Next-Latent Prediction Transformers Learn Compact World Models (2025) (arxiv.org)

Interesting, I'll test that out, I have been working on retrieval approximations and this is the learned version of it, so hopefully I can push this beyond what I'm working on.