OpenAI Introduces PaperBench for AI Research Replication

OpenAI has launched PaperBench, a new benchmark aimed at evaluating the ability of AI agents to replicate state-of-the-art AI research. This initiative is part of OpenAI's Preparedness Framework and involves replicating 20 ICML 2024 Spotlight and Oral papers from scratch. The process includes understanding the paper contributions, developing a codebase, and successfully executing experiments.

To ensure objective evaluation, PaperBench uses rubrics that break down each replication task into smaller, gradable sub-tasks. These rubrics are co-developed with the authors of each ICML paper to ensure accuracy and realism. An LLM-based judge has been developed to automatically grade replication attempts against these rubrics, and its performance is assessed through a separate benchmark for judges.

The benchmark has been tested with several frontier models, with the best-performing agent, Claude 3.5 Sonnet, achieving an average replication score of 21.0%. Despite these advancements, human experts still outperform AI models in this task. OpenAI has open-sourced the code to facilitate further research into AI engineering capabilities.