CausalSpatial

CausalSpatial is a visual question answering benchmark for evaluating object-centric causal spatial reasoning in vision-language models. Each question presents a 3D-rendered scene and asks the model to reason about physical outcomes — not just what is visible, but what would happen given a specific action or trajectory.

Dataset Structure

Each subset shares the same schema:

Field	Type	Description
id	string	Unique sample identifier
question	string	Multiple-choice question
answer	string	Correct option label (e.g. A, B)
image	image	Rendered scene image
not_sure	string	Option label corresponding to "Not sure"

Subsets

collision — 413 samples

A toy car is placed on a floor with objects nearby. The model must judge whether the car will collide with something if it moves forward, and if so, which object to remove to prevent the collision.

Note: Floor strip spacing encodes depth perspective.

compatibility — 99 samples

An object is suspended above a container. The model must judge whether the object can fit into the container if it falls freely.

Note: The falling object is positioned directly above the container.

occlusion — 189 samples

A car is parked in front of a cabinet. The model must judge whether an item inside the cabinet can be retrieved without being blocked, depending on the car's direction of movement.

physics — 311 samples

Ball trajectory prediction across three sports scenarios:

• Soccer: Will the ball enter the goal along the shown trajectory?
• Basketball: Will the ball go through the basket along the shown arc?
• Billiard: Will a ball reach a pocket given the cue ball's direction?

Usage

from datasets import load_dataset

# Load a specific subset
ds = load_dataset("Mwxinnn/CausalSpatial", name="collision")
ds = load_dataset("Mwxinnn/CausalSpatial", name="physics")
ds = load_dataset("Mwxinnn/CausalSpatial", name="compatibility")
ds = load_dataset("Mwxinnn/CausalSpatial", name="occlusion")

print(ds["train"][0])

License

MIT