ReFusion-8B-MDPO

ReFusion 8B trained with MDPO (Masked Diffusion Policy Optimization). Best result in the paper: 91.9% nonzero rate / 0.445 average reward on 124 test tasks (+31.4pp over SFT). Temporal advantage decomposition with mu=1.

Paper

Concentrate or Collapse: When Reinforcement Learning Meets Diffusion Language Models for Web Planning

• Author: Muhammad Enrizky Brillian
• Institution: University of Toronto Scarborough
• Code: https://github.com/billy-enrizky/openbrowser-ai

Training Details

• Dataset: FormFactory (992 train / 124 val / 124 test tasks, 25 form types, 8 domains)
• Infrastructure: NVIDIA L40S (ReFusion) / A10G (FS-DFM) on Modal.com
• Framework: PyTorch + PEFT (LoRA/QLoRA)
• Training prompts: 50 (sequence-level), G=4 rollouts per prompt

Citation

@article{brillian2026flowgrpo,
  title={Concentrate or Collapse: When Reinforcement Learning Meets Diffusion Language Models for Web Planning},
  author={Brillian, Muhammad Enrizky},
  year={2026}
}