Pith.

Robotics · arXiv:2610.28122 v1 · submitted 20 Oct 2026 · 24 references

De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control

Shaoyi Zheng, Zhongtao Jiang, Yue Peng, Wenbo Lu, Shunchi Zhang, Xi Wang, Shengji Wan

The paper says

“Cascade world-action models (WAMs) use predicted visual futures to condition robot actions, but inherit costly iterative inference from diffusion-pretrained experts. We ask what a single-step predicted future must preserve to support accurate single-step action prediction.”

more of the abstract

In the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both streams to one step substantially reduces task success. Video-space supervision only partially closes this gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. These findings motivate DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling through action-sufficient distillation, while retaining the explicit future-to-action cascade. After learning a one-step action model, we train the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert, with an optional future-latent reconstruction loss. The resulting objective evaluates the student-generated future through its downstream control consequences rather than relying on representation matching alone. DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget. This advantage has also extended to three real-world tasks that we designed. Compared with the full iterative teacher, it reduces inference latency by 28.0 ×.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 21 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Central claim is supported by well-matched experiments; the genuine boundary is that action-sufficient transfer is validated through one frozen action expert, which the authors only partly own. This line is wrong

2 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 24 references say.

The paper says, at §3.3, Eq. (2); §5 (b193/b196)

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The entire Stage-2 objective (Eq. 2) is gated through the frozen single-step action expert G_psi, which is itself trained in Stage 1 under teacher forcing on demonstrated futures, and the paper validates the resulting transfer for exactly one realization of that conduit: all task-level evidence in Tables 2, 3, and 8 passes through this same G_psi.

more of objection 1

The authors do state in Section 5 (b193, b196) that action sufficiency is relative to the downstream mapping, and that caveat is genuine; however, the abstract, the central-claim sentence (b158), and the Section 6 conclusion present the principle without noting that only a single instantiation of the downstream mapping has been tested, and no experiment varies the action expert. This untested dependence is a genuine boundary of the claim rather than an implementation detail: a differently initialized or differently trained single-step action head (for example, a Stage-1 checkpoint from a different seed, or an action expert trained with a different teacher-forcing schedule) could in principle make the action-error gradients a misleading training signal. I recommend (i) an explicit boundary statement in the abstract or Section 1 and Section 6, and (ii) a robustness check in which Stage 1 is re-run with a different seed (or a different stage-1 checkpoint is used) and the Stage-2 advantage over vision-supervised learning is re-measured on RoboTwin; if the advantage is insensitive to the action-expert realization, that evidence would substantially strengthen the general claim.

The paper says, at §3.2, Fig. 3b; App. C.2–C.3, Tables 7–8

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The sensitivity analysis that motivates action sufficiency measures 'action-sensitive' directions through the local Jacobian of the same frozen single-step action expert through which Eq. (2) is optimized (Eqs. 1, 15, and 18). Because Stage 2 explicitly moves the future predictor along the action-error gradients of this model, the finding that a component holding 0.63–0.85% of the discrepancy energy recovers 19.8–34.0% …

more of objection 2

… of the offline action-error improvement is partly produced by the training procedure itself: the selected component is, by construction, aligned with directions in which this particular G_psi is sensitive. The closed-loop success rates in Table 8 (78.2% to 87.3% from 1.4% energy) are external to the training loss and mitigate this concern substantially, and Appendix C.3 does state that the construction of P favors sensitivity (b401); but the shared-model circularity is not acknowledged anywhere in the paper. To make the mechanistic claim independent, the decomposition should be re-run with a probe that did not participate in training, such as an independently initialized Stage-1 action expert or a single-step action head trained with a different objective, and the energy/recovery imbalance re-reported; alternatively, the paper should explicitly re-scope the diagnostic as a description of the trained system rather than evidence about the general structure of future-to-action mappings.

6 minor objections

The paper says, at Abstract; §4.2, Table 2

Pith says: minor objection 1 of 6

The abstract's 'outperforming Flash-WAM at the same one-evaluation-per-stream budget' is stated without the qualifier that the LIBERO Flash-WAM numbers are quoted from the original paper rather than produced by a matched reproduction (Table 2, dagger, b143); the matched head-to-head evidence is RoboTwin only, and the LIBERO claim should be qualified in the abstract.

The paper says, at Fig. 3b caption vs. App. C.3

Pith says: minor objection 2 of 6

The energy and recovery ranges for the teacher-referenced sensitivity diagnostic are inconsistent between the main text and the appendix: Fig. 3b reports 0.63–0.85% energy and 19.8–34.0% recovery, while Appendix C.3 reports 0.31–0.85% energy and 16.3–34.0% recovery; these numbers should be reconciled.

show 4 more

The paper says, at §4.2, Tables 2–3; Table 9

Pith says: minor objection 3 of 6

The headline success rates (Table 2) and visual-quality metrics (Table 9) are reported without confidence intervals, even though Appendix A.3 shows the evaluation pipeline supports bootstrap intervals; adding intervals to Table 2 (especially the RoboTwin 85.4 vs. 78.2 comparison) and to Table 3 would make the head-to-head claims easier to calibrate, given that the real-world trials use only 20 attempts per cell.

