Open Problems¶

Keep only questions that could plausibly be tested here; do not try to collect every open problem in the field.

1. When does a video model become an interactive world model?¶

Move beyond attractive video generation toward action-conditioned prediction, long-term state consistency, executable planning, and explicit failure boundaries.

Minimal experiment: in a simplified simulator, compare how pixel prediction, latent prediction, and true environment-state prediction affect planning.

2. Do VLAs need explicit long-term memory?¶

Success on short tasks does not imply success on tasks that span hours. Separate the contributions of memory, state estimation, planning, and action policy.

Minimal experiment: hold the policy fixed while changing only the history window, external memory, and task decomposer.

3. Can verifiers replace large-scale human feedback?¶

This can be tested in math, code, and simulated environments. Open-world and aesthetic tasks require studying verifier bias and incompleteness.

4. How can self-generated tasks avoid distribution collapse?¶

The task generator must keep finding tasks the model cannot solve but that can still be evaluated reliably, rather than repeating simple variants.

5. How can automated discovery prove that progress is real?¶

Keep an independent holdout, preregistered metrics, rollback versions, and a compute-budget record.