Ask-Solve-Generate-BAGEL

Ask-Solve-Generate BAGEL Adapters

This repository contains BAGEL role adapters trained with the Ask-Solve-Generate self-evolving framework.

Contents

• role_proposer.pt: proposer role adapter.
• role_solver.pt: solver role adapter.
• role_generator.pt: generator role adapter.
• adapter_roles.json: role-to-adapter manifest used by the BAGEL loader.
• release_manifest.json: release metadata.

The package contains adapter weights only. Use these adapters with the public codebase and the upstream BAGEL base model.

Usage

git clone https://github.com/mbzuai-oryx/Ask-Solve-Generate.git
cd Ask-Solve-Generate

Download this model repository with huggingface_hub or huggingface-cli, then pass the snapshot path as LORA_CHECKPOINT_PATH:

MODEL_PATH=/path/to/BAGEL-7B-MoT \
LORA_CHECKPOINT_PATH=/path/to/Ask-Solve-Generate-BAGEL \
DATA_DIR=/path/to/unlabeled/images \
OUTPUT_DIR=/path/to/output/bagel \
NPROC_PER_NODE=8 \
bash Bagel/scripts/B1_unified_training.sh

The BAGEL loader accepts this folder directly because it contains role_<role>.pt files and adapter_roles.json.

Links

• Code: https://github.com/mbzuai-oryx/Ask-Solve-Generate
• Project page: https://mbzuai-oryx.github.io/Ask-Solve-Generate/
• Paper: https://arxiv.org/abs/2606.27376

Citation

@article{thawkar2026asksolvegenerate,
  title={Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards},
  author={Thawkar, Ritesh and Venkatraman, Shravan and Thawakar, Omkar and Shaker, Abdelrahman and Khan, Fahad and Cholakkal, Hisham and Khan, Salman and Anwer, Rao Muhammad},
  journal={arXiv preprint arXiv:2606.27376},
  year={2026}
}