World Action Models in August 2026 — Monthly Progress Issue

August 2026 was the month world action models (WAMs) stopped being a slogan and turned into an engineering agenda. The shared template that emerged across the screened corpus is a video- or latent-dynamics backbone that predicts scene evolution together with an action expert that reads it — and the month's work concentrates on the two places that template hurts: what the predicted future should be made of, and how to keep it cheap enough for a real control loop. Latent and structured futures (pointmaps, geometry, semantics, motion) were the most active answer to the first question, with several papers showing that supervision beyond RGB changes robustness rather than fidelity; asynchronous lookahead, cached future keys, compute-selectable streams and post-training that costs no extra sampling steps were the most active answer to the second. A third, quieter change is that the field began auditing itself: off-expert action following, hidden-state tracking, intervention fidelity and simulator contracts were all measured in August, and the first results say that visual quality and control capability are only loosely coupled.

Window: papers first submitted 2026-08-01 to 2026-08-31, read from the library. Coverage: 341 candidates from four retrieval routes, 204 kept by screening, 48 read at full-text depth (all 48 completed, in two passes), 15 expanded below with their figures. This is the first issue of the series, so it is a baseline rather than a comparison.

Figure 1. The five themes of this issue and the featured papers inside them; the numbers are reference numbers used in the text.

1 Making the imagined future affordable

The month's most consistent engineering story is that test-time imagination was decoupled from the control clock. Three papers attack it at three different points of the same pipeline: how often the future is refreshed, how the refreshed future reaches the action head, and whether the policy needs a future at all once training is over.

1.1 GlanceWAM: predict rarely, act continuously

GlanceWAM's argument is that the speed–success dilemma of WAMs comes from coupling video generation to the control clock, not from visual foresight itself [1]. Its fix is a two-timescale scheme on one shared SkyReels-V2-1.3B video DiT: a single sparse lookahead latent ~3 s ahead is generated on a slow background clock and held, while a flow-matching action head decodes action chunks against that stale latent at control rate, never decoding it to pixels. Co-training uses a randomized offset so the action head learns inverse dynamics toward a lookahead frame at varying temporal distance, which is what makes a held latent usable rather than merely stale.

Figure 2. GlanceWAM: (a) isolated-lookahead co-training and (b) asynchronous latent-space inference on one shared video DiT — the sparse lookahead latent and the held-state schedule are the elements to note [1] (reproduced from the original paper).

The reported evidence is unusually well controlled for this literature: 72.2% on a 24-task RoboCasa kitchen suite against 67.1% for Cosmos Policy at the same 50 demos/task, 79.2% on real-robot trials against 69.6% for π0.5 and 48.0% for Fast-WAM, and 48 ms per 16-action chunk against 1133–3812 ms for synchronous WAMs at comparable success [1]. Two ablations carry the mechanism: zeroing the lookahead tokens on the trained checkpoint drops RoboCasa to 61.6%, below the 64.4% no-lookahead baseline, and a horizon sweep peaks near 3 s — so the future is used, and roughly three seconds of it is enough. The caveat the authors themselves flag is that simulation numbers come from a blocking evaluation client, so fully pipelined real-time closed-loop success is inferred rather than shown.

1.2 ForeWAM: keep the future's keys, not its pixels

ForeWAM (Foresight Without Seeing) targets the interface rather than the refresh rate [2]. It keeps the direct-policy interface — actions conditioned on the current observation only — but runs one video-DiT prefill over the clean current latent concatenated with Gaussian-initialized future slots, caches the resulting key–value states, and lets every action-denoising step reuse them; dynamics registers are supervised during training and emitted at inference so future-indexed computation can reach the action pathway without pixel decoding. This is a sharper distinction than "latent vs pixels": the paper argues the field conflates forecasting with rendering and that a video DiT already forms hidden states at future positions.

Figure 3. ForeWAM: VAE-encoded video and action streams with dynamic registers, plus the training and inference masks that route future keys into action denoising [2] (reproduced from the original paper).

On the authors' protocol the payoff is concentrated in distribution shift: 77.6% on LIBERO-Plus against 51.5% for the matched Fast-WAM (both without policy-level embodied pretraining) and 59.2% versus 49.5% on RoboCasa at 50 demos/task, with a 6.3× latency reduction (88.7 ms vs 556.3 ms) attributed almost entirely to the action stage [2]. The component ablation makes the interface the point: latent-action supervision alone moves LIBERO-Plus 53.6 → 54.8, Future-KV alone 58.5, and the combination 76.7. On real-robot OOD tasks the same model still trails π0.5 (40.0% vs 45.0%), which the paper reports and which is a useful reminder that beating the WAM baseline is not the same as beating a strong VLA.

1.3 WAM-OPD: keeping the two-output structure during post-training

WAM-OPD asks a different question: if a WAM has two outputs, what should a teacher supervise? Its answer is both, jointly [3]. A frozen stronger WAM is queried on the student's own rollout histories and returns paired targets — a predicted future video and the action chunk generated from that video — and the student is trained on both while keeping the exact video-to-action computation it uses at deployment; the delta over VLA-style on-policy distillation is that action supervision is video-conditioned rather than token-level, and the delta over RL-style world-model post-training is that rewards are replaced with labelled targets.

Figure 4. WAM-OPD's three stages — student inference, teacher labelling, on-policy distillation — with video and action both supervised from the student's own histories [3] (reproduced from the original paper).

Reported gains over the same student checkpoint are large and consistent: 33.8% → 65.7% averaged over 12 RoboTwin 2.0 tasks and 51.4% → 64.6% on four real-robot tasks, at an unchanged deployment budget of one video and one action step [3]. Joint supervision beats either single modality on all three ablation tasks (+5.0 / +1.7 / +5.0 points over the better single-modality mode), which is the specific evidence for the paper's claim that the video pathway carries usable information. Two things temper it: the teacher history source is task-dependent (student histories far better on one task, teacher histories on another), and the main table and the ablation tables appear to rest on different evaluation draws for one task — the analysis records this as an inference from the reported numbers rather than a claim by the authors.

