Part III
Research directions
What you should noticeOpen problems: richer envs, fusion stamps, scale, and staying honest when intrinsic dimensionality is tiny.
Open directions that matter for this project and for the broader world-model literature:
- Richer environments — TwoRoom’s low intrinsic dim fights high-dim Gaussian targets; PushT-scale worlds recover more of LeWM’s behaviour.
- Multi-sensor fusion — shared real-world stamps
(x,y,t)let attention bind camera and lidar tokens; mismatched stamps do not. - Planning + scale — latent CEM is fast when
zis small; keeping latents informative under SIGReg at scale is still active research. - Honesty — toys must show claims (shuffle/stamp, neighbourhoods, surprise), not only assert them.
Desktop next step in JEPADot: finish fusion as a capstone after the single-sensor web story is solid. See also docs/leworldmodel-fidelity.md in the repo.