01 / ORIGINAL
MODEL SAYSNot trained
Train the model to reveal its answer.
SHORTCUT LEARNING, MADE TANGIBLE
Same shape.
Different answer.
A model can get every answer right for the wrong reason.
Train one. Change a single clue. See what breaks.
01 Learn a shortcut02 Change the world03 Teach a better clue
HOLD THE SHAPE STILL. CHANGE THE BACKDROP.
Ask a better question.
→shape
unchanged
unchanged
02 / INTERVENTION
MODEL SAYSNot trained
Only background pixels will change.
The two images begin identical. Train, then flip the backdrop to test which clue the model follows.
NOW TEST BEYOND ONE SPECIMEN
A good score can hide a bad rule.
The shifted test keeps every shape, position and texture fixed, and swaps only the backdrop. Four shape × backdrop groups cover both test environments. These are empirical scores on a toy task, not generalization guarantees.
CHANGE THE LESSON, NOT THE ANSWER
Let every shape live in both worlds.
Pair each training specimen with both backdrops. Retrain from zero using the same optimizer. The held-out images never enter training.
+Same shape. Same label.
Both backdrops.
Both backdrops.
256 training views · 128 original training tiles
REAL MODEL. DELIBERATELY SMALL WORLD.
The lesson is bigger
than the classifier.
A trained linear classifier.
768 normalized RGB features feed logistic regression. Full-batch gradient descent minimizes binary cross-entropy with L2 regularization. No prediction is manually assigned.
Change exactly one factor.
The counterfactual preserves the foreground pixels, geometry and texture. Only backdrop pixels change. The model receives pixels, never the label or backdrop metadata.
Accuracy is not certainty.
Model scores are uncalibrated. Balanced training can solve this toy shift; it does not prove robustness to new shapes, corruptions or real-world images.