We can look at examples of video generating models. I’d argue they have to have a meaningful and persistent representation of the world internally. Consider something like Genie as an example https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/

It doesn’t have volition, but it does have intelligence in the domain of creating consistent simulations. So, it does seem like you can get a domain specific intelligence through reinforcement training.