penguin-G/VisFactor

Human Cognitive Benchmarks Reveal Foundational Visual Gaps in MLLMs

👁️Overview

This repository contains the official dataset of VisFactor, a novel benchmark derived from the Factor-Referenced Cognitive Test (FRCT) that digitizes 20 vision-centric subtests from established cognitive psychology assessments. Our work systematically investigates the gap between human visual cognition and state-of-the-art Multimodal Large Language Models (MLLMs).

🎯 Key Features

• Comprehensive Evaluation: 4 core domains of human visual cognition
  • Visualization and Spatial Processing: Mental rotation, spatial relations
  • Perceptual and Closure: Figure-ground discrimination, pattern completion
  • Memory: Visual working memory, recognition tasks
  • Reasoning: Abstract visual reasoning, analogical thinking
• Extensive Model Coverage: 39 frontier MLLMs from GPT, Gemini, Claude, LLaMA, Qwen, and SEED families
• Rigorous Assessment: Based on well-established psychometric tests (FRCT)

🚀 Quick Start

Note: This evaluation framework is built on VLMEvalKit. For detailed information beyond this guide, please refer to their repository. We extend our sincere gratitude to the VLMEvalKit team for their excellent work.

Installation

git clone https://anonymous.4open.science/r/VisFactor-NeurIPS-ED-2026.git
cd VisFactor
pip install -r requirements.txt

Preparation

1. Download the VisFactor dataset

   mkdir -p ~/LMUData
   cd ~/LMUData
   # Download files will be placed here
   

   Place VisFactor.tsv and VisFactor_CoT.tsv in this directory.

2. Configure API credentials

   cd VisFactor/
   vim .env  # or use any text editor
   

   Example .env configuration:

   # OpenAI
   OPENAI_API_KEY=your_openai_key
   OPENAI_API_BASE=https://api.openai.com/v1
   
   # Google
   GOOGLE_API_KEY=your_google_key
   
   # Other Services
   STEPAI_API_KEY=your_stepai_key
   REKA_API_KEY=your_reka_key
   GLMV_API_KEY=your_glmv_key
   SENSENOVA_API_KEY=your_sensenova_key
   MOONSHOT_API_KEY=your_kimi_key
   DOUBAO_VL_KEY=your_doubao_key
   
   # Hunyuan-Vision
   HUNYUAN_SECRET_KEY=your_hunyuan_key
   HUNYUAN_SECRET_ID=your_hunyuan_id
   
   # Deployment Services
   CW_API_BASE=your_congwang_base
   CW_API_KEY=your_congwang_key
   LMDEPLOY_API_BASE=your_lmdeploy_base
   

Evaluation

Standard Evaluation

python3 run.py --data VisFactor --model GeminiPro2-5 --verbose

Chain-of-Thought (CoT) Evaluation

python3 run.py --data VisFactor_CoT --model GeminiPro2-5 --verbose

Command Arguments

Argument	Type	Default	Description
--model	list[str]	required	VLM names supported in VLMEvalKit (see supported_VLM in vlmeval/config.py)
--mode	str	'all'	Evaluation mode: 'all' (inference + evaluation) or 'infer' (inference only)
--api-nproc	int	4	Number of threads for API requests
--work-dir	str	'.'	Directory to save evaluation results
--reuse	flag	False	Use previously generated results if available

⚙️Generate testcases

We also provide a script to automatically generate some test cases, including CF1-3, CS1-3, MA1, S1-2, SS3, VZ1-2.

First, prepare some images:

mkdir visfactor/Collected_Figures

Place your images in this folder, then run the script to generate new questions:

cd visfactor
python3 generate_images.py