On a 2D Gaussian bump task, disentangled latents alone fail to generalize out-of-distribution, but forcing latents to be rendered into pixel space restores compositional generalization.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Compositional Generalization via Forced Rendering of Disentangled Latents
On a 2D Gaussian bump task, disentangled latents alone fail to generalize out-of-distribution, but forcing latents to be rendered into pixel space restores compositional generalization.