Understanding components tolerate substantial compression with minimal quality loss.
Generation components are highly sensitive to pruning.
We propose MoE Adaptation: partition the generation module into multiple experts and activate them sparsely, recovering performance while reducing active parameters.

The compressed models reduce the active parameters in the generation module by half while maintaining comparable or even improved GenEval scores.

Model	Active Experts	Repo
BAGEL-MoE-7B-GEN-32to16	16 / 32	LLM-Drop/BAGEL-MoE-7B-GEN-32to16
BAGEL-MoE-7B-GEN-16to8	8 / 16	LLM-Drop/BAGEL-MoE-7B-GEN-16to8

GenEval Results

Model	SO	TO	CT	CL	POS	ATTR	ALL
BAGEL-7B-MoT (original)	0.99	0.94	0.81	0.95	0.72	0.77	0.86
BAGEL-MoE-7B-GEN-32to16	0.99	0.94	0.87	0.93	0.79	0.78	0.89
BAGEL-MoE-7B-GEN-16to8	1.00	0.92	0.82	1.00	0.77	0.83	0.89

Installation

bash

conda create -n efficient_ug python=3.10
conda activate efficient_ug
pip install -r requirements.txt

Evaluation

bash

# GenEval evaluation
bash scripts/eval/bagel/run_geneval_wr.sh

Citation

bibtex

@article{he2025sparsity,
  title   = {Understanding and Harnessing Sparsity in Unified Multimodal Models},
  author  = {He, Shwai and others},
  journal = {arXiv preprint arXiv:2512.02351},
  year    = {2025}
}

LLM-Drop/BAGEL-MoE-7B-GEN-32to16

BAGEL-MoE-7B-GEN

GenEval Results

Installation

Evaluation

Citation

LLM-Drop/
BAGEL-MoE-7B-GEN-32to16