Measuring Agreeableness Bias in Multimodal Models

This repository contains the code and resources for replicating the experiments and analysis from our paper Measuring agreeableness bias in Multimodal Models.

Table of Contents

1. Introduction
2. Repository Structure
3. Setup
4. Generating Experiments
5. Evaluating Models
6. Analyzing Results
7. Models Tested
8. Benchmarks
9. Results
10. Contributing
11. License

Introduction

This project investigates the phenomenon of "agreeableness bias" in multimodal language models - the tendency of these models to disproportionately favor visually presented information, even when it contradicts their prior knowledge. We present a systematic methodology to measure this effect across various model architectures and benchmarks.

Repository Structure

• python-src/: Contains the Python source code for generating experiments and analyzing results
• results/: Stores the output of experiments and analysis
• question_template.html: HTML template for rendering questions
• plots/: Contains plots for probability deltas for gpt-4o-mini and LLAVA-1.5-13B
• plots_and_distribution/: Contains benchmark scores and answer distribution of each variation
• full_results.csv: Comprehensive results in csv format

Setup

To set up the project environment:

1. Ensure you have Python 3.10 installed on your system.

2. Clone this repository:

git clone https://github.com/jasonlim131/looksRdeceiving.git
cd looksRdeceiving

3. Create a virtual environment, e.g. "agreeable_venv":
4. Activate this environment in your terminal:

source /path/to/env/agreeable_venv/venv/bin/activate

5. Install dependencies

pip3.10 install -r requirements.txt

If you keep getting dependency errors, try:

• pip installing the missing packages
• uninstalling and reinstalling with specific version written in the requirements.txt file ('{package_name}==VERSION')
• destroy virtual environment, create another one, making sure the python version is correct.

Generating Experiments

To generate the experimental prompts:

1. For vMMLU:

python python-src/generate_vmmlu.py 

This will create each rendered prompts in default directory output_directory/vmmlu_{variation}_rendered/question{i}.png

2. For vSocialIQa:

python python-src/generate_vsocialiqa.py

This will create rendered prompts in:

• output_directory/vmmlu_centered_{variation}_rendered
• output_directory/vmmlu_{variation}_rendered

Evaluating Models

python3.10 evaluate_{model}_{task_format}.py

Analyzing Results

python3.10 calculate_bias_{task_format}.py

Models Tested

We evaluated agreeableness bias in the following models:

• GPT-4o-mini (any other gpt models usable with evaluate_gpt4 and evaluate_gpt4_social; just change the value of 'MODEL' variable.
• Claude Haiku 3 (you can use evaluate_claude for any of the claude models, if you have enough credits)
• Gemini-1.5-flash
• LLAVA-1.5-13B (vMMLU only)

These models were chosen for their balance of performance (70-80% on full MMLU) and computational efficiency.

Benchmarks

1. Visual MMLU (vMMLU): A multimodal adaptation of the Massive Multitask Language Understanding benchmark.
2. Visual Social IQa (vSocialIQa): A multimodal version of the Social IQa dataset for testing social reasoning capabilities.

Contributing

We encourage further research in the following areas:

• Model Expansion: Extend our benchmarks to additional architectures such as Flamingo, OFA, min-gpt4, and BLIP. Apply our analysis methodology to these models to broaden our understanding of agreeableness bias across different architectures.
• Mechanistic Analysis: Conduct in-depth investigations of open-source models exhibiting visual bias. This could begin with embedding projections to understand the overlap in representation.

We welcome contributions to this project! Please see our Contributing Guidelines for more information.

License

This project is licensed under the MIT License - see the LICENSE.txt file for details.