update

Author: Tianwei Zhao

README.md

@@ -1,145 +1,118 @@
-# RewardAnything GitHub Pages
+# Core Knowledge Deficits in Multi-Modal Language Models
 
-This directory contains the GitHub Pages website for the Core Knowledge Deficits in Multi-Modal Language Models project.
+**Official website for the ICML 2025 paper submission**
 
-## 🏗️ Structure
+🌐 **Website**: [https://williamium3000.github.io/core-knowledge](https://williamium3000.github.io/core-knowledge)  
+📄 **Paper**: [https://arxiv.org/abs/2410.10855](https://arxiv.org/abs/2410.10855)  
+🤗 **Dataset**: [https://huggingface.co/grow-ai-like-a-child](https://huggingface.co/grow-ai-like-a-child)
 
-```
-pages/
-├── _config.yml              # Jekyll configuration
-├── _layouts/
-│   └── default.html         # Main layout template
-├── index.html               # Homepage content
-├── assets/
-│   ├── images/              # Logo and image placeholders
-│   └── favicon.svg          # Site favicon
-├── Gemfile                  # Ruby dependencies
-├── setup.sh                 # Local setup script
-└── README.md               # This file
-```
+## 📖 About
 
-## 🚀 Automatic Deployment
+This repository contains the official website for our paper "Core Knowledge Deficits in Multi-Modal Language Models". The website presents our comprehensive evaluation of 230 multi-modal language models using the **CoreCognition** benchmark, which assesses 12 foundational cognitive concepts grounded in developmental cognitive science.
 
-1. **Changes are pushed** to the `main` branch in the `pages/` directory
-2. **Manual trigger** via GitHub Actions tab
+## 🔍 Key Findings
 
-The deployment is handled by the GitHub Actions workflow in `.github/workflows/deploy-pages.yml`.
+Our research reveals four critical shortcomings in state-of-the-art Multi-modal Large Language Models (MLLMs):
 
-## 🏠 Local Development
+1. **Core Knowledge Deficits**: MLLMs excel at higher-level abilities but struggle with lower-level cognitive abilities
+2. **Misaligned Dependency**: Core abilities show weak cross-stage correlations, lacking developmental scaffolding
+3. **Predictability**: Performance on core knowledge predicts higher-level abilities
+4. **Limited Scaling**: MLLMs show minimal scalability improvements on low-level abilities compared to high-level ones
 
-### Quick Setup
+## 🧠 CoreCognition Benchmark
 
-```bash
-# Navigate to pages directory
-cd pages
+The **CoreCognition** benchmark evaluates twelve foundational cognitive concepts:
+
+1. **Permanence** - Objects persist when not perceived
+2. **Continuity** - Objects remain unified across space and time
+3. **Boundary** - Transitions between objects
+4. **Spatiality** - Understanding Euclidean properties
+5. **Perceptual Constancy** - Appearance changes ≠ property changes
+6. **Intuitive Physics** - Laws of physical interaction
+7. **Perspective** - Seeing what others see
+8. **Hierarchy** - Inclusion/exclusion of objects and categories
+9. **Conservation** - Property invariances despite transformations
+10. **Tool Use** - Manipulating objects to achieve goals
+11. **Intentionality** - Understanding what others want
+12. **Mechanical Reasoning** - Inferring actions from system states
+
+## 🔬 Concept Hacking
 
-# Run setup script (macOS/Linux)
-chmod +x setup.sh
-./setup.sh
+We introduce **Concept Hacking**, a novel controlled evaluation method that systematically manipulates task-relevant features while preserving task-irrelevant conditions. This reveals that MLLMs fail to develop genuine core knowledge understanding and instead rely on shortcut learning as they scale.
+
+## 📊 Evaluation Scale
+
+- **230 MLLMs** evaluated across different model families and sizes
+- **11 different prompts** to ensure robust evaluation
+- **>26,000 total judgments** across all models and tasks
+- **2,530 image-question pairs** in the benchmark
+
+## 🏗️ Website Structure
 
-# Start development server
-bundle exec jekyll serve
 ```
+├── _config.yml              # Jekyll configuration
+├── _layouts/
+│   └── default.html         # Main layout template
+├── index.html               # Homepage with full paper content
+├── assets/
+│   ├── images/              # Paper figures and illustrations
+│   ├── growai.png          # Site favicon
+│   └── favicon.svg         # Backup favicon
+├── Gemfile                  # Ruby dependencies
+└── README.md               # This file
+```
+
+## 🚀 Local Development
 
-### Manual Setup
+To run the website locally:
 
 ```bash
-# Install Ruby dependencies
+# Install dependencies
 gem install jekyll bundler
 bundle install
 
