AutoReproduce: AI Framework Automates Research Paper Reproduction

Researchers have introduced AutoReproduce, a multi-agent framework that autonomously reproduces experimental code from research papers. The system uses a novel 'paper lineage' algorithm to mine implicit knowledge from cited literature, enabling end-to-end code reproduction. A sampling-based unit testing strategy ensures executability. The team also developed a benchmark called AutoBench with verified implementations and metrics for evaluating reproduction and execution fidelity. Evaluations on PaperBench and AutoBench show AutoReproduce consistently outperforms existing methods. The work aims to address the labor-intensive nature of reproducing increasingly complex research methods.