CiteVQA - Multi-Doc (N-Gold): 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 covers questions whose evidence is spread over several documents, counting a question as correct only when a judge scores the answer at least 4 of 5 and the cited evidence is either judged relevant (at least 4 of 5) or recovers at least 60 percent of the key evidence; higher is better.

Metric: Strict attributed accuracy (%): share of questions with answer score at least 4 and either relevance at least 4 or evidence recall at least 0.6, on the multi-document questions with several gold documents (22.3 percent of the set), 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)High71.6#102
2Gemini 3 Flash (Preview)High60.5#156
3GPT-5.4xHigh55.1#138
4Gemini 2.5 ProThinking, 32K39.2#280
5Seed 2.0 Pro36.2#226
6GPT-5.2xHigh30.5#190
7Gemma 4 31B17.8#390
8Qwen 3 VL 235B A22B Instruct17.8#541
9Qwen 3.5 397B A17B15.2#270
10Kimi K2.514.2#272
11Qwen 3 VL 32B Instruct14#607
12GLM-5V Turbo13#365
13Qwen 3.5 27B11.6#380
14Qwen 3.5 9B10.7#838
15Qwen 3.6 Plus9.8#245

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=citevqa-multi-doc-n-gold · How It Works · Data refreshed daily, snapshot 2026-11-22.