CiteVQA - Evidence Relevance: 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 how well the cited evidence collectively supports the answer, judged on a 0 to 5 scale and multiplied by 20, averaged over all questions; higher is better.

Metric: Holistic relevance of the cited evidence to the answer, judged on a 0-5 scale and multiplied by 20 (0-100), 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)High83.6#99
2Gemini 3 Flash (Preview)High75.7#152
3GPT-5.4xHigh67.5#135
4Gemini 2.5 ProThinking, 32K59.8#276
5GPT-5.2xHigh56.6#185
6Seed 2.0 Pro54.9#222
7Qwen 3 VL 235B A22B Instruct35.3#532
8Gemma 4 31B35#382
9Qwen 3 VL 32B Instruct30.5#606
10GLM-5V Turbo29.2#360
11Kimi K2.526.8#266
12Qwen 3.5 27B25.3#375
13Qwen 3.6 Plus25#238
14Qwen 3.5 397B A17B24.6#264
15Qwen 3.5 122B A10B19#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-relevance · How It Works · Data refreshed daily, snapshot 2026-11-21.