SIN-Bench - SIN-Find: leaderboard
Evidence discovery in SIN-Bench (Ren et al., 2026): given a scientific paper as interleaved text and figures and a question, the model must return the ordered chain of figure and text anchors that supports the answer. The score averages evidence matching judged by Qwen3-8B, relevance F1 and Kendall-tau order similarity against the reference chain, in percent; higher is better.
Metric: Evidence Discovery Score (%; mean of evidence matching, relevance F1 and Kendall-tau logic). Source: arxiv.org. Saturation forecast: Around 2029. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 (Thinking) | 46 | #244 (Claude Sonnet 4.5) |
| 2 | Gemini 3 Pro (Preview) | 39.9 | #111 |
| 3 | GPT-5 | 37.8 | #163 |
| 4 | Grok 4 | 37.8 | #306 |
| 5 | Qwen 3 VL 30B A3B (Thinking) | 35.2 | #710 |
| 6 | Gemini 2.5 Pro | 34.6 | #262 |
| 7 | Qwen 3 VL 8B (Thinking) | 34.6 | #887 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sin-bench-sin-find · How It Works · Data refreshed daily, snapshot 2026-11-12.