SplitJEPA: Disentangling Invariant-Variant Structure Without Reconstruction
udmrzn · x · 2026-10-11
New work SplitJEPA explores identifying invariant-variant structure in latent representations without reconstructing observations. While DSReg recovers every latent of the true world model, the authors show two blocks often suffice for reasoning and collaboration — an invariant part (content, objects) and a variant part (noise, backgrounds) — both learned without reconstruction.
Related event: SplitJEPA Separates Invariant and Variable Factors Without Reconstruction(2 posts)→
More from Research
- 3 papers on adversarial AI attacks vs flight control, CCTV, 6G at IEEE conf — WeisiG · 2026-10-11
- Are ML Papers Getting Too Wordy? Reddit Debates 40-Page Bloat — NeighborhoodFatCat · 2026-10-11
- Anders Sandberg's Claude-made video shines with Tromp's Lambda Diagrams notation — anderssandberg · 2026-10-11
- Lean Co-creator Jeremy Avigad on AI, Verification, and the Future of Mathematics — EchoShao8899 · 2026-10-11
- Reader finds 10+ results in OpenAI's math repo whose headline claims fall outside their Lean checks — MysteriousAvocado580 · 2026-10-11
- NeurIPS 2026 position paper argues LLMs lack explicit cognitive control for true autonomy — AndrewLampinen · 2026-10-11