ScholScan (Image Input) - Inference and Conclusions: leaderboard
Metric: Flaw-detection score (0-1, scaled to 0-100): per question, zero unless the model finds the annotated scientific error, otherwise the geometric mean of evidence-location Dice and reasoning-chain prefix match, times a penalty for unrelated errors listed, as extracted by a GPT-4.1 evaluator, averaged over the inference and conclusions questions of ScholScan's 1,800 expert-reviewed questions that ask a model to scan a whole academic paper (ICLR 2024-2025 and Nature Communications, 715 papers) for a planted or reviewer-reported scientific flaw without being told where to look, with the paper given as page images; higher is better. Source: arxiv.org. Saturation forecast: Around July 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 25.3 | #91 |
| 2 | Seed-1.6 | 17.9 | #257 |
| 3 | Gemini 2.5 Pro | 14.7 | #145 |
| 4 | Doubao-Seed-1.6 (Thinking) | 12.2 | |
| 5 | Llama 4 Maverick | 6.7 | #451 |
| 6 | Mistral Small 3.1 | 2.2 | #600 |
| 7 | Grok 4 | 1.9 | #169 |
| 8 | Gemma 3 27B | 1.3 | #596 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=scholscan-image-input-inference-and-conclusions · How It Works · Data refreshed daily, snapshot 2026-10-11.