 # Serve the site locally
 bundle exec jekyll serve --livereload
 ```
 
-Then visit: `http://localhost:4000/RewardAnything`
+Then visit: `http://localhost:4000/core-knowledge`
 
-## 📝 Configuration
+## 👥 Authors
 
-### GitHub Pages Settings
+**Yijiang Li¹**, **Qingying Gao²,§**, **Tianwei Zhao²,§**, **Bingyang Wang³,§**, **Haoran Sun²**, **Haiyun Lyu⁴**, **Robert D. Hawkins⁵**, **Nuno Vasconcelos¹**, **Tal Golan⁶**, **Dezhi Luo⁷,⁸,†**, **Hokin Deng⁹,†**
 
-1. Go to **Repository Settings** → **Pages**
-2. Source: **GitHub Actions**
-3. The workflow will handle the rest automatically
+¹University of California San Diego, ²Johns Hopkins University, ³Emory University, ⁴University of North Carolina at Chapel Hill, ⁵Stanford University, ⁶Ben-Gurion University of the Negev, ⁷University of Michigan, ⁸University College London, ⁹Carnegie Mellon University
 
-### Environment Variables
+§Equal Contribution, †Corresponding author
 
-The following are configured in `_config.yml`:
+## 📄 Citation
 
-- `github_username`: Your GitHub username
-- `paper_url`: Link to your arXiv paper
-- `huggingface_url`: Link to model weights
-- `pypi_url`: Link to PyPI package
+If you find this work useful in your research, please consider citing:
 
-## 🎨 Customization
-
-### Replacing Placeholder Images
-
-Replace the SVG placeholders in `assets/images/` with your actual logos:
-
-- `logo-placeholder.svg` → Navigation logo
-- `logo-placeholder-white.svg` → Footer logo (white version)  
-- `hero-logo-placeholder.svg` → Large hero section logo
-- `favicon.svg` → Browser favicon
-
-### Updating Content
-
-- **Homepage**: Edit `index.html`
-- **Navigation**: Modify `_layouts/default.html`
-- **Site settings**: Update `_config.yml`
-- **Styling**: Customize Tailwind classes in templates
-
-### Adding New Pages
-
-Create new `.html` or `.md` files with front matter:
-
-```yaml
----
-layout: default
-title: "Page Title"
-description: "Page description"
----
-
-Your content here...
+```bibtex
+@article{li2025core,
+    title={Core Knowledge Deficits in Multi-Modal Language Models}, 
+    author={Li, Yijiang and Gao, Qingying and Zhao, Tianwei and Wang, Bingyang and Sun, Haoran and Lyu, Haiyun and Luo, Dezhi and Deng, Hokin},
+    journal={arXiv preprint arXiv:2410.10855},
+    year={2025}
+}
 ```
 
-## 🔧 Troubleshooting
+## 📧 Contact
 
-### Local Development Issues
+For questions about the paper or dataset, please contact the corresponding authors:
+- Dezhi Luo: [dezhi@umich.edu](mailto:dezhi@umich.edu)
+- Hokin Deng: [hokindeng@cmu.edu](mailto:hokindeng@cmu.edu)
 
-```bash
-# Clean build files
-bundle exec jekyll clean
-
-# Rebuild dependencies
-bundle install --force
-
-# Verbose build for debugging
-bundle exec jekyll serve --verbose
-```
-
-### Deployment Issues
-
-1. Check **Actions** tab for build logs
-2. Ensure `pages/` directory changes are pushed to `main`
-3. Verify GitHub Pages settings are correct
+## 🔧 Technical Details
 
-## 📊 Performance
-
-The site is optimized for:
-- ✅ Mobile responsiveness
-- ✅ Fast loading (Tailwind CSS via CDN)
-- ✅ SEO optimization
-- ✅ Accessibility
-- ✅ Modern browsers
-
-## 🤝 Contributing
-
-When making changes:
-
-1. Test locally first: `bundle exec jekyll serve`
-2. Commit changes to `pages/` directory
-3. Push to `main` branch
-4. Automatic deployment will trigger
+The website is built with:
+- **Jekyll** for static site generation
+- **Tailwind CSS** for styling
+- **GitHub Pages** for hosting
+- **Responsive design** optimized for all devices
+- **SEO optimization** for better discoverability
 
 ---
+
+*This website presents the official results and findings from our comprehensive evaluation of multi-modal language models on core cognitive abilities.*

_config.yml

@@ -1,7 +1,7 @@
-title: Core Cognition"
+title: "Core Cognition"
 description: "Core Knowledge Deficits in Multi-Modal Language Models"
-url: "https://grow-ai-like-a-child.github.io"
-baseurl: "/CoreCognition"
+url: "https://williamium3000.github.io"
+baseurl: "/core-knowledge"
 
 # Build settings
 markdown: kramdown
@@ -30,7 +30,7 @@ exclude:
   - README.md
 
 # Social links
-github_username: grow-ai-like-a-child
+github_username: williamium3000
 paper_url: "https://arxiv.org/abs/2410.10855"
 huggingface_url: "https://huggingface.co/grow-ai-like-a-child"
 pypi_url: ""