DocHop: leaderboard
Out-of-domain multi-hop reasoning over information-dense documents and charts: value retrieval, counting, numeric reasoning, ranking, hypotheticals and fact checking; overall accuracy. The human and random-guessing references are not published. Higher is better. Source: arxiv.org/abs/2609.02059.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around June 2028. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Thinking) | 40.6 | #159 (Gemini 2.5 Pro) |
| 2 | GPT-5.2 (2025-12-11) (Non-reasoning) | 40.36 | |
| 3 | Gemini 2.5 Flash (Thinking) | 32.02 | #255 (Gemini 2.5 Flash) |
| 4 | GPT-5 Mini | 28.25 | #193 |
| 5 | Gemini 2.5 Flash | 24.88 | #255 |
| 6 | Qwen 2.5 VL 72B Instruct | 24.4 | #383 |
| 7 | Qwen 3 VL 8B Instruct | 24.16 | #419 |
| 8 | Qwen 2.5 VL 32B Instruct | 21.79 | #463 |
| 9 | Qwen 2.5 VL 7B Instruct | 19.05 | #650 |
| 10 | Claude Sonnet 4.5 | 17.94 | #148 |
| 11 | Claude Haiku 4.5 | 16.35 | #287 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=dochop · How It Works · Data refreshed daily, snapshot 2026-11-05.