DSReg paper: provably recovering individual world latents without reconstruction
randall_balestr · x · 2026-10-09
- Yujia Zheng, Randall Balestriero, Bernhard Schölkopf et al. release the DSReg paper on arXiv, tackling a core question: JEPA-style methods without reconstruction can only identify the world's latent state up to a linear transformation — can individual latents be provably recovered?
- Yes: under a "Structural Diversity" condition (distinct latents leave distinct dependency footprints on observations), DSReg (Dependency-Sparsity Regularization) recovers individual world latents up to signed permutation — no reconstruction, no decoder, no labels.
- The method applies post hoc to any linearly identified pretrained representation (e.g., LeJEPA checkpoints), with no loss versus joint training, yielding the first fully identifiable JEPA that recovers every world latent.
- Structural Diversity is strictly weaker than all structural conditions of prior identifiable latent variable models; validated on synthetic regimes, world model probes, and visual encoders.
Related event: DSReg Provably Recovers Latent Variables Without Reconstruction(3 posts)→
More from Research
- Test-Time Structure-Space Search lifts antibody design success from 16% to 78% — anshulkundaje · 2026-10-09
- aimotive driving dataset: train labels come from a tracker that sees the future — RexDouglass · 2026-10-09
- Yale-led study finds symbolic structure inside LLM representations — tallinzen · 2026-10-09
- Discrete Diffusion Meetup at COLM: Oct 8, Grand Ballroom — yuntiandeng · 2026-10-09
- Astra 机器人控制实测:Johns Hopkins 深度评测其运动与控制理解 — jmin__cho · 2026-10-09
- Tavily's 93% live-retrieval agent study challenged: prompts were domain-anchored — edwin · 2026-10-09