What We Don't C: Manifold Disentanglement for Structured Discovery

ICLR 2026 Conference Submission18374 Authors

Keywords: supervised representation learning, disentanglement, flow matching

TL;DR: We use latent flow matching to disentangle latent subspaces from VAE latent variables, recovering meaningful residual representations.

Abstract: Accessing information in learned representations is critical for annotation, discovery, and data filtering in disciplines where high-dimensional datasets are common. We introduce What We Don't C, a novel approach based on latent flow matching that disentangles latent subspaces by explicitly removing information included in conditional guidance resulting in meaningful residual representations. This allows factors of variation which have not already been captured in conditioning to become more readily available. We show how guidance in the flow path necessarily represses the information from the guiding, conditioning variables. Our results highlight this approach as a simple yet powerful mechanism for analyzing, controlling, and repurposing latent representations, providing a pathway toward using generative models to explore what we don't capture, consider, or catalog.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 18374