SciDocBench - Evidence Verification: leaderboard
SciDocBench workflow-centred scientific document understanding: the evidence alignment and verification capability group over 496 evaluation instances from 124 questions across arXiv, bioRxiv and OpenReview papers. Higher is better. Source: arxiv.org/abs/2609.05141.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 78.16 | #22 |
| 2 | Claude Opus 5 | 74.47 | #16 |
| 3 | GPT-5.6 Terra | 66.28 | #40 |
| 4 | GPT-5.6 Luna | 65.15 | #79 |
| 5 | Claude Opus 4.8 | 58.63 | #39 |
| 6 | Claude Sonnet 4.6 | 54.25 | #94 |
| 7 | Kimi K2.5 | 53.05 | #152 |
| 8 | Claude Haiku 4.5 | 40.33 | #287 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=scidocbench-evidence-verification · How It Works · Data refreshed daily, snapshot 2026-11-05.