The masking question · World Action Models since Fast-WAM

This is the question you asked about, and it needs to be stated carefully, because the literature answers three different versions of it and the answers diverge.

3.1 The reverse edges are genuinely settled

There is near-unanimous agreement on the other direction of the mask. Video tokens must not read action tokens, and clean anchor or conditioning tokens must not read future tokens. Fast-WAM states it ("the clean first-frame tokens do not attend to any other tokens"); the second Faster-WAM (arXiv:2608.04404) states it twice — "future slots may attend to the clean anchor and one another, while the anchor cannot attend to them" and "across experts, action queries may read video features, whereas video queries cannot read action tokens"; UNIVERSE blocks the video→action direction outright; GlanceWAM blocks all non-lookahead tokens from reading lookahead tokens; Next Forcing forbids clean context tokens from attending to noisy ones; Discrete-WAM enforces a causal context mask so that "noisy action tokens... cannot directly attend to their clean targets."

The reasons are two, and both are structural rather than empirical. First, if the video hierarchy does not depend on the action trajectory, it can be computed once and cached across every denoising step — which is precisely what makes the second Faster-WAM's one-pass conditioning and ForeWAM's Future-KV prefill possible. Second, if clean anchors could read the future, video prediction degenerates into copying; GlanceWAM's mask is designed so that "the backbone representations for observations and future targets remain mathematically identical to standard video co-training," which is what lets it attribute its gain to the lookahead channel alone.

3.2 The forward edge is not settled, and is now actively contested

On whether action tokens should be permitted to attend to future video tokens, the field has fragmented rather than converged. Grouping by what the action branch is actually allowed to see:

Position	Papers	What the action tokens see
Hard block, no future channel at inference	Fast-WAM, Enfold, World Tokens, ImageWAM, Discrete-WAM (policy mode)	Only present-derived features; futures shape weights during training only
Explicit hard block, argued on causal-validity grounds	UNIVERSE	Bidirectional V↔A block; historical context shared
Latent surrogate channel, pixels never decoded	ForeWAM, LAWA, GlanceWAM, Faster-WAM (2608.04404)	Future-position tokens as KV, latent actions, or asynchronous lookahead latents
Full joint attention over decoded futures	Efficient-WAM, MaskWAM, MWM, Next Forcing, RISE	Noisy or denoised future frame tokens directly
Continuous gate replacing the binary mask	NoiseGate	All future frames, attenuated by per-frame noise level

NoiseGate deserves separate mention because it reframes the question rather than answering it. Under the noise-as-masking view, a per-latent timestep vector $t_v = [t_1, \dots, t_F]$ with independent per-frame values makes each future frame's key–value contribution a continuous function of its noise level, so a frame at high t_f is nearly masked and one at low t_f is nearly visible. The paper trains a small gating policy network to emit per-frame time increments via GRPO against binary episodic reward, and verifies empirically that action-token attention to each frame decays monotonically with that frame's t_f. Its ablation is the cleanest evidence that the degree of future visibility is worth optimising: 57.5 with a shared scalar schedule, 61.3 with independent per-latent noise training, 63.4 with a hand-crafted monotone schedule, 67.5 with the learned policy. The learned schedules do not collapse to either extreme — near futures are opened early, distant futures often stay partially masked, and the ordering changes by task.

3.3 What the evidence actually shows

The honest summary is that almost nobody has run the decisive experiment, which would hold a single model, training recipe, and parameter count fixed and toggle only the action→future-video attention edge, measuring both in-distribution success and out-of-distribution robustness. The nearest approaches, in descending order of cleanliness:

UNIVERSE (arXiv:2607.05133) runs it in the driving domain, and the accuracy result is null. Its Modality-Decoupling Visibility Mask sets $M(i,j) = 1$ for $i \in V \cup A$ and $j \in C$ (both future modalities read shared history) and $M(i,j) = 0$ in both directions between future video V and trajectory tokens A, justified explicitly as preventing "future-target leakage." Its own ablation reports 91.0 PDMS with the mask and 90.9 without — well inside noise. What the mask buys is not planning accuracy but train/test consistency: because trajectory tokens never depended on future video during training, the video branch can be dropped at inference with "minimal distribution shift," which is where its 4.3× speedup comes from. Ablating video co-training, by contrast, costs 2.8 PDMS (91.0 → 88.2), and replacing the shared DiT with Fast-WAM's two-DiT arrangement costs 1.4 (91.0 → 89.6). So on this evidence the mask is a deployment convenience with no measured accuracy cost — in a domain with short horizons and dense in-distribution data.

