SIN-Bench: leaderboard
Evidence-grounded reading of long scientific papers given as interleaved text and figures (SIN-Bench, Ren et al., 2026): models find evidence chains, verify whether evidence supports a claim, answer questions with cited evidence and write evidence-anchored summaries, under a 'No Evidence, No Score' rule. Each task score averages its metrics (evidence matching judged by Qwen3-8B, relevance F1, Kendall-tau order similarity, and answer or verification accuracy), and the score is the mean of the four task scores in percent; higher is better.
Metric: Average Overall Score (%; mean of the SIN-Find, SIN-Verify, SIN-QA and SIN-Summary task scores). Source: arxiv.org. Saturation forecast: Around July 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 56.6 | #111 |
| 2 | Claude Sonnet 4.5 (Thinking) | 54.9 | #244 (Claude Sonnet 4.5) |
| 3 | GPT-5 | 54.4 | #163 |
| 4 | Gemini 2.5 Pro | 51 | #262 |
| 5 | Grok 4 | 49.5 | #306 |
| 6 | Qwen 3 VL 8B (Thinking) | 45.2 | #887 |
| 7 | Qwen 3 VL 30B A3B (Thinking) | 44.8 | #710 |
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 · How It Works · Data refreshed daily, snapshot 2026-11-12.