An automated system that discovers, analyzes, and organizes the most promising Large Language Model (LLM) research papers daily.
- Automated Discovery: Scans arXiv and other sources for new LLM papers
- Intelligent Scoring: Uses Google Gemma to evaluate paper significance
- Smart Organization: Categorizes papers into subject-specific folders
- Daily Updates: Runs automatically or on-demand via GitHub Actions
- PR Integration: Creates pull requests with detailed annotations
Papers are automatically organized into:
- Architectures: Model designs, attention mechanisms, transformers
- Training: Learning methods, optimization, fine-tuning
- Multimodal: Vision, audio, cross-modal capabilities
- Reasoning: Logic, inference, problem-solving
- Alignment: Safety, ethics, human feedback
- Efficiency: Compression, quantization, acceleration
- Evaluation: Benchmarks, metrics, testing
- Applications: Real-world deployment, tools
- Theoretical: Mathematical analysis, complexity
- Agents: Autonomous systems, planning
- Generation: Text synthesis, creativity
- Knowledge: Retrieval, memory, factual reasoning
days_back: Number of days to look back for papers (default: 1)min_score: Minimum significance score threshold (default: 70)
Set these in your GitHub repository settings:
GITHUB_TOKEN: Automatically provided by GitHub Actions
The workflow runs automatically at 9 AM UTC daily.
- Go to Actions tab in your GitHub repository
- Select "LLM Paper Curation" workflow
- Click "Run workflow"
- Optionally adjust parameters
# Install dependencies
pip install -r requirements.txt
# Set environment variables
export DAYS_BACK=3
export MIN_SCORE=75
# Run the curation script
python .github/scripts/curate_papers.pypapers/
├── architectures/
│ ├── README.md
│ ├── metadata.json
│ └── [paper-id].md
├── training/
│ ├── README.md
│ ├── metadata.json
│ └── [paper-id].md
└── [other-categories]/
Each paper gets:
- Individual markdown file with full analysis
- Entry in category metadata.json
- Link in category README.md
- Paper Discovery: Fetches from arXiv API and RSS feeds
- Deduplication: Removes duplicates based on URL and content hash
- Analysis: Queries Gemma to evaluate and summarize findings
- Classification: Categorizes into subject areas using Gemma
- Filtering: Only includes papers above significance threshold
- Organization: Creates structured file system with metadata
- PR Creation: Generates pull request with detailed summary
Edit the subject_categories dictionary in curate_papers.py:
self.subject_categories = {
"your_category": ["keyword1", "keyword2", "keyword3"],
# ... existing categories
}Extend the fetch_papers_with_rss() method with additional RSS feeds or APIs.
The system includes multiple quality checks:
- Duplicate detection across sources
- Minimum score thresholds
- Content validation
- Error handling and logging
- Manual review via pull requests
No papers found:
- Check if arXiv is accessible
- Verify date range settings
- Lower the minimum score threshold
Workflow fails:
- Check GitHub Actions logs
- Verify repository permissions
- Ensure all required files are present
Categories not working:
- Review keyword lists in subject_categories
- Check paper classification logic
- Verify directory creation permissions
- Fork the repository
- Create a feature branch
- Make your changes
- Test locally
- Submit a pull request