What it is
Studies why world action models generalize and shows test-time future conditioning matters. It proposes Simple-WAM: one pass over fully noised future tokens instead of iterative video denoising. Instead of either fully denoising future video at test time or dropping it entirely, keep future-token conditioning but use a single pass on fully noised future tokens. This isolates that preparing future representations, not generating clean frames, drives WAM generalization.
Key results
- 79.5 perturbation success vs 67.7 explicit and 53.8 latent
- 3.8× faster than explicit WAM, with only 1.2× latent latency
- Explicit beats latent on task gen w/ video: 69.9 vs 5.9 success
Inference-time access to future-token representations, not clean future-frame generation, is what preserves WAM generalization.