Feed
Do vision models know that three is three?
Vision models learn a clean direction for '2 to 3' per image type, but rarely share it.
0.89
τ
Top local-global linearity correlation
Models with stronger local linearity escape fragmentation
How it works
Same concept, different looks
The number 3 appears as a digit, three dots or three fingers. Do vision embeddings encode one shared direction for it?
Measure direction consistency
Pair matched images across a concept change in 46 encoders. Check if the change vectors align within and across image types.
Fragmented linearity found
Directions are consistent within each image type but often fail to transfer. CLIP-like models transfer best.
From the paper