H-JEPA: First End-to-End Learned Hierarchical World Model for Long-Horizon Visual Planning
randall_balestr · x · 2026-10-06
H-JEPA is presented as the first end-to-end learned hierarchical world model for long-horizon visual planning, built on SIGReg/LeWM. Key finding: forcing higher levels to discard fast-varying details yields temporal compression — representations evolve on slower timescales and predict far ahead while lower levels handle local execution. SIGReg prevents dimensional collapse across levels without complicated heuristics. Paper and code are open.
Related event: LeCun Team Releases H-JEPA, an End-to-End Hierarchical World Model(5 posts)→
More from Research
- 11-Square Packing Optimality Proved and Formalized in Lean With Help From Astra and Claude — ctjlewis · 2026-10-07
- Researcher unveils RSI paradigm: generic disposable agents plus an evolving knowledge base — yisongyue · 2026-10-07
- John Urschel proves Gaussian elimination growth factor is ~n^1/2, settling a Trefethen conjecture — ctjlewis · 2026-10-07
- Where Does Memory Live? RNNs, Transformers, and SSMs Compared Through Working Memory — Pretty_Upstairs9035 · 2026-10-07
- KNOWS Benchmark Jointly Evaluates Agents on Search, Tools and Visual Understanding — anmarasovic · 2026-10-07
- ID-Forcing Keeps Long Video Generation In-Distribution, Enabling Minute-Scale Videos Without Fine-Tuning — _akhaliq · 2026-10-07