1.4 The rest of the cost axis

Two further papers show the range of the same move. RISE replaces a fixed rollout depth with a learned Scheduler — a latent evaluator that predicts a risk profile and a planning-gain profile, plus a rollout gate that trades predicted gain against compute — and reports NAVSIM-v1 PDMS 91.5 and v2 EPDMS 90.8 at 2.40 average rollouts [4]. Faster-WAM finds that deep action modules are not needed: an architecture keeping one action block reports LIBERO 98.3% at 52 ms per action chunk versus 196 ms for Fast-WAM, a 3.8× latency cut mostly inside action generation [5]. At the radical end, SLIM-0.5B deletes test-time prediction entirely, keeping future prediction as a latent training signal inside a 0.47B two-stream trunk and reporting LIBERO 97.5% with 60.6 ms per chunk against 193.1 ms for π0.5 and 360.6 ms for Fast-WAM — while sitting 2.05 points below VLA-JEPA on LIBERO-Plus [6]. Its reusable artefact is a collapse diagnostic: a no-EMA variant shows lower future-latent MSE with effective rank falling 61.28 → 13.95, i.e. a good-looking loss hiding a collapsed representation.

Table 1. Where each paper removes test-time cost, with the authors' reported result and the evidence basis for this issue.

Takeaways

• Test-time imagination is being converted from a per-chunk cost into a scheduled or cached resource; the reported latency gains (6× to 3.8×) come with unchanged or better success on the authors' benchmarks.
• The mechanism claims are supported by ablations that remove the future channel (zeroing lookahead, freezing the cache path, dropping joint supervision), which is stronger evidence than a new state of the art on a saturated benchmark.
• The open question is closed-loop realism: GlanceWAM evaluates with a blocking simulator client, and no featured paper here reports a matched-embodied-pretraining, matched-compute comparison against the VLAs it beats on OOD tasks.

2 What the imagined future is made of

The second concentration of effort is about the target, not the schedule. Several papers keep the backbone and change what the video expert is asked to represent — an argument that RGB futures entangle task-relevant state change with nuisance appearance, and that other projections of the future carry the control signal more efficiently.

2.1 DreamWAM: the future as motion, geometry and semantics

DreamWAM keeps the Fast-WAM two-expert backbone and adds three supervised future views — optical flow, depth-derived geometry and DINOv2 semantics — as complementary projections of action-relevant future state [7]. Its routing insight is that these targets are not interchangeable: VAE-native motion is denoised jointly with the RGB latent, whereas heterogeneous encoder features enter through gated residual branches, and at inference all beyond-RGB pathways are removed so the deployed policy stays RGB-only. The paper's sharpest single result is a negative control: jointly denoising all three targets reaches 97.85% on LIBERO, below the 98.00% RGB-only baseline, while the proposed mixed routing reaches 98.90% — mechanism matters more than adding supervision.

Figure 5. DreamWAM's training-time world states (RGB, flow, depth, DINO) and gated world residual prediction; the gates are what keep the extra targets from hurting [7] (reproduced from the original paper).

Where the change shows up is robustness: LIBERO-Plus rises from 69.16% to 75.47% in joint-inference mode and from 51.36% to 63.44% in the no-rollout mode the paper cares about, improving on all seven shift dimensions; on a real dual-arm platform the perturbed-condition average goes 55.6% → 74.4% [7]. All simulation baselines are the authors' own re-implementations, and the controlled comparison is the matched Fast-WAM pair rather than a cross-method ranking — the analysis notes this explicitly and treats the robustness claim as well supported while the mechanism attribution (which of motion, geometry, semantics drives the shift gains) is not ablated under shift.

2.2 Flex-π: geometry and semantics as selectable streams

Flex-π pushes the same intuition further by predicting pointmaps and object-centric DINO features, but its enabling observation is infrastructural: the frozen Wan-2.2 video VAE encodes 3D pointmaps accurately even though it was never trained on them, so geometry supervision needs no new sensor, no new visual prior and no inference cost [8]. Per-stream presence masks and an output generation mask turn one 6B checkpoint into a selectable latency/accuracy frontier — from ~60 ms action-only to ~193 ms with all streams.

Figure 6. Flex-π: RGB, pointmap and DINO streams through a shared video VAE, with the stream-presence and output masks that turn one checkpoint into a latency/accuracy frontier [8] (reproduced from the original paper).

The reported ablations are the clearest quantitative statement of the theme: from one checkpoint, adding DINO to a video-only stream gives +6.8% and adding pointmaps a further +20% on five RoboTwin tasks at 50 demos; the largest real-world margins appear at low demonstration budgets (78.8% at 50 demos/task versus 31.4% for π0.5 and 41.9% for Fast-WAM) [8]. Two limits are stated by the authors: the extra modalities slow convergence (≥10 fine-tuning epochs), and full joint generation remains slower than comparable VLAs. The claim that geometry helps even when never generated at inference is the falsifiable core, and the paper's own evidence for it is that action-only decoding already beats every real-world baseline.

2.3 JEPA-WAM: supervise the transition, not the frame

JEPA-WAM changes the target's nature rather than its modality [9]. A frozen V-JEPA 2.1 encoder encodes the current and a future observation stacked in time into a single patch-structured target, and one shared 0.5B predictor is supervised by patch-wise cosine distance while simultaneously producing the representations that condition a flow-matching action expert. The design claim is that the transition relation is not recoverable from independently encoded endpoints: supervised endpoint differencing reaches 70.9% on LIBERO-Plus against 79.2% for the joint target, and disabling transition prediction while keeping the V-JEPA features drops to 73.2%.

Figure 7. JEPA-WAM: a frozen V-JEPA transition target, an LLM predictor and a flow-matching action expert sharing one backbone [9] (reproduced from the original paper).

The results position latent transition supervision as a cheap add-on to a strong VLA rather than a replacement for robot pretraining: 79.2% on LIBERO-Plus without large-scale policy pretraining, 86.3% when the same supervision is added to π0.5 (from 84.5%), 59.8%/54.2% real-robot ID/OOD against π0 at 51.8%/22.5% [9]. Standard LIBERO stays near-saturated, so the honest summary is that this family buys robustness and deployment cost, not peak in-distribution success.

