LLMs, Reinforcement Learning, and Physical World Models
LLMs and reinforcement-learning systems both expose the same limitation when they are asked to act in the physical world: prediction alone is not a complete world model.
A useful physical model needs state estimation, object persistence, constraints, contact, mass, energy, delayed effects, and the ability to test whether an action changed the world the way the model expected.
Reinforcement learning supplies a way to connect action and outcome, but it can become brittle when the environment is poorly represented or the reward leaves important structure out. LLMs can carry broad prior knowledge, but text prediction is not the same as grounded control.
The important research question is how to combine learned representations with explicit physical state, uncertainty, feedback, and safe action. A machine that acts needs a model it can correct when reality disagrees.