Fanchao Chen

I am a final-year CS PhD student at the University of Wisconsin–Madison, advised by Prof. Shivaram Venkataraman. I am currently a research intern on the Nemotron Post-training Team at NVIDIA. Previously, I worked at ByteDance Seed, Moonshot AI (Kimi), and Microsoft Research Asia. I received my bachelor's degree in Intelligence Science and Technology from Fudan University. My research interests are about Efficient AI Systems and Continual Learning.

Email  /  CV  /  LinkedIn  /  Twitter  /  GitHub

Fanchao Chen

Work Experience

Selected Publications

* indicates equal contribution.

Deriving a Quantization-Aware Distillation Recipe for NVFP4 Inference Accuracy Recovery
Fanchao Chen, Minghao Yan, Jiantao Jiao, Shivaram Venkataraman
arXiv, 2026

We develop an NVFP4 quantization-aware distillation (QAD) recipe from open data, using domain-specific failure analysis and small-scale experiments to guide data selection and iteration. On Nemotron models, it uses up to 4.7× fewer training tokens than NVIDIA's official QAD recipe while recovering more of the quantization-induced accuracy loss.

Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards
Fanchao Chen*, Hengyu Fu*, Shivaram Venkataraman, Jiantao Jiao
arXiv, 2026

HDL turns the hindsight signal from token-level supervision into trajectory branching: it uses hindsight-induced changes in token log-probabilities to pick branch points and samples new continuations from shared prefixes. Compared with GRPO, it speeds up rollouts by up to 1.8× while improving performance on math, code, and agentic tasks.

Towards Full Pipeline FP8 Reinforcement Learning for LLMs
Fanchao Chen, Ziheng Jiang, Ziyun Wei, Zheng Zhong, Du Li, Chi Zhang, Haibin Lin, Shivaram Venkataraman
COLM, 2026

We enable both rollout and training in FP8 for LLM post-training. We find that FP8 quantization distorts importance ratios and causes over-clipping, and propose Calibrated Clipping to stabilize training, recovering BF16-level performance with up to 1.5× training throughput.

PipeTune: Tuning Pipeline Parallelism for Efficient Vision-Language Model Training
Fanchao Chen, Shivaram Venkataraman
arXiv, 2026

Heterogeneous sequence lengths and modality mixes in VLM training data create pipeline bubbles. PipeTune jointly tunes micro-batch packing, ordering, and sizing, and moves vision-encoder computation into idle time, guided by a lightweight simulator that adapts as the data mix shifts. It speeds up training by up to 1.25× without changing training semantics.

Teaching & Service