ProjectionBench: leaderboard
ProjectionBench evaluates large language models on scientific discovery and reasoning using recent-paper topics and research questions, with technical details revealed progressively across stages. The inputs are text. The pinned metric is AUC, measuring alignment of generated hypotheses with the original paper’s conclusions via semantic similarity of atomic claims. Source: BenchmarkList’s attributed result table.
Metric: AUC. Source: benchmarklist.com. Saturation forecast: Around June 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 1.65 | #83 |
| 2 | Gemini 3.1 Pro (Preview) | 1.62 | #64 |
| 3 | GPT-5 | 1.57 | #105 |
| 4 | Gemini 2.5 Pro | 1.55 | #159 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=projectionbench · How It Works · Data refreshed daily, snapshot 2026-11-05.