original paper
PaperBench: Evaluating AI’s Ability to Replicate AI Research (v1)
OpenAI · Published 2025-04-02
Evidence from this source
| Type | Evidence |
|---|---|
| Result | PaperBench · o3-mini-high / BasicAgentSource detailsSection 5.2, Table 4 |
| Result | PaperBench · GPT-4o / BasicAgentSource detailsSection 5.2, Table 4 |
| Result | PaperBench · Gemini 2.0 Flash / BasicAgentSource detailsSection 5.2, Table 4 |
| Result | PaperBench · o1-high / BasicAgentSource detailsSection 5.2, Table 4 |
| Result | PaperBench · Claude 3.5 Sonnet (New) / BasicAgentSource detailsSection 5.2, Table 4 |
| Result | PaperBench · o3-mini-high / IterativeAgentSource detailsSection 5.3, Table 5 |
| Result | PaperBench · Claude 3.5 Sonnet (New) / IterativeAgentSource detailsSection 5.3, Table 5 |
| Result | PaperBench · o1-high / IterativeAgentSource detailsSection 5.3, Table 5 |
| Result | PaperBench · o1-high / IterativeAgent, 36 hSource detailsSection 5.3, Table 5 |
| Benchmark | PaperBenchSource detailsSections 2–3 and Sections 5.2–5.3; Tables 4–5 |
| Paper / report | PaperBench: Evaluating AI’s Ability to Replicate AI ResearchSource detailsPrimary publication source |
Document tracking details
- Source ID
- src-paperbench
- Source type
- Primary
- Original URL
- https://arxiv.org/html/2504.01848v1
- Last updated by publisher
- Not reported
- Access status
- available
- Retrieved
- 2026-09-26T04:47:27Z
- Document fingerprint
- 1704aec76a9740f8f4af9e47a9008d79aaddc17a0e16f72d3eccc39db568d27d
Based on original bytes. Used to detect changes to the source.
Availability checks
- 2026-09-26T04:47:27Z · success: Original source inspected by Codex; extraction check, not experimental replication.