The Problem
Transforming ML papers into runnable code repositories requires coordinating multiple LLM agents across planning, analysis, and generation stages. This repo attempts that automation but inherits the typical risk: code-generation paths are untested, and there’s no automated gate to catch pipeline breaks when agents interact or when prompts shift.
What This Does
Paper2Code implements a three-stage pipeline across codes/: Planning (codes/1planning.py, codes/1planningllm.py) Analysis (codes/2analyzing.py, codes/2analyzingllm.py) Code generation (codes/3coding.py, codes/3codingllm.py, codes/3.1codingsh.py, codes/4debugging.py)
Prompt files data/prompts/refbased.txt and data/prompts/reffree.txt steer agent behavior. The system supports both OpenAI API and vLLM-backed open-source models (deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct default).