Jiaxiang Yu

I am currently a second-year EECS PhD student in Sky Computing Lab at UC Berkeley, advised by Prof. Alvin Cheung. My research focuses on the gap between algorithmic promise and systems reality in large language model inference. I am currently working on Speculative Decoding for Large Language Model inference optimization.

[NEW] Our work “Speculative Decoding: Performance or Illusion?” have been accepted by MLSys 2026. Here is the paper, code, and blog. If you are curious about how and why speculative decoding perform in more realistic settings (e.g. with continuous batching and CUDA Graph), this paper is worth a read!

Selected Publications

Speculative Decoding: Performance or Illusion?

Xiaoxuan Liu*, Jiaxiang Yu*, Jongseok Park, Ion Stoica, and Alvin Cheung

2025

MLSys 2026 Best Paper Honorable Mention

arXiv Bib Blog Code

@arvix{liu2025speculativedecodingperformanceillusion,
  title = {Speculative Decoding: Performance or Illusion?},
  author = {Liu*, Xiaoxuan and Yu*, Jiaxiang and Park, Jongseok and Stoica, Ion and Cheung, Alvin},
  year = {2025},
  eprint = {2601.11580},
  archiveprefix = {arXiv},
  primaryclass = {cs.CL},
  note = {MLSys 2026 Best Paper Honorable Mention},
}

Melange: Cost Efficient Large Language Model Serving by Exploiting GPU Heterogeneity

Tyler Griggs*, Xiaoxuan Liu*, Jiaxiang Yu, Doyoung Kim, Wei-Lin Chiang, Alvin Cheung, and Ion Stoica

2024

arXiv Bib Code

@arvix{griggs2024melangecostefficientlarge,
  title = {Melange: Cost Efficient Large Language Model Serving by Exploiting GPU Heterogeneity},
  author = {Griggs*, Tyler and Liu*, Xiaoxuan and Yu, Jiaxiang and Kim, Doyoung and Chiang, Wei-Lin and Cheung, Alvin and Stoica, Ion},
  year = {2024},
  eprint = {2404.14527},
  archiveprefix = {arXiv},
  primaryclass = {cs.DC},
}