AI
DSReg: Provably Recovering Individual World Latents without Reconstruction
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 06/10/2026, 20:00:00 UTC
Briefing created: 11/11/2026, 20:42:17 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 19:18:24 UTC.
Briefings
A new method called DSReg recovers individual world latents without reconstruction, decoder, or labels. It applies to any linearly identified representation and establishes the first fully identifiable JEPA that recovers every world latent. The key condition is Structural Diversity, which is strictly weaker than all structural conditions of prior identifiable latent variable models. This approach preserves dense prediction while improving individual-latent recovery and downstream use with scales.
Key points
- DSReg recovers individual world latents without reconstruction, decoder, or labels by applying Structural Diversity as a condition on dependency footprints.
- The method applies post hoc to any linearly identified representation, reusing trained checkpoints at no loss over joint training.
Context
DSReg closes the gap between joint-embedding predictive architectures (JEPAs) that identify latent states only up to a linear transformation and methods that anchor latents to observations through reconstruction. By proving that under Structural Diversity, DSReg recovers individual world latents up to signed permutation, it establishes the first fully identifiable JEPA that recovers every world latent, offering a post hoc solution for any linearly identified representation.
Limitations
The evidence is limited to synthetic regimes, world model probes, learned visual encoders, and external renderers. The method applies to any linearly identified representation, but the specific conditions and constraints of the evidence are not fully detailed in the provided text.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 11/11/2026, 20:42:17 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeify4p252nytsdq2zasv5wtucpgnery73dcr3sxenzloianz64flfm