DSReg: Provably Recovering Individual World Latents without Reconstruction
DSReg claims individual latent variables can be identified without a decoder, labels, or reconstruction loss.
The paper’s condition is “Structural Diversity”: each latent leaves a distinct dependency pattern in the observations. Under that condition, the authors say DSReg can recover latents up to signed permutation from a linearly identified representation such as LeJEPA. They frame the method as usable after training, reusing existing checkpoints with no loss versus joint training. The experiments cited cover synthetic settings, world-model probes, visual encoders, and external renderers.
HF Daily Papers' note
The paper’s condition is “Structural Diversity”: each latent leaves a distinct dependency pattern in the observations. Under that condition, the authors say DSReg can recover latents up to signed permutation from a linearly identified representation such as LeJEPA. They frame the method as usable after training, reusing existing checkpoints with no loss versus joint training. The experiments cited cover synthetic settings, world-model probes, visual encoders, and external renderers.
HF Daily Papers' note
score 5