CiteVQA - Evidence Recall: leaderboard

CiteVQA (2026) asks multimodal models to answer questions over long PDFs, 1,897 questions on 711 documents in seven domains, and to cite the exact page elements that support each answer as bounding boxes. This board is the recall of the key evidence elements, a cited box matching one at IoU 0.5 or more, in percent over all questions; higher is better.

Metric: Recall (%) of the crucial evidence elements matched by a cited box at IoU 0.5, over all 1,897 questions, the model answers from the document pages and cites element-level bounding boxes, unified prompt, temperature 1.0, at most 4,096 output tokens, Qwen3-VL-235B-A22B judge; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 20 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)High66#99
2Gemini 3 Flash (Preview)High45.4#152
3GPT-5.4xHigh31#135
4Seed 2.0 Pro28.5#222
5Gemini 2.5 ProThinking, 32K27.4#276
6GPT-5.2xHigh18.2#185
7GLM-5V Turbo14.9#360
8Gemma 4 31B11.6#382
9Qwen 3 VL 235B A22B Instruct11.3#532
10Qwen 3.6 Plus7.7#238
11Qwen 3 VL 32B Instruct6.6#606
12Kimi K2.56.2#266
13Qwen 3.5 397B A17B5.4#264
14Qwen 3.5 27B5.3#375
15Qwen 3.5 122B A10B3.9#346

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

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