The paper says, at Table 9

Pith says: minor objection 4 of 6

DirectWAM exceeding the full 25-step teacher on PSNR (31.79 vs. 30.37) and SSIM (0.949 vs. 0.938) is surprising for a model trained primarily on action error; one or two sentences of explanation (for example, smoothing effects or the effect of the anchor loss) would preempt skepticism, particularly because the FID/KID metrics still favor the teacher.

The paper says, at §3.1, Table 1; App. A.1

Pith says: minor objection 5 of 6

The attenuation statistics in Table 1 are computed with the learned single-step action head rather than the teacher's native multi-step action sampler, a fact stated only in Appendix A.1; since the abstract's 'resampling video noise has little effect' is a statement about the evaluated WAM, the main text should state which action pathway each statistic uses, given that the closed-loop Table 5 interventions do cover the teacher's own pathway.

The paper says, at Throughout

Pith says: minor objection 6 of 6

The names 'DirectWAM', 'LingBot-VA', and 'WAMs' appear with inconsistent spacing ('DirectW AM', 'LingBot-V A', 'W AMs'); this should be cleaned up in the camera-ready version.

What would settle it

Retrain DirectWAM with the Stage-1 action expert varied, e.g., (a) a different seed or slightly different Stage-1 training budget, and (b) an action expert trained from a differently initialized checkpoint while keeping the video expert, data, and Stage-2 budget fixed; report RoboTwin success of the resulting DirectWAM. If the 85.4% result moves by more than a few points across these conduits, the headline does not support a general action-sufficiency principle for cascade WAMs; if it is stable, the conduit concern is resolved for the tested class.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 21 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 12 s · 19,531 tokens read, 1,535 written
  3. Editor wrote the verdict above 51 s · 20,486 tokens read, 7,676 written
  4. Referee major revision: 2 major objections, 6 minor 1 min 31 s · 19,791 tokens read, 15,308 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Two things to know. The central claim survives the right experiments: a one-step future for this cascade WAM should be learned through the action model that consumes it rather than through representation matching. On RoboTwin 2.0 the matched comparison gives DirectWAM 85.4% vs Flash-WAM's 78.2% over roughly 2,500 episodes (the gap is several standard errors), and the controlled ablation holding initialization, data, budget, and the frozen action expert fixed puts vision-supervised students at 76.0-78.2% against DirectWAM's 87.0%. That ablation isolates exactly the factor the paper claims matters. And the closed-loop selective-replacement test is the strongest single piece: with the action model fixed, transferring 1.4% of the prediction-difference energy along locally action-sensitive directions recovers 87.3% success versus 87.0% for full replacement. Because the headline numbers are closed-loop success on external benchmarks rather than outputs of the training loss, the circularity burden is genuinely low. The attenuation analysis (video-side sampling variance dropping from 6.6% to 0.03% of total variance in action space) is also a real empirical characterization, and the citation lineage to objective mismatch and value-aware learning is honest.

Soft spots, in proportion. The stress-test concern lands but lands softly. Stage 2 is gated through one teacher-forced single-step action model, and the sensitivity analyses share that same model as both training conduit and evaluation probe, so the paper cannot say how the recipe transfers to a differently trained or initialized action expert. The authors do state that action sufficiency is relative to the downstream mapping, so they acknowledge the relativity — but they don't test the robustness, and for a method pitched as a model-conversion principle, one instantiation is thin. A second action-expert realization would materially strengthen the revision. Missing code and weights is a real cost for a deployment-oriented paper. The real-world evidence is twenty trials per task on three self-designed tasks; the pouring gap (11/20 vs 2/20) is suggestive but noisy. The quoted LIBERO numbers are properly flagged with a dagger, so the cross-benchmark claim rests on RoboTwin, which is matched. One thing I'd ask a referee to chase: Table 9 shows DirectWAM beating the 25-step teacher on PSNR/SSIM — plausible as regression-to-the-mean inflating pixel metrics, and the FID still favors the teacher, but it deserves a check.

Who it's for: anyone working on WAM acceleration or control-oriented representation learning. It deserves a serious referee. Send it out, and push on the action-expert generality boundary and on releasing code.

The strongest objection

The central claim (b158) is that a future should be learned through its downstream action consequences, and DirectWAM's headline results (b7) rest on Eq. 2, where G_psi is a single-step action model trained in Stage 1 under teacher forcing on demonstrated futures, then frozen. The load-bearing assumption is that this particular G_psi is a faithful, smooth conduit: action-error gradients propagated through it shape futures that transfer to closed-loop success. The paper validates this conduit empirically for its one instantiation (Table 2, Table 8, real-world trials), but it does not characterize when the conduit assumption fails. Since the analysis in Sec. 3.2 uses the same frozen single-step action model as both training conduit and evaluation probe, the sensitivity measurements (Fig. 3b, Tables 7--8) share the model that the training objective optimizes through; the offline analysis and Eq. 2 therefore cannot independently confirm that action-sufficient futures transfer to a different action-expert realization. In contrast, the closed-loop Table 8 replaces only the future representation while fixing the same action model, so the strongest evidence for task-level transfer still leaves the action-expert dependence untested. The paper's own conditions (b193, b196) state that action sufficiency is relative to the downstream mapping, which makes the untested dependence on which action expert is used a genuine boundary of the claim, not a mere implementation detail.

