Cross-modal Information Flow in Multimodal Large Language Model

This is the official repository for our CVPR paper: Cross-modal Information Flow in Multimodal Large Language Models

Installation

1. git clone https://github.com/FightingFighting/cross-modal-information-flow-in-MLLM.git
2. cd cross-modal-information-flow-in-MLLM
3. Please following LLaVANEXT(https://github.com/LLaVA-VL/LLaVA-NeXT) to install the environment: llava

After installing llava environment, you will find an LLaVA-NeXT folder in cross-modal-information-flow-in-MLLM

Dataset

Our dataset is collected from GQA. The collected datasets are in datasets. For the images, please download from here.

Use

Information flow

1. Open scripts/informationFlow.sh

2. Setting:

   current_window: how many layers you want to block for the attention knock at a time;

   current_block_desc: which kind of information flow you want to block;

   model_path: which kind of model you want to explore;

   convmode: different kind of model has different convmode;

   dataset: which kind of task you want to explore;

   imagefolder: the image folder

3. Run sbatch scripts/informationFlow.sh

current_block_desc can be chosen from:

  "Question->Last"
  "Image->Last"
  "Image->Question"
  "Last->Last"
  "Image Central Object->Question"
  "Image Without Central Object->Question"

model_path and convmode can be chosen from:

  model_path="liuhaotian/llava-v1.6-vicuna-7b" convmode="vicuna_v1"
  model_path="lmms-lab/llama3-llava-next-8b"  convmode="llava_llama_3"
  model_path="liuhaotian/llava-v1.5-7b"   convmode="vicuna_v1"
  model_path="liuhaotian/llava-v1.5-13b"   convmode="vicuna_v1"

dataset can be chosen from:

  datasets/GQA_val_correct_question_with_choose_ChooseAttr.csv
  datasets/GQA_val_correct_question_with_positionQuery_QueryAttr.csv
  datasets/GQA_val_correct_question_with_existThatOr_LogicalObj.csv
  datasets/GQA_val_correct_question_with_twoCommon_CompareAttr.csv
  datasets/GQA_val_correct_question_with_relChooser_ChooseRel.csv
  datasets/GQA_val_correct_question_with_categoryThatThisChoose_objThisChoose_ChooseCat.csv

Probability of answer word tracking

1. Open scripts/last_position_answer_prob.sh

2. Setting:

   model_path: which kind of model you want to explore;

   convmode: different kind of model has different convmode;

   dataset: which kind of task you want to explore;

   imagefolder: the image folder

3. Run sbatch scripts/last_position_answer_prob.sh

Visulization

if you want to merge several lines into one figure, you can run python vil/merge_lineplot.py.

For example, you might already get the results of the information flow: Question->Last,Image->Last,Last->Last, and you want to merge these three lines in one Figure, and then you could run python vil/merge_lineplot.py.

Cite

If this project is helpful for you, please cite our paper:

@article{zhang2024cross,
  title={Cross-modal Information Flow in Multimodal Large Language Models},
  author={Zhang, Zhi and Yadav, Srishti and Han, Fengze and Shutova, Ekaterina},
  journal={arXiv preprint arXiv:2411.18620},
  year={2024}
}

Acknowledgement

The code is built upon https://github.com/google-research/google-research/tree/master/dissecting_factual_predictions and LLaVA.

Our used datasets are collected from GQA