Paper2Code is a multi-agent LLM system that automates code generation from scientific papers in machine learning.
The tool addresses the challenge of translating research papers into working code implementations. It operates through a three-stage pipeline where specialized agents handle planning, analysis, and code generation sequentially. The system accepts papers in multiple formats—LaTeX source, PDF, or JSON—and produces a complete code repository as output. It supports both commercial APIs like OpenAI's o3-mini and open-source models deployed via vLLM, with the default being DeepSeek-Coder-V2-Lite-Instruct.
Developers should consider this tool if they need to rapidly prototype implementations from academic papers or validate research reproducibility. It suits machine learning researchers and practitioners who want to bridge the gap between paper descriptions and executable code without manual implementation. The project includes benchmark datasets for evaluation and provides model-based assessment of generated repositories, making it useful for those studying code generation quality.
The project maintains active development with structured documentation covering quick-start examples, detailed setup instructions for different deployment scenarios, and clear guidance on environment configuration. The codebase includes example workflows using well-known papers like "Attention Is All You Need" to demonstrate functionality. The tool provides transparent cost estimates for API-based usage and offers flexibility in model selection for users preferring open-source alternatives. Documentation addresses common installation challenges and provides fallback references to upstream dependencies.
