PaperBench: Evaluating AI's Ability to Replicate AI Research
This addresses the problem of assessing AI's engineering capabilities for AI researchers, providing a scalable evaluation framework, though it is incremental in benchmarking methodology.
The authors introduced PaperBench, a benchmark to evaluate AI agents' ability to replicate state-of-the-art AI research papers from scratch, including understanding contributions, coding, and experiments. They found that the best-performing AI agent achieved an average replication score of 21.0%, and models did not outperform human ML PhDs on a subset of tasks.
We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism. To enable scalable evaluation, we also develop an LLM-based judge to automatically grade replication attempts against rubrics, and assess our judge's performance by creating a separate benchmark for judges. We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline. We open-source our code (https://github.com/openai/preparedness) to facilitate future research in understanding the AI engineering capabilities of AI agents.