Zengzhi Wang 王增志

PhD student at GAIR Lab, Shanghai Jiao Tong University.
Advised by Prof. Pengfei Liu.

I work on the engineering and science of data for foundation language models. Specifically:

  • Data Science: building scaling ladders, evaluating how data recipes scale, and designing scalable data recipes.
  • Data Engineering: parsing, refining, auditing, organizing, and synthesizing data.

Currently, I am a research intern on an industry pre-training team.

Always happy to connect—feel free to reach out.

Zengzhi Wang smiling in an ornate interior in the UAE.
We should dream big.

Selected work

Explore research
  • OctoThinker

    An early systematic study of why mid-training matters for scalable reinforcement learning.

    Connects data quality and format to RL performance, with a two-stage recipe that prepares base models for stronger downstream RL.

  • MegaMath

    The largest open math pre-training corpus at release.

    Used in NVIDIA Nemotron 3 Nano, AMD Instella-MoE, Apertus, and SmolLM3.

    ~968K total downloads on Hugging Face as of
  • ProX

    Making expert-style data refinement scalable with small language models.

    Generates and executes edits tailored to each training example to improve pre-training efficiency.

Training as connection

How training organizes knowledge.

A conceptual view of training: points represent knowledge, and connections suggest how a model uses it.

Build broad knowledge Connect diverse knowledge units into a general representation.

Training as the progressive organization of a knowledge network Dots represent knowledge units. Pre-training creates broad connections, mid-training strengthens useful multi-step paths, and post-training shapes which paths are followed.