World models don’t compress enough. I think LLMs generalise because they are extremely compressed. “Ship” means both rocket ship and naval ship, which forces the abstraction “large vessel.” RL on boats can then transfer to rockets when the task is about large vessels. And that’s just one token. WMs currently use huge representations that strictly do not require this kind of abstraction. I think it should be possible to train WMs to create factorised representations: s, r. s is the super compressed state: ‘ship’. R is the residual, everything required to reconstruct that specific instance: ‘wooden on water’. When it’s not required one could drop the r and model dynamics purely in the abstract space. One could then decode the instance with the residual of an observation.