2.4 The rest of the target-design space

Three further papers mark the edges. GaussianWAM shows the same lesson with a different teacher: a transient 3D Gaussian field built at training time carries VGGT-Omega depth and CLIP semantics, and distilling those targets into the WAM's current-observation latents lifts FastWAM from 52.05% to 71.29% on LIBERO-Plus and Cosmos Policy from 71.52% to 77.30%, at no inference cost — with the unification itself worth 1.92 points over naive independent distillation [10]. LEON argues the transition operator, not its target, is the underexplored axis: replacing a transformer predictor with a context-modulated low-rank operator plus additive forcing raises VLA-JEPA's LIBERO average from 97.2% to 99.05%, but leaves the policy-facing LaWAM instantiation neutral (84.13% vs 84.50%), so the paper's own framing accepts a near-baseline outcome there [11]. Vid2WAM attacks the same question from the data side, distilling video-diffusion priors into a WAM so that withheld task trajectories become usable: 54.7%/55.3% on novel RoboTwin tasks against 45.0%/42.8% for Fast-WAM, and 30% versus 5% on one real novel task, at 212.5 ms per chunk [12].

Table 2. Future-target design choices and what they change, as reported by the authors.

Takeaways

• Across five independent groups the same finding recurs: supervision beyond RGB improves robustness under shift rather than in-distribution success, and the gain survives when the extra stream is discarded at deployment.
• Routing is load-bearing. DreamWAM's all-denoise variant is worse than RGB-only, and GaussianWAM's unified carrier beats naive multi-teacher distillation by a small but consistent margin — adding targets naively does not work.
• Everything in this theme is measured on LIBERO/LIBERO-Plus, RoboTwin or RoboCasa, often with the authors' own re-implementations of baselines; cross-paper percentages are indicative only.

3 Scale, data and one model with several interfaces

A third group of papers treats the WAM as the endpoint of a pretraining pipeline and asks what makes it general: how much video, what action interface across embodiments, and how many roles one weight set can play.

3.1 ZimaBlue: video scale as the axis

ZimaBlue's thesis is that scaling action-free embodied video, not a new policy architecture, is the main axis for generalization [13]. It adapts a Wan2.2 text-to-video DiT through three stages — causal egocentric and robot video pretraining up to 120,000 hours, cross-embodiment mid-training over a unified 100-dimensional state–action space where each embodiment activates only its native slots, and target-robot post-training — then deploys the result as an asynchronous Slow (5B) / Fast (0.5B) pair whose link is a key–value cache exported from the Slow model's early layers.

Figure 8. ZimaBlue's Slow 5B video DiT and Fast 0.5B action DiT, linked by layered KV injection from the Slow model [13] (reproduced from the original paper).

Reported evidence is broad and, for this corpus, unusually real-robot-heavy: 77.8% zero-shot average over 12 held-out real tasks versus 36.1% from scratch, 46.1% with action-labelled data only and 66.9% at 60K hours of video, against π0.5 at 54.4% and DreamZero at 53.6%; 57.5% versus 15.0% on the perturbed suite; LIBERO-Plus 92.0% with per-benchmark fine-tuning and 86.7% zero-shot; and 16.5% versus 7.9% for ABot-M0.6 on RoboCasa365's unseen-composition split, where a VLA trained on 100K hours of real-robot data still leads overall [13]. The latency story is the deployment half: 449.6 ms → 33.0 ms across Slow-only, dual-system and distilled+compiled configurations at a 2.8-point success cost. The analysis is explicit that the four training configurations are cumulative, so video volume is not cleanly isolated from multi-embodiment action data, and no matched-compute baseline spends the same budget on robot data instead.

3.2 Riemann-1.0: one causal sequence, action before consequence

Riemann-1.0 serializes multi-view visual latents, robot states and actions into a single causal sequence and predicts them action-first, then their visual consequence, with the same mask and objective at pretraining, post-training and deployment [14]. That ordering is the design choice: the paper positions it against joint denoising, video-first autoregression and decoupled two-DiT designs, and uses latent-action-model pseudo-actions on mixed human–robot trajectories, a λ schedule from 0.1 to 0.9, and a 230,000-hour corpus to make weakly supervised video usable under one objective.

Figure 9. Riemann-1.0's unified trajectory tokens with action-first causal visibility, and its separate latent-prediction and action-prediction branches [14] (reproduced from the original paper).

The policy half is well evidenced: 62.6% on RoboCasa-365 against 54.2% for the strongest named baseline, 94.3% on RoboTwin 2.0, 99.0% on LIBERO, 85.0% success with 94.43% progress on four real task families, and 85.0% on held-out real OOD tasks against 50.0% for π0.5 [14]. The world-model half is not: multi-embodiment rollouts are shown qualitatively (Figure 10), with no FVD/PSNR or action-faithfulness metric, and there is no ablation separating the causal factorisation from joint or decoupled alternatives. The claim "one checkpoint is both policy and simulator" should be read as a design position with a strong policy and an unmeasured simulator.

3.3 Motus2: policy, simulator and evaluator in one weight set

Motus2 keeps one shared parameter set and exposes it as three interfaces through different query modes — a policy (the WAM), an action-conditioned simulator, and a value evaluator — factored action-first [15]. The three are wired into a closed loop: propose action chunks, imagine their consequences, score the outcomes, then use the scores both for Best-of-N test-time selection and for DiffusionNFT-style policy updates, so the policy can improve beyond the demonstrations. It also carries substantial data and scaling content: ~130,000 hours of egocentric corpus, a tactile expert, and a stereo video scaling law that is log-linear over nested 2k–20k-hour subsets.

Figure 10. Motus2: one parameter set exposed as policy, simulator and evaluator, with the model-based RL and planning loop on the right [15] (reproduced from the original paper).

On five physical tasks with matched target-robot fine-tuning, Motus2 reports an 84% macro-average success rate against 51% for its own pretrained-only initialization and 0% for both a video-model baseline and π0.5, with the improvement loop adding further points (65.0% → 75.0% with MBRL and planning on two tasks) and the tactile expert worth 12.5 points on two contact-rich tasks [15]. The evidence gap is the one the paper's own framing invites: the loop is run with offline prefixes plus imagination rather than real online robot data, so "self-evolving" here means policy improvement inside the learned simulator.

