Updates

Recent research releases, milestones, and selected talks.

News

Recent milestones

  • MegaScience accepted to COLM 2026.
  • MegaMath accepted to COLM 2025.
  • OctoThinker accepted to the AI4Math Workshop @ ICML 2025.
  • Released the OctoThinker technical report and MegaMath-Web-Pro-Max (70B+ tokens).
  • ProX accepted to ICML 2025.
  • Launched OctoThinker, our mid-training effort for RL-friendly reasoning base models.
  • Started my PhD at SJTU.

Talks

Selected Research Talks

  • OctoThinker: Mid-training Incentivizes RL Scaling

    • ERNIE team, Baidu
    • Hunyuan team, Tencent
    • FAIR/GenAI team, Meta
    • MLNLP Community
  • Pre-training Data Engineering in LLMs

    • Ant Group
    • MLNLP Community
  • Generative AI for Math: MathPile

    • Doubao Team, ByteDance
    • OpenDataLab, Shanghai AI Lab