Coder at Meta Superintelligence Lab.
PhD from MMLab@NTU, advised by Prof. Ziwei Liu.
I'm fortunate to work with talented people on frontier large language models, where I focus on post-training. Recently, I contributed to Muse Spark 1.3 and Muse.
Shipped
The latest version of Meta's Muse Spark model.
Current Interests
Agentic LLM Ability: Training models that plan over long horizons, use tools reliably, and learn from environment feedback to correct their own mistakes.
Information-dense Video Reasoning: Building models that identify salient details, connect dispersed evidence, and infer causal structure in complex visual narratives.
Benchmarking Model–Human Gaps: Developing diagnostic benchmarks that reveal systematic differences between models and humans, and suggest better optimization targets.
Haichen He*, Jiayi Zhou*, Sifeng Shang, Yihan Hu, Yuanhan Zhang, Kaiyang Zhou
An omni-modal benchmark on videos averaging 109 minutes, with questions that require continuous long-span viewing — Gemini-3.1-Pro scores 56.3% vs. 84.4% for humans.
Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Xiang Yue, Bo Li, Yuanhan Zhang, Ziwei Liu
A multi-discipline video benchmark for knowledge acquisition — adopted by Gemini-3 as part of their evaluation suite.
Yuanhan Zhang*, Yunice Chew*, Yuhao Dong, Aria Leo, Bo Hu, Ziwei Liu
A video reasoning benchmark that eliminates key-frame sampling bias — exposes a 48-point gap between GPT-4o (36.6%) and humans (84.3%).
Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, Chunyuan Li
An open video LMM built entirely on synthetic data — 1.2M+ HuggingFace downloads.
Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, Chunyuan Li
One model for image, multi-image, and video — SOTA across all three.
