How the field split · World Action Models since Fast-WAM

The July 2026 tutorial From World Models to World Action Models (arXiv:2607.00836) sorts the space into four paradigms — imagine-then-execute, video-feature-conditioned action prediction, joint video-action modelling, and auxiliary video prediction for policy learning — and places Fast-WAM in the last. Since March the follow-up work has organised itself into four distinguishable responses to it, and it is worth noting that they are not simply "for" and "against."

Papers that accept the premise and push the efficiency argument further. Faster-WAM (arXiv:2608.02365, Huawei Noah's Ark, 3 Aug) argues that Mixture-of-Transformers designs needlessly tie the action module's depth to the video backbone's, and replaces layer pairing with a "Dock of Transformers": a single-layer action head reading key–value states from all thirty video layers through a docking interface. Channel mixing applies learned $W^K, W^V \in \mathbb{R}^{D' \times D'}$ , layer mixing applies a per-head aggregation matrix $A_h \in \mathbb{R}^{L_a \times L_v}$ , and the fused states enter mixed-context attention as $O_a^{(j)} = \mathrm{Softmax}\big(Q_a^{(j)}[K_a^{(j)}, \tilde{K}_v^{(j)}]^{\mathsf{T}}/\sqrt{d}\big)[V_a^{(j)}, \tilde{V}_v^{(j)}]$ . A separate contribution undoes the 3D RoPE on cached video keys, fuses in the canonical space, then applies the action head's 1D RoPE — a correction the authors argue is impossible inside a single shared attention, and therefore structurally unavailable to MoT designs. The result is 98.5 LIBERO average at 66.5 ms against their measured 211.7 ms for Fast-WAM, and 75.0 on LIBERO-Plus against 51.5, a 23.5-point improvement in out-of-distribution robustness. Notably, the paper contains no ablation of action-module depth despite the title question, and its sequential ablation (49.5 → 60.3 with docking → 66.8 with KV-fusion → 71.3 with RoPE alignment → 75.0 after removing text cross-attention) is reported as a bar chart rather than a table.

Enfold (arXiv:2607.26657) takes Fast-WAM's logic to its conclusion by arguing that the reusable asset of a video generator is not its output but its internal computation, distilling multi-level generator states into a present-only encoder via $L_{\mathrm{G2R}} = \mathrm{SmoothL1}(\hat{r}^{G}, \mathrm{sg}(r^{G}))$ and testing the result by feeding it back to condition generation. Task heads read the representation through a stop-gradient, so task loss cannot reshape the encoder. It reaches 97.8 LIBERO average at 134 ms, and its Flash variant at 49 ms — and its ablation is directly relevant here, since conditioning on rendered future pixels scores 96.0, action-only scores 94.9, and multi-level distilled states score 97.8. ImageWAM (arXiv:2606.19531) makes the related argument that a target-frame edit is a better-matched prior than a video rollout, conditioning the action expert on the KV caches of an image-editing denoiser without ever decoding the edited frame, and reporting 83.1 on LIBERO-Plus with a FLUX.2 4B backbone at 263 ms and 9.72 TFLOPs. World Tokens (arXiv:2608.09730) routes all visual-linguistic context through 256 learned tokens that also condition a future-video denoiser deleted at deployment, and C³ache (arXiv:2606.08962) accelerates what remains by caching denoising residuals across inference chunks, reaching 2.51× on LIBERO with no accuracy cost — though the same configuration collapses to 75.85 on RoboTwin, fifteen points down, which is a useful reminder that acceleration tricks validated on short-horizon benchmarks do not transfer automatically.

Papers that keep imagination but make it cheap. Efficient-WAM (arXiv:2606.10040) shrinks the video expert to under a billion parameters, downsamples future frames to roughly 192×160 before VAE encoding, and allocates asymmetric denoising budgets — two steps for video against ten for actions — reaching about 98 ms per chunk against Motus's 3215 ms while keeping full joint video-action attention. ForeWAM (arXiv:2608.11605) performs a single video-DiT prefill over the current latent concatenated with stochastic Gaussian future slots, caches the layer-wise key–value states, and reuses them throughout action denoising, so the action branch sees future-position tokens that were never decoded into pixels. LAWA (arXiv:2608.24882) replaces the future-video channel with a compact discrete latent-action sequence that the action expert may attend to, explicitly designed so the policy "use[s] the evolving future intention without leaking future-video information." GlanceWAM (arXiv:2608.23927) moves the rollout off the critical path entirely, generating a lookahead latent asynchronously at roughly a three-second horizon so that one prediction serves about four consecutive 0.8-second action chunks, and consuming it directly in latent space at 48 ms per chunk.

Papers that make imagination conditional. RISE (arXiv:2608.20430) frames the budget as a per-scene decision, learning a Roll/Stop gate over rollout depth and producing the debate's most counterintuitive datum: on simple driving scenes, zero imagination scores 89.9 EPDMS while forcing four steps of rollout drops it to 88.4, whereas complex intersections improve substantially. When to Trust Imagination (arXiv:2605.06222) uses the predicted future as a verifier rather than a conditioner, cutting WAM forward passes by 69.10% and execution time by 34.02% while improving success by 2.54% in simulation and 35 points in the real world; its benefit concentrates entirely on the hard subset where the base policy is below 65%. NoiseGate (arXiv:2605.07794) dissolves the binary question altogether, which is discussed in the next section.

Papers that measure rather than build. Beyond Task Success (arXiv:2606.01095) applies behavioural rollout analysis and sparse-autoencoder feature labelling to seven policies and finds that Fast-WAM's success rate hides VLA-like behaviour: a wrong-object path rate of 0.673 against π₀.₅'s 0.687 and LingBot-VA's 0.174, and a predictive-feature share in the action stream of 2.0% against LingBot-VA's 4.3% and VLA-JEPA's 59.8%. Its conclusion is that "auxiliary WAMs compress or weaken the future signal at inference time." BadWAM (arXiv:2607.15207) attacks imagination–action desynchronisation and finds the action-only variant most fragile on LIBERO, dropping 96.5 → 43.1 against 36.6- and 32.3-point drops for the imagination-retaining variants — though on RoboTwin all three degrade nearly identically, so that critique does not replicate across benchmarks.