Tag
This paper introduces a Bayesian framework for hierarchical LLM agents to decide when to escalate to stronger models during reasoning, formulating it as an optimal-stopping problem and providing theoretical guarantees on performance.
Proposes interaction locality, a task-geometry-aware framework for measuring whether information flow in spatial reasoning models stays within local cells or crosses into global structure, and applies it to HRM, TRM, and MTU3D models on grid benchmarks and embodied 3D grounding.
Yann LeCun predicts that within 12-18 months, a general method for training hierarchical world models will emerge, learning from video and real-world data to aid planning in robotics, healthcare, and beyond, scaling toward a universal world model.