3.4 Human video, compact stacks and 3D state

Four more papers round out the theme. LD4WAM learns a motion-aligned latent dynamics variable from human video — a next-semantic-feature transition explicitly regressed onto end-effector motion — and pairs it with a 5,086-hour, 8-dataset corpus unified to LeRobot; it reports 93.4% on 50 RoboTwin tasks and 70.5% real-world against π0.5's 63.3%, while its own ablation attributes most of the real-world gain to the pretraining corpus rather than the latent bridge [16]. Zero-WAM specifies a new task by an in-context human demonstration video instead of language, generating paired human–robot data automatically (74.2K human–robot ICL pairs, 8.6K tasks) and reporting real-robot unseen-configuration trials without any test-time training [17]. Hydra-0 conditions a video world model on action flow — sparse, visibility-aware image-plane trajectories of the robot's body — so one interface trains across embodiments and backbones, and reports a 90% cut in gripper-motion error within a fixed backbone plus an open-loop policy-evaluation correlation of r = 0.96 against reference success rates [18]. 4DGS-WAM replaces pixels with an explicit decomposed 4D Gaussian state, predicting per-actor SE(3) actions and transporting only object splats; it beats two video world models at a given camera on KITTI-MOT (PSNR 18.80 vs 17.46) but loses at the predicted camera and is evaluated for short horizons only [19]. Finally, AcrossWAM1.0 is the month's engineering paper: it makes the policy–world–action boundary explicit in a latent WAM stack, shows a 0.8B backbone can replace a 2B one within a prespecified two-point criterion (97.45% vs 98.00% on 2,000 paired LIBERO episodes, McNemar p = 0.27), and exports a checkpoint with 42.4% fewer inference-reachable parameters — while flagging a 27.4% instruction-adherence weakness in its own language grounding [20].

Table 3. Scaling axes and headline results, as reported by the authors.

Takeaways

• Scale claims in this theme are real-robot-heavy and correspondingly strong on transfer (ZimaBlue, Riemann-1.0), but the comparisons are usually against VLAs with different pretraining budgets and nobody reports a matched-compute control.
• "One checkpoint, many interfaces" is now a common claim; in this month's set only the policy interface is quantitatively validated, with simulator fidelity left qualitative or unmeasured.
• Human video is the shared scaling currency — in-context prompts, latent dynamics, action flow and unified action spaces are four different ways of turning it into supervision.

4 Outside the table-top: driving, navigation and interactive worlds

Manipulation dominates the corpus, but the same template is being ported to domains with different failure modes — and those ports are where the month's most interesting negative controls are.

4.1 GeoWAM: forecast geometry, not pixels, for driving

GeoWAM's premise is that a driving WAM's useful future is a future point map, not a future image [21]. The model encodes multi-view history, keeps a multi-level historical memory, and has a geometry branch that decodes a future geometry token into a point map plus a planning branch that produces a trajectory, with the pose decoder cross-attending to cached geometry keys.

Figure 11. GeoWAM: multi-view geometry encoding feeding a future-geometry branch and a planning branch that share a pose/geometry decoder through cached keys [21] (reproduced from the original paper).

The evidence is unusually clean because the geometry claim can be tested without labels: predicting future geometry on nuScenes without ever training on nuScenes gives Abs Rel 0.202 against 0.325 for a dedicated geometry world model and 0.274 for geometry reconstructed from a video world model, and the planning ablations are causal — removing future-geometry prediction costs 1.0 EPDMS on navtest and 4.9 on navhard, and navhard EPDMS rises monotonically with forecast horizon [21]. Zero-shot nuScenes planning (L2 0.89 m, collision rate 0.12%) is reported without nuScenes fine-tuning. The caveat is scope: all planning numbers are open-loop or pseudo-closed-loop NAVSIM variants, and the paper's own strongest claims concern a single benchmark family.

4.2 Hydra: search inside the world model's own manifold

Hydra makes the planner and the model share one embedding [22]. Visual states, poses and actions are encoded together and then compressed by modality-specific VQ codebooks (2048 visual codes; 64 pose and 64 action codes), so planning searches a finite set of learned kinodynamic intents and scores them with a latent kinematic-perceptual cost — never decoding to pixels — before the winning intent is expanded into smooth commands by conditional flow matching.

Figure 12. Hydra: intent seeding, VQ codebook quantization, top-k selection, KPC ranking and flow-matching decode, iterated N times [22] (reproduced from the original paper).

The reported separation between mechanism and capacity is the point: at the same search budget, plan time is ≈0.9 s against >500 s for the continuous-latent baseline it is compared to, and obstacle-scenario success is 100%/80%/80% (unobstructed/occluded/blind corner) against 0/0/0, 40/0/0 and 20/10/10 for the baseline, its own deterministic base and a no-codebook variant respectively [22]. KPC ablations show each latent cost term is needed (mean collisions 0.20 → 0.40–0.70 when one is removed). Three limits matter for reading the headline: 10 trials per scenario, off-board inference on a workstation, and the comparison runs the baseline under Hydra's own search budget and hyperparameters rather than its tuned configuration.

4.3 GameWAM: the WAM label reaches games and GUIs

GameWAM applies the joint video-action template to native first-person game control and GUI control, with parallel causal video and action diffusion transformers coupled through a shared clean visual prefix, and deploys "predict long, execute short" so that future-video denoising is omitted online [23]. Its reported gains are large in absolute terms — 50.7 (Mini) / 46.6 (All) average success on a >800-task Minecraft suite against 36.8/31.5 for OpenHA, with 138 executed steps per success against 287 — and its ablations are pointed: removing future-video supervision drops average success from 50.7 to 35.7, and the modality-decoupled mask raises online frequency to 12.51 Hz from 8.12 Hz [23].

Figure 13. GameWAM's video–action transformer pair: joint attention over a shared clean visual prefix in training, memory-only video state at inference [23] (reproduced from the original paper).

