ヽ(•‿•)ノ Building multimodal foundation models, one token at a time.
(ﾉ◕ヮ◕)ﾉ*:･ﾟ✧ Hi there!

I'm Xuehan Xiong

Senior Staff Software Engineer @ Google Cloud AI

I work on large language and vision-language models. I currently lead post-training quality on Gemini Enterprise Agent Platform training clusters. Before that, I led multimodal LLM pretraining at ByteDance Seed and was part of the initial Gemini team at Google.

✉ Email me 🎓 Google Scholar in LinkedIn
LLM · VLM
( •̀ ω •́ )✧
Research by the numbers
24K+citations
27h-index
34i10-index
via Google Scholar

Things I've worked on

(⌐■_■)

Over about 15 years I've moved from 3D scene understanding to face alignment, video understanding, and now multimodal LLMs.

✦Gemini 1 & 1.5→ ◕‿◕Seed1.5-VL→ ▶Video Transformers→ ✂Video object segmentation→ ʘ‿ʘSupervised Descent Method→ 👀Face & gaze tracking→ 🏆Epic-Kitchens 2022 winner→ 🏆Kinetics 2020 winner↗ 🏅Test of Time Award 2026→ ▦Semantic 3D building models→

Blog

✍(◔◡◔)
GEAP Training Cluster

Post-Training Conversational Agents via Self-Evolving Data Synthesis

Guanghao Mei, Weiran Zhao, Jun-Liang Lin, Siqi Miao, Xuehan Xiong

A self-evolving pipeline turns policy documents and tools into verifiable multi-turn tasks for training conversational agents with SFT then GRPO. On τ3-Banking, Nemotron-3.5-Super-VL-120B-A12B improves from 15.1% to 44.0% and Lightning-30B-A3B from 7.6% to 32.5%.

Read the post → GEAP Training Cluster

Model Distillation Best Practices

Xuehan Xiong, Youbao Tang, Fei Xia, Bao Thach, Joseph Pagadora

The first post in a series on off-policy distillation. We fine-tune Qwen3 base models (1.7B–14B) on teacher-generated reasoning traces for math, instruction following, ARC-AGI 1, and τ²-bench. Covers dataset curation, teacher selection, rollout counts, rejection sampling, and learning-rate scaling rules.

Read the post →

Publications

φ(゜▽゜*)♪

Selected papers. Citation counts come from Google Scholar. See the full list on Scholar.

No papers found for that search! (╥﹏╥)

My journey

✈ ┌(ㆆ㉨ㆆ)ʃ

From Wuhan to the Bay Area. Hover over a stop, or click one in the list to fly there.

    Education & fun facts

    (｡♥‿♥｡)

    🎓Education

    • PhD, Robotics
      Carnegie Mellon University      2015
    • M.S., Robotics
      Carnegie Mellon University      2011
    • B.S., Computer Science
      University of Arizona, Summa Cum Laude      2009
    Was I really a sprinter? ε=ε=┌( >_<)┘
    Yes! I was on the University of Arizona Men's Track & Field team (2005–2009), an NCAA All-American (2007), and the NCAA Western Regional Champion (2008).
    Even before college? (ง •̀_•́)ง
    In 2004 I was the National Junior Champion in the 200m dash, a member of the National Junior Team, and a bronze medalist at the Asian Junior Track & Field Championships.
    Where's SDM used? (☞ﾟヮﾟ)☞
    The Supervised Descent Method (CVPR 2013) was the core technology at FacioMetrics. After Facebook acquired the company, it powered face effects in Facebook and Instagram.