Hierarchical Planning

Human planning is multi-level: a trip from New York to Tunis is not planned at the muscle-movement level. High-level ("go to airport") decisions compose into mid-level ("hail a taxi") then low-level ("stand up, walk to door") then subconscious (actual motor control).

For AI systems, hierarchical planning means the world model operates at multiple levels of abstraction — higher levels make long-term, coarse predictions; lower levels manage short-term detailed actions. The payoff: horizons that are computationally intractable at the low level become tractable when most of the search happens at the abstract level.

Key Claims

• Hierarchical world models mitigate error accumulation — H-WM's symbolic-plus-visual architecture reduces drift in long-horizon TAMP problems by using the symbolic layer as a stabilizing prior. Evidence: moderate (H-WM)

• Structured sparse prediction is a form of hierarchy — StructVLA predicts physically meaningful keyframes rather than every frame; this is a temporal hierarchy (skip unimportant frames) analogous to spatial hierarchies (symbolic-plus-visual). Evidence: moderate (StructVLA)

• Most current world models are single-level — V-JEPA 2 and Genie 3 operate at a single abstraction level; hierarchical extension is an open research direction. Evidence: inferred from paper descriptions

Open Questions

• How many levels of hierarchy are optimal? Two (symbolic + visual) is common; is more better?
• Should the levels be hand-designed or learned?
• Can JEPA representations at different layers serve as a natural hierarchy, or does hierarchy require explicit architectural separation?
• How do you train a hierarchical world model end-to-end without the levels collapsing into each other?

Related Concepts

• World Models — hierarchy is a structural property
• System 2 Reasoning — hierarchical planning is a specific System 2 recipe
• Joint Embedding Predictive Architecture (JEPA) — candidate substrate for hierarchical world models

Changelog

• 2026-04-22 — Initial compilation from H-WM and StructVLA.