026!
  • 2026.01Released four preprints: TALON, KALE, Double, and HIPPO!
  • 2025.10Released nano-PEARL, an engineering follow-up to PEARL with multi-GPU draft-target disaggregation.
  • 2025.07Released LogitSpec, a training-free, retrieval-based speculative decoding method, on arXiv!
  • 2025.01PEARL was accepted to ICLR 2025!
  • 2025.01One paper accepted to NAACL 2025. Many thanks to Qitan for leading it!
  • 2023.09REST was accepted to NeurIPS 2023!

Selected Publications

†: corresponding author; *: equal contribution

PEARL illustration

ICLR 2025

PEARL: Parallel Speculative Decoding with Adaptive Draft Length

Tianyu Liu, Yun Li†, Qitan Lv, Kai Liu, Jianchen Zhu, Winston Hu, Xiao Sun†

[Paper]  [Project]  [Code] [BibTeX]

Check out nano-PEARL, a draft-target disaggregated LLM serving system built on parallel speculative decoding!

D-cut illustration

Preprint

D-cut: Adaptive Verification Depth Pruning for Batched Speculative Decoding

Tianyu Liu, Yuhao Shen, Rui Cen, Junhan Shi, Jiebin Zhang, Guangshuo Qin, Hong Liu, Song Liu, Guanghua Yu†, Jianchen Zhu

[Paper]  [Code]  [Docs] [BibTeX]

Followed up by DeepSeek's DSpark!

KVShot illustration

COLM 2026 ER WorkshopSpotlight

When Hidden States Drift: Can KV Caches Rescue Long-Range Speculative Decoding?

Tianyu Liu, Yuhao Shen, Xinyi Hu, Baolin Zhang, Hengxin Zhang, Jun Dai, Jun Zhang†, Shuang Ge†, Lei Chen, Yue Li, Mingcheng Wan

[Paper] [BibTeX]

TALON illustration

NeurIPS 2026

TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees

Tianyu Liu*, Qitan Lv*, Yuhao Shen, Jun Zhang, Xiao Sun†, Xiaoyan Sun

[Paper] [BibTeX]

LogitSpec illustration

ACL 2026 Findings

LogitSpec: Accelerating Retrieval-based Speculative Decoding via Next Next Token Speculation

Tianyu Liu*, Qitan Lv*, Hao Li, Xing Gao, Xiao Sun†, Xiaoyan Sun

[Paper]  [Code] [BibTeX]

REST illustration

NeurIPS 2023

Learning Rule-Induced Subgraph Representations for Inductive Relation Prediction

Tianyu Liu, Qitan Lv, Jie Wang†, Shuling Yang, Hanzhu Chen

[Paper]  [Code] [BibTeX]

Other Publications

Tech ReportAngelSpec: Towards Real-World High Performance Inference with Speculative Decoding

Hong Liu, Rui Cen, Junhan Shi, Guangshuo Qin, Jiebin Zhang, Tianyu Liu, Runzhi Fan, Guoliang Zhao, Ruobing Xie, Kai Zhang, Song Liu, Guanghua Yu†, Jianchen Zhu

[Paper]  [Code]  [Blog] [BibTeX]

arXiv 2026Graft — Draft Less, Retrieve More: Hybrid Tree Construction for Speculative Decoding

Yuhao Shen‡, Tianyu Liu‡, Xinyi Hu‡, Quan Kong, Baolin Zhang, Jun Dai, Jun Zhang†, Shuang Ge, Lei Chen, Yue Li, Mingcheng Wan, Cong Wang†

‡: core contribution

[Paper] [BibTeX]

EMNLP 2026 MainHIPPO: Accelerating Video Large Language Models Inference via Holistic-aware Parallel Speculative Decoding

Qitan Lv*, Tianyu Liu*, Wen Wu, Xuenan Xu, Bowen Zhou, Feng Wu, Chao Zhang†

[Paper] [BibTeX]

arXiv 2026KALE: Enhancing Knowledge Manipulation in Large Language Models via Knowledge-aware Learning

Qitan Lv*, Tianyu Liu*, Qiaosheng Zhang†, Xingcheng Xu†, Chaochao Lu

[Paper] [BibTeX]

ACL 2026 MainOralDouble: Breaking the Acceleration Limit via Double Retrieval Speculative Parallelism

Yuhao Shen, Tianyu Liu, Junyi Shen, Jinyang Wu, Quan Kong, Li Huan, Cong Wang

[Paper]  [Code] [BibTeX]

ICLR 2026SpecBranch: Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism

Yuhao Shen, Junyi Shen, Quan Kong, Tianyu Liu, Yao Lu, Cong Wang

[Paper]  [Code] [BibTeX]

NAACL 2025Exploiting Edited Large Language Models as General Scientific Optimizers

Qitan Lv*, Tianyu Liu*, Hong Wang†

[Paper] [BibTeX]

Education

  • 2022.09 – PresentPh.D. in Information and Communication EngineeringUniversity of Science and Technology of China
  • 2018.09 – 2022.06B.Eng. in Computer Science and TechnologyCentral University of Finance and Economics

Academic Service

  • ReviewerICLR'25, ICLR'26, ICLR'27, WWW'25, NeurIPS'25, NeurIPS'26