DSReg: provably recovering individual latents without reconstruction, post hoc
ShahabBakht · x · 2026-10-09
Researchers propose DSReg, a simple regularization that can provably recover individual latent variables of the true world even without reconstruction objectives like JEPA.
- Key idea: exploit the sparse support structure of the Jacobian (e.g., moving an object doesn't change the whole image) to fully disentangle latents
- Goes beyond prior results like LeJEPA, which recover latents only up to linear mixing
- Works post hoc on already-pretrained models, no retraining needed
- Collaborators call the approach an elegant route to full disentanglement
Related event: DSReg Provably Recovers Latent Variables Without Reconstruction(3 posts)→
More from Research
- Ditto-Bench stress-tests GPT-6 Astra for robot control: simple tasks shine, physics stalls — anand_bhattad · 2026-10-09
- LLMs can access truncated context info; bottleneck is elicitation, not storage — Sauers_ · 2026-10-09
- 16 VPD weight edits boost model accuracy 10x, revealing the attention head that suppresses introspection — Sauers_ · 2026-10-09
- What learning from experience looks like: an agent learns to value draws in chess — anirudhg9119 · 2026-10-09
- One attention head drives sandbagging-like introspection in Qwen3-1.7B; ablating it helps — Sauers_ · 2026-10-09
- How models introspect to recover deleted CoT tokens — and why some do the opposite — Sauers_ · 2026-10-09