The more transferable contribution is diagnostic: the paper isolates a source-sensitivity failure mode of generative action policies, where the dominant DCT band of the action output follows the conditioning source rather than the instruction — donor-swapping that band moves the output in 94.8% of trials and zeroing it removes 99.25% of the associated variance [23]. Headline numbers depend on protocols re-used across papers, and per-map reward values live only in a figure; that is a reminder to treat cross-paper game-agent comparisons cautiously.

4.4 Driving, aerial and interactive variants

The same month produced a dense driving cluster with a shared benchmark and a shared caution. SimWAM reports NAVSIM navtest PDMS 91.9 with a single front camera and shows its isolated attention mask cuts trajectory inference from 1512 ms to 518 ms at essentially equal PDMS, while its co-training ablation (86.5 → 90.3 with future-video supervision, 91.9 with RL) is a clean statement of the future-supervision effect [24]. DA-WAM makes the future candidate-specific and finds that a shared global future is worse than no future prediction at all (92.81 vs 93.31 PDMS), the sharpest negative control in the driving set [25]. BrainWAM reports that letting VLM semantic tokens share attention with the predictive branch hurts relative to the world-model-only branch (87.8 vs 89.5 PDMS) and that zero video-denoising steps collapses PDMS to 79.3 while one step restores 89.3 [26]. WA-JEPA retargets V-JEPA's masking objective at future prediction and reaches 91.7 EPDMS on NAVSIM-v2 with a zero-shot HUGSIM closed-loop HD-Score of 0.4462 against 0.3252 for the strongest baseline [27].

Beyond cars, FlowPilot ports the paradigm to a quadrotor with a depth-conditioned mixture-of-transformers and reports 4.0% collision rate against 26.0% for a depth-frozen control, 16.3 ms mean onboard inference on a Jetson Orin NX, and forest flight up to 5.5 m/s [28]. UniNav jointly denoises future video latents, continuous waypoints and camera-geometry tokens, reporting an average ATE of 0.491 against 0.593 for the strongest navigation world model and a 0.109 s one-step path against a ~900 s planning baseline — but every main result is offline trajectory error on pre-collected datasets, with no closed-loop navigation success metric [29]. On interactive rendering, ReWorld trains control and long-horizon memory under different attention windows and shows the interference directly (adding action injection improves control but drops revisit SSIM 0.3898 → 0.3376; per-head routing restores recall while keeping unrouted-level control) [30], while Matrix-Game 3.5 builds a geometry-aware patch-level memory and reports 1.63° camera-rotation error against 4.50° for the next best, best revisit SSIM, and real-time streaming on one H100 after three-step distillation — with the honest caveat that its geometry comes from an external estimator and dynamic entities have no persistent state [31].

Table 4. Domain ports and their distinctive mechanisms, as reported by the authors.

Takeaways

• Ports to other domains are converging on NAVSIM-style open-loop or pseudo-closed-loop scoring; only Hydra, FlowPilot, GameWAM and ZimaBlue-class systems report physical deployment, usually with small trial counts.
• The driving cluster produced the month's best negative controls: shared futures, premature semantic fusion and over-aggressive denoising all hurt, which is useful guidance for anyone assembling a WAM stack.
• Terminology drift is real: several of these papers (GameWAM excepted) never use "world action model", and the interactive-rendering systems control cameras rather than infer actions.

5 The field starts auditing itself

The least expected change in August is how many papers ask whether the WAM's central object actually works as advertised. Four distinct audits appeared in the window, and their results are consistent: rendering quality and behavioural fidelity are only loosely coupled.

5.1 WorldExam: capability, not appearance

WorldExam is a four-level diagnostic hierarchy — visual quality, control adherence, spatial consistency, world reactivity — instantiated as eight tasks over 1,474 curated cases and applied to 20 video world models [32]. Its design choice is to withhold the consequence of an interaction from the prompt, so a model that merely renders what it was told cannot pass; the reported result is a sharp dissociation between control and reactivity: action-driven models lead subject control (LingBot-World 55.47 vs Veo 3.1's 37.28) while language-driven models lead terrain and interaction tasks by a wide margin (Vidu Q3 64.39 terrain against the best action-driven 27.49), and the VLM judge tracks human scores at ρ = 0.8614 over 800 instances.

Figure 14. WorldExam: the four diagnostic levels mapped to their evaluation tasks, metrics and representative cases [32] (reproduced from the original paper).

The caveat is definitional rather than empirical: WorldExam probes the "world" half of a WAM and never tests action-space fidelity, policy success or closed-loop control, and no evaluated system calls itself a WAM — so its headline capability split is partly an artefact of which interface each model exposes [32].

5.2 Do robotic world models really follow actions?

The most directly WAM-relevant audit asks whether action-conditioned world models respond faithfully to arbitrary valid actions, not just replayed demonstrations [33]. Its protocol quizzes six expert-trained models across five action-query categories, from demonstrated replay to increasingly off-expert actions, and scores rollouts by a visual-integrity gate plus SE(3) end-effector agreement against simulator replays of the queried actions. The diagnosis is consistent: integrity-gated error rises by 0.029–0.099 m from demonstrated to off-expert queries, with visual-failure rates up 6.3–28.1 points. The paper then offers a recipe — expanded action coverage plus intervention-effect supervision — which the authors report raises simulated policy improvement from 51–52% to 65% against 56–57% for the same improvement procedure on the closest baseline, and real-robot success from 48% to 68% [33].

Figure 15. The WorldSync recipe: action coverage expansion, the action-forcing expert and intervention-effect supervision, all feeding one action-conditioned world model [33] (reproduced from the original paper).

The balance of evidence is asymmetric and worth stating plainly: the diagnostic is broad (six models, 50 tasks) while the improvement claim rests on a four-task ablation, one real-robot task and unmatched update counts; the paper's headline win on gated error is marginal (0.0661 vs 0.0670) and it trades that against a worse raw trajectory error [33].

