huangrh9/AlphaGRPO-srr

AlphaGRPO-SRR

🤗 Models | 📄 Paper | 🌐 Project Page | 💻 GitHub | 🧩 Base Model

Model Summary

AlphaGRPO-SRR is a PEFT LoRA adapter for BAGEL-7B-MoT, trained with AlphaGRPO for self-reflective refinement.

This repository contains adapter weights only. Please load it together with the BAGEL base model. The adapter uses LoRA rank 32 and alpha 64.

Usage

Set the adapter path when running reflection-enabled AlphaGRPO/BAGEL inference or evaluation:

export BAGEL_LORA_PATH=/path/to/AlphaGRPO-srr

For installation, evaluation scripts, and full usage examples, please see the GitHub repository.

Citation

@inproceedings{huang2026alphagrpo,
  title={AlphaGRPO: Unlocking Self-Reflective Multimodal Generation in Unified Multimodal Models via Decompositional Verifiable Reward},
  author={Huang, Runhui and Wu, Jie and Yang, Rui and Liu, Zhe and Zhao, Hengshuang},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}