Atlas

Benchmarks

← All benchmarks

VQAv2

Multimodal · 2016-12-02

VQA v2 asks open-ended questions about COCO images, balanced so each question pairs with similar images that lead to different answers. This row is the test-dev split of 107,394 questions, scored on the official server with the standard VQA accuracy against ten human answers; higher is better.

Top models (higher is better)

ModelScore
NVLM-D-72B85.4
Gemini 1.5 Pro (May 2024)80.2
Gemini 1.5 Flash (May 2024 report)80.1
Gemini 1.5 Pro (February 2024 report)73.2
Gemini Nano 267.5
Gemini Nano 162.7
Loading Atlas data…