5.3 WorldSimProbe: judge the model as a simulator

WorldSimProbe makes the simulator role explicit with an "observable simulator contract" — supplied actions must induce the corresponding agent motion, and environment responses must be grounded in that realized motion — operationalized as five suites over more than 18,000 instances on three platforms [34]. Its most useful finding for WAM work is a threshold asymmetry: the perturbation magnitude at which a model's behaviour persistently breaks is 0.374 for one leading system and 1.78 for another, against a simulator boundary of 0.05, i.e. both tolerate far more action distortion than the reference simulator does; false-contact grounding is the weakest trigger (macro-mean 38.6 against 75.0 for distractors), and policies trained on generated data reach 78–89% under standard controls but 21–53% under out-of-distribution controls, ordering consistently with the benchmark [34].

Figure 16. WorldSimProbe's five suites, from local action calibration through global coverage and source preservation to interaction grounding and dynamics, with RMFA-based scoring [34] (reproduced from the original paper).

5.4 Hidden state, intervention fidelity, and what the theory says

Three further audits close the loop. An action-conditioned visual Shell Game shows that standard video backbones — transformers, Mamba2, linear attention, RNNs — render plausible frames yet fail to compose hidden swaps beyond the training horizon, that more denoising steps do not help, and that two specific state-carrier changes do (a reflection-capable linear attention reaches 0.95 state accuracy at 30 swaps against chance-level for the backbone; a fast-weight TTT layer scores 1.00/1.00/0.98/0.69 at 5/10/20/30 swaps) [35]. A capture-gated intervention audit shows that a world model can carry the task query and resolve real effects while its own five-step imagined effect is worse than predicting no effect — predicted-effect R² of −5.65/−3.36/−6.33 for three checkpoints whose capture R² is 0.58–0.61 — and that the reward channel is silent about it, while the size of a TD-MPC2 model ranks against episode return by operator error at ρ = −0.90 [36]. And a theory paper argues that imitated world-action controllers flatten into ordinary stochastic policies with the same closed-loop distribution, with worst-case regret ≥ 1/4 for observational learners and 1/2 for the exact optima, unless an informative intervention is available [37]; it reports no experiments, and concedes that finite parametric classes may still enjoy a representational advantage. A companion position paper reframes world modelling as an agent-facing information proxy and offers a taxonomy slot for WAMs without making WAM-specific claims [38].

Table 5. The month's audits: the question each asks and its key reported finding.

Takeaways

• The audits agree that visual fidelity is a poor proxy for what a WAM needs: state tracking, action following and intervention fidelity all fail at levels invisible in rendering metrics.
• Two of the four audits were run on the authors' own protocols against frozen third-party models, which is much stronger evidence than a new benchmark alone.
• Nothing here says WAMs do not work; it says the field's success-rate style of evaluation does not yet show that they work for the reason claimed.

Also noted

Efficiency and post-training variants. DELE-w0.5 makes the strongest anti-video-generation claim of the month — that video generation is an unnecessary intermediate objective — and infers actions from a future latent state instead; it reports 62.5% full-task success against 30.0% for the best baseline on four long-horizon dual-arm tasks, with the advantage growing with horizon and an 87.5 ms median inference latency [44]. Vid2WAM [12] and Faster-WAM [5] pursue the same goal by distillation and depth reduction respectively.

Contact-rich manipulation. HiTac-WAM attaches a tactile-forecasting branch to a frozen WAM and uses the forecast both as a candidate reranker and as an execution-time verifier, reporting 72.2% mean real-robot success against 31.1% for the same backbone without the tactile pathway at a cost of ~1.4% LPIPS and ~0.9% action error [42]. Disentangling Visuo-Tactile Foresight takes the complementary diagnostic route, replacing the learned predictor with an oracle demonstration to isolate the interface through which a visuo-tactile future reaches the action expert; its same-budget comparison puts the routing interface 8.3 points above an independent-fusion arm on 350 rollouts, a result the analysis flags as a hypothesis worth replicating because of a task-level inversion [43].

Systems and orchestration. HarnessWAM argues that finite-horizon WAMs cannot carry planning, occlusion state, verification and recovery, and adds these as a model-external layer — a VLM task manager, an evidence-grounded state, and a compiler that projects plans onto skills the WAM can execute; on a 26-task memory arena it reports 64.4% full-task success against 38.5% for the strongest benchmark-specific baseline, with gains of +15.4 to +26.0 points on the same backbones and 95.0% on one physical tidying task [41].

Cross-embodiment simulators and navigation representation studies. CLAP harmonizes three action interfaces (absolute end-effector poses, language actions, learned latent actions) in one action-conditioned video model and reports zero-shot planning on a real Franka that beats two strong policies (80% vs 40% and 75% on two tasks), while conceding that perceptual-metric parity with the single-embodiment state of the art comes with worse FVD/FID and that absolute actions beat relative ones for cross-embodiment end-effector conditioning [39]. A navigation study on representation and history choices reports that reconstruction quality is anti-correlated with prediction quality in its framework (lowest reconstruction error 0.069 for one representation family, highest 0.335 for another), that the best representation flips between direct prediction and long rollouts, and that cached-history attention cuts latency and memory by roughly 7× — all in open-loop evaluation only [45].

Benchmarks. PlayWorld evaluates nine interactive video models with a closed-loop agent player that adapts its actions online instead of executing a fixed trajectory; the best model scores 2.12 out of 5 overall, and the two evolution dimensions never exceed 1.95/1.81, with one model passing 80.4% of trajectory validation yet scoring 1.48 on the rubric — the clearest statement that reaching the goal and preserving world state are different abilities [40].

Open questions to watch

• Matched-compute baselines. Every scaling claim this month (video hours, corpus size, distillation) is compared against systems with different pretraining budgets; nobody reports spending the same compute on robot data instead.
• Does the future channel survive third-party evaluation? The strongest mechanism evidence is internal ablation. The audit papers suggest a cheap external test: recompute action-following and hidden-state metrics on the published checkpoints.
• Deployment arithmetic. Reported latency gains are benchmark-specific and often exclude decoding, guidance, VAE and networking; the field still lacks a shared measurement protocol for a "real-time WAM".
• Naming and deduplication. Within the screened set, several distinct August papers share a name (two 4D-WAM, two Faster-WAM, two JEPA-WAM, two SG-WAM), and by mid-month the "-WAM" suffix spans manipulation, driving, navigation, games, surgery, radio and aerospace. Reading the papers rather than the titles is currently unavoidable; a stable taxonomy would help.

