NYU and Meta FAIR introduce ID-WM, a hierarchical world model beating baselines by 38.5%

ylecun · x · 2026-11-05

Yann LeCun amplified ID-WM (Implicit Dynamics World Models), a NeurIPS 2026 paper from NYU Courant and Meta-FAIR.

Core idea: a hierarchical world model decouples "where to go" from "how to get there" — a high-level model samples feasible transitions across the state manifold as subgoals, while a low-level world model acts as an inverse dynamics solver grounding them into precise actions.
Results: across three navigation and manipulation suites, it improves over the strongest policy/diffusion-based planners by 38.5% and over single-level world-model MPC by 135%, especially on unseen long-horizon configurations.
Why it matters: the hierarchy mitigates compounding errors, letting the planner synthesize novel long-horizon plans from short segments beyond trained behavior modes. Code is not yet released; demo videos are on the project page.

More from Embodied