Implications

For architecture. The design that has effectively won on the efficiency-robustness frontier is neither Fast-WAM's nor the imagine-then-execute paradigm's: compute the future once, cheaply, in latent space, expose it to the action branch as cached keys and values, and keep every reverse edge masked so the cache is action-independent and reusable across all denoising steps. The second Faster-WAM's SparseMoT with eight interaction stages out of thirty, ForeWAM's single prefill at σ = 1.0, GlanceWAM's asynchronous lookahead, and LAWA's latent-action channel are four independent arrivals at the same shape. The reverse-edge masking that Fast-WAM introduced is what makes this shape possible, so its architectural contribution outlives its headline claim.

For deployment latency. The cost of restoring the forward channel is now small and sometimes negative. LAWA pays 196.5 → 338.5 ms; the second Faster-WAM reports being faster than its own Fast-WAM reimplementation (252.95 vs 320.97 ms) because sparse interaction speeds the action denoising more than the extra video pass costs; GlanceWAM runs at 48 ms per chunk. The argument that keeping a future channel is prohibitively expensive no longer holds.

For evaluation. LIBERO is saturated and cannot discriminate on this question — every method discussed lands between 96 and 99. The mask's effect is visible only on LIBERO-Plus, RoboCasa in few-shot, RoboTwin 2.0 randomised, and real-robot long-horizon tasks. Any future claim about test-time imagination that reports only LIBERO average should be treated as uninformative about the question.

For interpretability and safety. Two consequences of a hard mask are not measured by success rate at all. When to Trust Imagination points out that without an inference-time future there is nothing to compare against reality, so plan-versus-outcome verification becomes impossible; its adaptive verifier gains 35 real-world points while cutting forward passes by 69%. BadWAM finds the imagination-free variant most adversarially fragile on LIBERO (a 53.4-point drop against 32–37 for the alternatives), though with the important caveat that this does not replicate on RoboTwin. And Beyond Task Success shows that a Fast-WAM-style model can match a WAM's success rate while behaving like a VLA in its object-directed metrics and carrying almost none of the predictive structure in its inference-time features.

---