Scope and method

Window and sources

This issue covers world action model (WAM) work first submitted in August 2026 (v1 dates 2026-08-01 to 2026-08-31), read from the local library rather than from arXiv listing dates. It is the first issue of this series, so it is a first-run baseline: there is no previous issue to compare against, and "new" below means new in the August window, not new to the field as a whole.

WAM has no single settled name in the literature, so the candidate pool was built from four routes over the window: (a) the literal name variants ("world action model(s)", "world-action model(s)", "action world model") in title and abstract — 66 hits; (b) the acronym WAM as a whole word — 56 hits; (c) papers whose title contains "world model" / "world-action" / "latent dynamics" / "future imagination" together with an action/policy/robot/agent term — 183 hits; and (d) a semantic route phrased as the WAM definition (predicting how the environment evolves in response to actions and using that to act), top 300 (reached; similarity fading from 0.59 to 0.37). Union after de-duplication: 341 candidates.

Screening

The 341 candidates were screened in bulk against an explicit WAM boundary: include if the core contribution is a learned model of environment evolution coupled to action — predicting future world state (pixels, latent, geometry, structured state) conditioned on actions, and/or deriving actions, policies or plans from predicted futures for an embodied agent; exclude non-action video/3D generation, vision-language-action policies with no predictive dynamics, world models for non-embodied text/web agents, agent-trajectory-only forecasting, and acronym-only matches. 204 of 341 passed (137 excluded). Screening used titles and abstracts, calibrated on two trials before the full run; the criterion is fixed in the screened working set, so counts here describe these checked routes, not the size of the field.

Reading depth

Forty-eight papers were then read at full-text depth in background subtasks: methods, experiments and stated limitations, with per-paper evidence notes recording the authors' reported numbers and their settings. Twenty-five analyses completed in the first pass and the twenty-three that hit model-call failures were re-run; all 48 are complete. That deep-reading set is the source of every featured item and of the "also noted" entries below, and the fifteen expansions in this issue exceed a normal tracking brief because the task behind it asked for a dozen-plus expanded entries over a read of at least thirty papers. Featured items below are marked as checked in the original (read at full-text depth) — every number quoted comes from the paper's own tables or text and keeps its setting (benchmark, task, protocol); where a comparison mixes heterogeneous backbones or data, that is said so.

Gaps and limits

• Screening is abstract-level: a paper can be included or excluded on claims its experiments do not fully support, and the counts are not a field census.
• The semantic route reached its top-300 cap and its tail (roughly beyond rank 170) drifts toward generic VLA, trajectory prediction and embodied-agent security; adjacent work outside the WAM framing is therefore under-represented by design, not proven absent.
• Benchmarks and success-rate numbers come from different simulators, embodiments and demo budgets. Cross-paper success percentages in this issue are indicative, not directly comparable, unless a paper reports a matched comparison.
• Where a claim rests on qualitative figures (for example "policy and simulator in one checkpoint"), this issue says so instead of reporting it as a measured result.

References

[1] Linhan Wang, Zijian An, Mingyuan Zhang, et al. GlanceWAM: Sparse Test-Time Imagination for World-Action Models. 2026. arXiv preprint. arXiv:2608.23927

[2] Jiakai Huang, Zhongbo Wu, Siyu Xu, et al. Foresight Without Seeing: Latent Futures for World Action Models. 2026. arXiv preprint. arXiv:2608.11605

[3] Liuhaichen Yang, Zhengyang Zhong, Hanshang Zhu, et al. WAM-OPD: Joint Video-Action Supervision for World Action Model Post-Training with On-Policy Distillation. 2026. arXiv preprint. arXiv:2608.22364

[4] Hongbo Lu, Liang Yao, Chenghao He, et al. RISE: Adaptive Imagination for World Action Models. 2026. arXiv preprint. arXiv:2608.20430

[5] Liheng Ma, Rui Heng Yang, Amin Abyaneh, et al. Faster-WAM: Do World Action Models Need Deep Action Modules? 2026. arXiv preprint. arXiv:2608.02365

[6] Jingkai Wang, Zihan Tang, Gu Zhang, et al. SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation. 2026. arXiv preprint. arXiv:2608.09771

[7] Shanglin Yuan, Weiheng Zhao, Xin Shi, et al. DreamWAM: Beyond RGB Future Prediction for World Action Models. 2026. arXiv preprint. arXiv:2608.04996

[8] Ge Yan, Jinghao Liu, Yuzhi Fan, et al. Flex-pi: A Multi-Stream World-Action Model with Compute Flexibility. 2026. arXiv preprint. arXiv:2608.10860

[9] Yihan Lin, Jiawei He, Shifeng Bao, et al. JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling. 2026. arXiv preprint. arXiv:2608.09381

[10] Zijian Zhang, Yuqing Jiang, Weitao Zhou, et al. GaussianWAM: Distilling Geometry and Semantics from 3D Gaussian Fields into World-Action Models. 2026. arXiv preprint. arXiv:2608.24714

[11] Xiaoxiao Lu, Yunlong Dong, Jiahao Shi, et al. Making Latent Evolution Explicit: Operator-Structured Transitions for World Action Models. 2026. arXiv preprint. arXiv:2608.27259

[12] Chenhao Qiu, Ruixiang Wang, Runyi Zhao, et al. Vid2WAM: Distilling Video Diffusion Priors into World Action Models. 2026. arXiv preprint. arXiv:2608.08558

[13] Xionghao Wu, Yijun Yang, Shiyang Zhou, et al. ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training. 2026. arXiv preprint. arXiv:2609.00188

[14] Haofeng Sun, Jiangbo Pei, Fei Kang, et al. Riemann-1.0: An Embodied World Action Model for Physical AI. 2026. arXiv preprint. arXiv:2608.27033

[15] Hongzhe Bi, Zihao Zhou, Yihang Tang, et al. Motus2: A Self-Evolving General World Model for Dexterous Manipulation. 2026. arXiv preprint. arXiv:2608.30237

