Common-O

measuring multimodal reasoning across scenes

Common-O, inspired by cognitive tests for humans, probes multimodal LLMs' ability to reason across scenes by asking "what’s in common?"

Common-O is comprised of household objects:

We have two subsets: Common-O (3 - 8 objects) and Common-O Complex (8 - 16 objects).

Multimodal LLMs excel at single image perception, but struggle with multi-scene reasoning

Evaluating a Multimodal LLM on Common-O

import datasets

# get a sample
common_o = datasets.load_dataset("facebook/Common-O")["main"]
# common_o_complex = datasets.load_dataset("facebook/Common-O")["complex"]
x = common_o[3]

output: str = model(x["image_1"], x["image_2"], x["question"])

check_answer(output, x["answer"])

To check the answer, we use an exact match criteria:

import re

def check_answer(generation: str, ground_truth: str) -> bool:
    """
    Args:
        generation: model response, expected to contain "Answer: ..."
        ground_truth: comma-separated string of correct answers

    Returns: bool, whether the prediction matches the ground truth
    """
    preds = generation.split("\n")[-1]
    preds = re.sub("Answer:", "", preds)
    preds = preds.split(",")
    preds = [p.strip() for p in preds]
    preds = sorted(preds, key=lambda x: x[0])

    # split into a list
    ground_truth_list = [a.strip() for a in ground_truth.split(",")]
    ground_truth_list = sorted(ground_truth_list)
    return preds == ground_truth_list

Some models have specific formatting outputs for their answers, e.g. \boxed{A} or Answer: A. We recommend checking a few responses as you may notice slight variations based on this. This public set also has slight variations with the set used in the original paper, so while the measured capabilities are identical do not expect an exact replication of accuracy figures.

If you'd like to use a single image model, here's a handy function to turn image_1 and image_2 into a single split image:

from PIL import Image

def concat_images_horizontal(
        image1: Image.Image, image2: Image.Image, include_space: bool=True, space_width: int=20, fill_color: tuple=(0, 0, 0)
        )  -> Image.Image:
    # from https://note.nkmk.me/en/python-pillow-concat-images/
    if not include_space:
        dst = Image.new("RGB", (image1.width + image2.width, image1.height))
        dst.paste(image1, (0, 0))
        dst.paste(image2, (image1.width, 0))
    else:
        total_width = image1.width + space_width + image2.width
        max_height = max(image1.height, image2.height)

        dst = Image.new("RGB", (total_width, max_height), color=fill_color)
        dst.paste(image1, (0, (max_height - image1.height) // 2))
        dst.paste(image2, (image1.width + space_width, (max_height - image2.height) // 2))
    return dst

For more details about Common-O see the

• dataset card
• ArXiv Paper

Please note we fixed an upload issue as of March 3, 2026, so please consider using the latest version if you cached an earlier version. Thank you to Milad Afshari for catching this!

Cite:

@inproceedings{Ross2025what0s,
  title  = {What’s in Common? Multimodal Models Hallucinate When Reasoning Across Scenes},
  author = {Candace Ross and Florian Bordes and Adina Williams and Polina Kirichenko and Mark Ibrahim},
  year   = {2025},
  url    = {https://openreview.net/attachment?id=d0F0N0cu4n&name=supplementary_material}
}