PaperBench (OpenAI): replicating 20 papers from the ICML 2024 conference, the best AI agent (Claude 3.5 Sonnet) scores 21.0% on average; models do not yet outperform human experts
translated by AI, reviewed by the editors
· Science
- technology
- Recursive self-improvement of AI
- source
- arxiv.org ↗
¶ c927jx6v4