[16] Zhenhao Shen, Jiaqi Liang, Jasper Lu, et al. LD4WAM: Learning Latent Dynamics from Human Videos for World Action Models. 2026. arXiv preprint. arXiv:2608.22403

[17] Jiaming Zhou, Qihang Zhang, Gangwei Xu, et al. Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization. 2026. arXiv preprint. arXiv:2608.26103

[18] Hongyu Li, Bowen Wen, Xinghao Zhu, et al. Hydra-0: Action Flow for Generalist World Modeling and Control. 2026. arXiv preprint. arXiv:2608.18077

[19] Yueen Ma, Zenglin Xu, Irwin King. 4DGS-WAM: Bridging Past and Future with an Object-Centric World Action Model based on 4D Gaussian Splatting. 2026. arXiv preprint. arXiv:2608.25956

[20] Yafei Zhang, Nan Wu. AcrossWAM1.0: A Modular Latent World-Action Stack for Compact Robot Policies. 2026. arXiv preprint. arXiv:2608.29937

[21] Yiren Lu, Xin Ye, Jiaming Liu, et al. GeoWAM: Visual Geometry World Action Models for Autonomous Driving. 2026. arXiv preprint. arXiv:2608.23486

[22] Mohammad Nazeri, Alexandyr Card, Samira Huber, et al. Hydra: A Navigation World Action Model with Discrete Latent Planning and Continuous Flow-Matching Execution. 2026. arXiv preprint. arXiv:2608.28995

[23] Yuncheng Guo, Zhanqiu Zhang, Yiwen Guo, Weijia Li. GameWAM: A World Action Model for Video Games. 2026. arXiv preprint. arXiv:2608.26200

[24] Zongchuang Zhao, Xin Zhou, Tianyang Xu, et al. SimWAM: A Simple World Action Model for End-to-End Autonomous Driving. 2026. arXiv preprint. arXiv:2608.07468

[25] Ruiguo Zhong, Benshan Ma, Xiaolong Chen, et al. DA-WAM: Decision-Aligned Future Latents for Driving World Models. 2026. arXiv preprint. arXiv:2608.19085

[26] Bing Zhan, Shuyao Shang, Shuo Lu, et al. BrainWAM: Action-Space Coordination of Semantic Priors and Predictive Dynamics for Autonomous Driving. 2026. arXiv preprint. arXiv:2608.12854

[27] Xinlin Wang, Yujiao Xiang, Yuheng Zhou, et al. WA-JEPA: Rethinking the Video JEPA Paradigm for World-Action Modeling in Autonomous Driving. 2026. arXiv preprint. arXiv:2608.20974

[28] Runqing Wang, Ding Yu, Pengyuan Min, et al. FlowPilot: Real-Time World-Action Modeling for Agile UAV Navigation. 2026. arXiv preprint. arXiv:2608.00635

[29] Changqing Zhou, Yueru Luo, Zeyu Jiang, Changhao Chen. UniNav: A Unified World-Action Diffusion Model for Visual Navigation. 2026. arXiv preprint. arXiv:2608.03244

[30] Zhifei Chen, Luozhou Wang, Guibao Shen, et al. ReWorld: An Interactive World Model with Long-Horizon Memory. 2026. arXiv preprint. arXiv:2608.23565

[31] Runjia Qian, Zile Wang, Jihai Zhang, et al. Matrix-Game 3.5: Enhancing Real-Time Streaming Interactive World Models with Patch Memory. 2026. arXiv preprint. arXiv:2608.29910

[32] Yuxue Yang, Shuyao Shang, Jiahe Wang, et al. WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity. 2026. arXiv preprint. arXiv:2608.02603

[33] Sixiang Chen, Jiaming Liu, Jixian Wu, et al. Do Robotic World Models Really Follow Actions? Diagnosing and Aligning Action-Conditioned Generation for Policy Learning. 2026. arXiv preprint. arXiv:2608.24885

[34] Peterson Co, Sicheng Hu, Chunxuan Jiao, et al. WorldSimProbe: Diagnosing Simulator Faithfulness in Action-Conditioned World Models for Embodied Manipulation. 2026. arXiv preprint. arXiv:2608.09298

[35] Joonghyuk Shin, Yicong Hong, Jaesik Park, Xun Huang. Can Video World Models Track Unobserved World States? 2026. arXiv preprint. arXiv:2608.30692

[36] Donna Vakalis. The Intervention Gap in Latent World Models. 2026. arXiv preprint. arXiv:2608.29998

[37] Yang Yu. On the Capability Separation Between World-Model Policy Learning and Imitated World-Action Models. 2026. arXiv preprint. arXiv:2608.22197

[38] Yu Yang, Xuemeng Yang, Licheng Wen, et al. Quo Vadis, World Modeling? 2026. arXiv preprint. arXiv:2608.02713

[39] Kechen Liu, Ola Shorinwa. CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators. 2026. arXiv preprint. arXiv:2608.27406

[40] Kaixin Ding, Xi Chen, Minghong Cai, et al. PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives. 2026. arXiv preprint. arXiv:2608.13552

[41] Zhaopeng Gu, Bingke Zhu, Tianxi Lin, et al. HarnessWAM: Bridging Prediction and Deliberation in World Action Models. 2026. arXiv preprint. arXiv:2608.09516

[42] Chao Xue, Chaofan Zhang, Wenxuan Ma, et al. HiTac-WAM: A Hierarchical Tactile World Action Model for Contact-Rich Robot Manipulation. 2026. arXiv preprint. arXiv:2608.19574

[43] Zihang Yao, Chaoyue Ding, Yingying Yu. Disentangling Visuo-Tactile Foresight: Oracle-Guided Interface Discovery for World Action Models. 2026. arXiv preprint. arXiv:2608.00547

[44] Fenghao Lei, Zhixiong Huang, Long Yang, et al. DELE-w0.5: Inferring Action from Future Latent State for Robotic Manipulation. 2026. arXiv preprint. arXiv:2608.22067

[45] Guangfu Guo, Xiaoqian Lu, Rui Liu, et al. Visual Representation and History Modeling for Navigation World Models. 2026. arXiv preprint. arXiv:2609.29555