VFA: Empowering Multilingual MLLMs via Vision-Free Adaptation

Yixia Li^1*◊, Yaqing Shi^3*, Zhiwen Ruan¹, Dongdong Zhang², Lingjie Jiang⁴,
Shaohan Huang², Yun Chen³, Guanhua Chen^1†, Furu Wei²

¹Southern University of Science and Technology   ²Microsoft Research Asia
³Shanghai University of Finance and Economics   ⁴Peking University
^*Equal contribution   ^†Corresponding author   ^◊Work done during internship at Microsoft Research Asia.

---

Overview

VFA (Vision-Free Adaptation) is an efficient framework for enhancing the multilingual capabilities of Multimodal Large Language Models (MLLMs) without relying on scarce multilingual image-text data.

VFA decouples multilingual learning from visual alignment through a two-stage vision-free strategy:

• Stage 1: Fine-tune the base LLM on multilingual text-only data to derive a multilingual task vector
• Stage 2: Merge the multilingual task vector into the vision-aligned MLLM, preserving visual grounding while injecting multilingual competence

Installation

1. Clone the repository

git clone https://github.com/sustech-nlp/VFA.git
cd VFA

2. Install LLaMA-Factory

LLaMA-Factory is included locally at src/LLaMA-Factory/. Install it into a virtual environment:

bash scripts/env_setup/uv_llamafactory.sh

3. Install evaluation frameworks

lmms-eval (for multimodal benchmarks):

bash scripts/env_setup/uv_lmms_eval.sh

OpenCompass (for text-only benchmarks):

bash scripts/env_setup/uv_opencompass.sh

Quick Start

Stage 1: Fine-tune Base LLM on Multilingual Text Data

Run training (example with Qwen2.5-7B):

bash scripts/train/run_train.sh configs/finetune_llm/Qwen2.5-7B/multilingual_sft_99999.yaml

Stage 2: Merge Multilingual Task Vector into MLLM

source uv_llamafactory/bin/activate

Task Arithmetic:

python scripts/merge/vlm_model_merge.py \
    --vlm_model_type Qwen/Qwen2.5-VL-7B-Instruct \
    --llm_model_type Qwen/Qwen2.5-7B \
    --vlm_model_path /path/to/Qwen2.5-VL-7B-Instruct \
    --llm_model_path /path/to/finetuned-Qwen2.5-7B \
    --base_model_path /path/to/Qwen2.5-7B \
    --output_dir outputs/Qwen2.5-VL-7B-VFA \
    --mode task_arithmetic \
    --alpha 0.8

Weighted Averaging:

python scripts/merge/vlm_model_merge.py \
    --vlm_model_type Qwen/Qwen2.5-VL-7B-Instruct \
    --llm_model_type Qwen/Qwen2.5-7B \
    --vlm_model_path /path/to/Qwen2.5-VL-7B-Instruct \
    --llm_model_path /path/to/finetuned-Qwen2.5-7B \
    --output_dir outputs/Qwen2.5-VL-7B-VFA \
    --mode weighted_average \
    --alpha 0.8

TIES-Merging:

python scripts/merge/vlm_model_merge.py \
    --vlm_model_type HuggingFaceM4/Idefics3-8B-Llama3 \
    --llm_model_type meta-llama/Llama-3.1-8B \
    --vlm_model_path /path/to/Idefics3-8B-Llama3 \
    --llm_model_path /path/to/finetuned-Llama-3.1-8B-Instruct \
    --base_model_path /path/to/Llama-3.1-8B \
    --output_dir outputs/Idefics3-8B-VFA \
    --mode ties \
    --alpha 0.8 \
    --density 0.2

Stage 3: Evaluate

Multilingual multimodal evaluation (lmms-eval):

export TASKS="maxm,xgqa,xmmmu,xm100,marvl,m3exam"
bash scripts/eval/eval_lmms.sh /path/to/merged-model outputs/eval_results

Text-only multilingual evaluation (OpenCompass):

export TASKS="tydiqa_gen,mmmlu_gen_d5017d,xnli_gen_973734,mhellaswag_gen_1a6b73,mifeval_gen_79f8fb,mlogiqa_gen_36c4f9,flores_gen_2697d7"
bash scripts/eval/eval_opencompass.sh /path/to/merged-model outputs/eval_results chat "$TASKS"

General multimodal evaluation (lmms-eval):

export TASKS="ocrbench_v2,mmbench_en_dev,mmmu_val,mathvista_testmini"
bash scripts/eval/eval_lmms.sh /path/to/merged-model outputs/eval_results

Released Checkpoints

The following trained checkpoints are available on ModelScope.

Supported Models

VLM-LLM Pairs

Merging Methods

Data

We fine-tune base LLMs on a 100K-example subset of the Multilingual-SFT dataset.

To download and preprocess the dataset (filters out multimodal samples):

python scripts/data_processing/process_multilingual_sft.py
# Output: data/multilingual_sft_99999.json

Citation

If you find this work useful, please cite our paper:

@inproceedings{li2026vfa,
    title={VFA: Empowering Multilingual MLLMs via Vision-Free Adaptation},
    author={Yixia Li and Yaqing Shi and Zhiwen Ruan and Dongdong Zhang and Lingjie Jiang and Shaohan Huang and Yun Chen and Guanhua Chen and Furu Wei},
    booktitle={Proceedings of ACL},
    year={2026}
}

Acknowledgments

This project builds upon the following excellent open-source projects:

• LLaMA-Factory for training infrastructure
• lmms-eval and Pangea for multimodal evaluation
• OpenCompass for text evaluation