DocScope - Fact Extraction: leaderboard

DocScope (2026) evaluates verifiable reasoning over long, visually rich PDF documents: for each human-annotated question of its 730-question test split the model reads the whole document and returns evidence pages, evidence regions, supporting facts and an answer. This board is the percent of extracted facts that a Qwen3.6-Plus judge finds consistent with the gold evidence on the correctly retrieved pages; higher is better.

Metric: Fact consistency rate (%): share of extracted facts that a Qwen3.6-Plus judge marks consistent with the gold evidence on the correctly retrieved pages, micro-averaged over the answerable questions of the 730-question DocScope test split (human-annotated questions on long, visually rich PDF documents; the model receives the whole document and returns evidence pages, evidence regions, supporting facts and an answer), each stage scored only on the gold pages the model retrieved; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.777.4#85
2Gemma 4 31B73.6#382
3Claude Sonnet 4.672.1#173
4Gemini 3.1 Flash Lite70.7#371
5Gemini 3.1 Pro (Preview)68.7#99
6Qwen 3.6 Plus66.4#238
7Qwen 3.5 27B61.3#375
8GPT-5.459.4#135
9Qwen 3.5 397B A17B58.7#264
10Qwen 3 VL 235B A22B58.6#485
11Gemma 4 26B A4B53.1#590
12Qwen 3 VL 32B48.9#622

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=docscope-fact-extraction · How It Works · Data refreshed daily, snapshot 2026-11-21.