SynthDocBench - Complex Reasoning: leaderboard
The complex multi-hop subset of SynthDocBench (ServiceNow, 2026): accuracy in percent on 597 questions that compose two to four pieces of evidence across long synthetic reports read as page images, judged by GPT-5 against deterministic references. Higher is better.
Metric: Accuracy (%) on multi-hop complex-reasoning questions: GPT-5 judge scores each answer 0 to 10 against a deterministic reference and an answer counts when the score is at least 6; pages rasterized at 144 DPI in 5-page strips, vision-only input, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 78.9 | #94 |
| 2 | Qwen 3 VL 235B A22B | 61.1 | #479 |
| 3 | GPT-5.4 | 45.7 | #123 |
| 4 | InternVL3-78B | 39.7 | #738 |
| 5 | GPT-4o | 36 | #741 |
| 6 | Claude Sonnet 4.5 | 33.7 | #250 |
| 7 | Qwen 2.5 VL 7B Instruct | 1.2 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=synthdocbench-complex-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-13.