GlanceWAM comes closest in manipulation, and the result points the other way. Because no non-lookahead token attends to the lookahead frame, its backbone representations are provably identical to plain video co-training, so the difference between its 72.2% on 24-task RoboCasa and the imagination-free co-training baseline's 64.4% is attributable to the lookahead conditioning channel alone. That is a 7.8-point gap at 50 demonstrations per task. (The paper's introduction states +8.4% while its own abstract figures give +7.8; I could not retrieve the Section 4.3 table that reconciles them, so treat the magnitude as approximate.)

ForeWAM's component ablation isolates the channel directly. Starting from a base direct-policy configuration at 53.6 on LIBERO-Plus, adding the Future-KV interface alone reaches 58.5, latent-action supervision alone reaches 58.0, and both together reach 61.6. Its per-perturbation comparison against Fast-WAM is the most dramatic single number in the debate: 62.5 against 16.4 on camera-viewpoint shift, a 46.1-point gap, and 58.8 against 37.7 on sensor noise — while Fast-WAM actually wins on robot perturbation (44.5 vs 37.4) and lighting (78.2 vs 74.1). The deficit is specific to geometric and viewpoint shift, not photometric shift.

LAWA supplies the data-scale gradient, and it is monotone. With matched implementations on RoboCasa, Fast-WAM trails a Joint-WAM baseline by 8.1 points in few-shot (56.0 vs 64.1, roughly 100 demonstrations per task) but by only 2.5 points at full data (76.3 vs 78.8, 1,000 per task) — a gap 3.2× larger in the low-data regime. On real-world tasks at 50 trajectories per task, Fast-WAM scores 8.8% average against LAWA's 40.0%, including 0.0% on both long-horizon tasks. On LIBERO-Plus its matched Fast-WAM reaches 60.0 against Joint-WAM's 70.4. The authors state plainly that their comparison shares "downstream splits, optimization, batch sizes, and training steps, but not paradigm-specific branches, objectives, parameter counts, or inference costs," so it is a matched-training comparison, not a matched-capacity one.

The second Faster-WAM makes the argument its thesis. Its abstract asserts that efficient alternatives that "remove future modeling at inference time... may lose the robustness benefits of temporal reasoning," and its Figure 1 caption states that "inference-time future conditioning is critical for robust WAMs." Its controlled reimplementation puts Fast-WAM at 49.14 on LIBERO-Plus against a Joint-WAM at 66.30 and its own one-pass design at 73.57, while running faster than both.

The Temporal Ratio paper (arXiv:2607.08127) gives the mechanism a measurable quantity. It defines TR as the attention mass the action head assigns to future latent frames relative to the anchored current frame, and reports two properties: a model's TR predicts its compositional generalisation capacity, and TR fluctuates within an episode, rising during planning phases and falling during precise manipulation. This matters for your question because Fast-WAM's mask sets TR identically to zero by construction — the quantity the paper identifies as predictive of compositional generalisation is not merely low in Fast-WAM, it is structurally unavailable. Fast-WAM scores 9.4% compositional out-of-distribution success in that paper's evaluation; TR-adaptive guidance takes a video-action model from 55.7% to 59.4% while holding in-distribution performance at 94.0%.

3.4 The consensus, stated as precisely as the evidence allows

There is no consensus that future frames must be masked out of the action tokens. What has emerged instead, and what I would describe as the closest thing to a settled position as of late August 2026, is a three-part view:

Video prediction as a training objective is necessary and uncontested. Every paper in this survey retains it. Fast-WAM's 97.6 → 93.5 and 91.8 → 83.8 ablations, its real-world collapse to 10%, and UNIVERSE's 91.0 → 88.2 all point the same way, and nobody has published a counterexample.
Decoding future pixels at inference is unnecessary and is now generally regarded as wasteful. This part of Fast-WAM's thesis has been absorbed rather than disputed. Enfold's ablation puts distilled generator states (97.8) above rendered future pixels (96.0); ImageWAM cuts FLOPs roughly sixfold by never decoding an edited frame; ForeWAM, LAWA and GlanceWAM all keep a future channel while explicitly refusing to render it. Even papers that keep decoded futures make them coarse, low-resolution and two-step.
Severing the attention edge between action tokens and future-position tokens is a separable choice, and the evidence now weighs against it outside the data-rich in-distribution regime. Fast-WAM conflated (2) and (3) — its mask and its inference-time deletion were introduced together and ablated only jointly, since the four-variant comparison varies the inference paradigm and the training objective but never the mask itself. The papers that separate them find that the cheap part is skipping generation and the expensive part is closing the channel.

The strongest form of the counter-position, which UNIVERSE's null ablation supports, is that the mask costs nothing in accuracy when the deployment distribution matches training and the horizon is short, and that what it buys — a future branch that can be dropped without distribution shift, and a video KV cache independent of the evolving action trajectory — is worth having for its own sake. That position is defensible on the current evidence for autonomous driving planning. It is not defensible for few-shot manipulation, viewpoint shift, or long-horizon contact-rich tasks.