MI-MMSearch

Multi-image deep-research VQA benchmark: 339 questions, each with 1-3 images (98 / 144 / 97 questions with 1 / 2 / 3 images). Answering typically requires identifying what the images show and then multi-hop web research.

Columns

column	description
index	question id
question	question text; refers to the images as "Image 1", "Image 2", "Image 3"
answer	ground-truth answer
num_images	number of images (1-3)
image, img_2, img_3	base64-encoded image files (JPEG or PNG; RGB, at most 1024x1024 pixels); null when absent
image_url, img_2_url, img_3_url	public source URL of each image (mostly Wikimedia Commons); null when absent

Usage

import base64, io
from datasets import load_dataset
from PIL import Image

ds = load_dataset("kartiknarayan/MI-MMSearch", split="test")
row = ds[0]
img = Image.open(io.BytesIO(base64.b64decode(row["image"])))

Evaluation code (direct and agentic inference on VLMEvalKit, LLM-judge scoring): https://github.com/Kartik-3004/MI-MMSearch