ReFusion-8B-ESPO

ReFusion 8B trained with ESPO v19 (ELBO-based Sequence-level Policy Optimization). Sequence-level RL prevents the training collapse seen in token-level methods. +1.6pp nonzero rate improvement on test split vs SFT. Part of the STAD80 project: Generative Action Planning via Discrete Flow Matching.

Paper

Generative Action Planning via Discrete Flow Matching with Online Reinforcement Fine-Tuning

• Author: Muhammad Enrizky Brillian
• Institution: University of Toronto Scarborough

Training Details

• Dataset: FormFactory (992 train / 124 val / 124 test tasks, 25 form types, 8 domains)
• Infrastructure: Single NVIDIA A10G GPU (24GB VRAM) on Anyscale
• Framework: PyTorch + PEFT (LoRA/QLoRA)

Citation

If you use this model, please cite:

@article{brillian2026flowgrpo,
  title={Generative Action Planning via Discrete Flow Matching with Online Reinforcement Fine-Tuning},
  author={Brillian, Muhammad Enrizky},
  year={2026}
}

This model was trained and evaluated on the FormFactory benchmark:

@misc{li2025formfactory,
  title={FormFactory: An Interactive Benchmarking Suite for Multimodal Form-Filling Agents},
  author={Bobo Li and Yuheng Wang and Hao Fei and Juncheng Li and Wei Ji and Mong-Li Lee and Wynne Hsu},
  year={2025},
  eprint={2506.01520},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2506.01520}
}