The reader's weakest_assumption names exactly this: the teacher-forced G_psi is a faithful and smooth conduit, with no bound or failure characterization and no generality test across action-expert architectures. I agree that this is the most load-bearing concern. The paper's controlled ablation (Section 3.2) and closed-loop selective replacement (Table 8) are real supporting evidence, and the dagger disclosure (b143) plus honest limitations (b329--b330, b378, b401) reduce the risk of over-reading ancillary claims. The remaining risk is not internal inconsistency; it is scope: the principle 'learn futures through the downstream action model' is demonstrated only for one instantiation of that downstream model, and the diagnostic probes used to justify the principle (Eq. 1, Eq. 15) are computed from the same frozen model that defines the Stage-2 loss (Eq. 2). A single replication of DirectWAM under a varied Stage-1 action expert would settle whether the mechanism is robust or specific to the reported checkpoint. Since the empirical validation is strong but the mechanism is not generalized, CONDITIONAL is the right verdict; the paper should be accepted with this boundary made explicit in the text.

The referee's report

The empirical core is well designed, and the paper is unusually honest about its own boundaries. The Section 3.2 comparison holds initialization, data, budget, and the frozen action expert fixed across future-learning objectives, isolating the learning criterion; the RoboTwin comparison with Flash-WAM (85.4 vs. 78.2) is a matched head-to-head at the same deployment structure over roughly 2,500 episodes; and the closed-loop selective replacement in Table 8 (78.2% to 87.3% success from 1.4% of prediction-difference energy) is a genuinely falsifiable, externally measured check of the action-sufficiency mechanism. The paper also reports intervention statistics (McNemar p-values, bootstrap intervals) in Appendix A.3, flags the quoted LIBERO Flash-WAM numbers with a dagger (b143), and repeatedly limits the sensitivity probe's interpretation (b378, b401). If the results hold, the contribution is substantial: pretrained diffusion WAMs can be converted into deterministic single-step cascades that keep the explicit, decodable future interface and achieve a 28.0x latency reduction, and 'action sufficiency' supplies a clean design criterion for such conversions.

Pith's reading of the paper (a machine's summary, not the authors')

Cascade world-action models predict a visual future and then decode robot actions from it, but the diffusion machinery that produces the future is slow. This paper asks what a single-step predicted future must preserve in order to keep control accurate, and answers: what the downstream action model needs, not the full visual detail of the teacher's iterative generation. The authors establish this with three observations: resampling video noise barely changes downstream actions, yet naively truncating both streams to one step cuts RoboTwin success from 91.2% to 54.9%; matching future representations in video space recovers only part of the gap (76.0–78.2% versus 87.0%); and correcting a small, locally action-sensitive component holding about 1.4% of the prediction-difference energy recovers essentially the full closed-loop control gain. They then introduce DirectWAM, which distills the pretrained video and action experts into deterministic one-step predictors by backpropagating demonstration-action error through the frozen action expert into the future predictor. DirectWAM reaches 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, ahead of Flash-WAM at the same one-step budget, with a 28.0x latency reduction, while keeping predicted futures visually decodable.

What carries the argument. The carrying mechanism is action-sufficient distillation, a two-stage model-conversion objective. Stage 1 freezes the diffusion video expert and trains a single-step action model $G_\psi$ under teacher forcing on demonstrated futures; Stage 2 freezes $G_\psi$ and trains the future predictor $F_\theta$ with $\mathcal{L}_{\mathrm{train}} = \mathcal{L}_{\mathrm{ASD}} + \lambda_v \mathcal{L}_{\mathrm{anchor}}$, where $\mathcal{L}_{\mathrm{anchor}} = \mathbb{E}_{\mathcal{D}}[\|F_\theta(h_t) - z^\star_{t+1:t+K}\|_2^2]$ optionally anchors the predicted future to the demonstrated latent chunk and $\lambda_v = 0$ gives purely action-supervised learning. The companion diagnostic is a local Jacobian decomposition $\Delta z = z_T - z_B$ into an action-sensitive part $P\Delta z$ and its orthogonal complement, where $P$ is an orthogonal projection built from 32 sampled action-output directions pushed back through the frozen action model; this projection is what quantifies the highly unequal action consequences of equal-magnitude representation errors. The paper's overall argument — that stochastic sampling can be separated from iterative refinement and that both can be replaced by direct, action-supervised computation — rests on these two objects working together: the distillation objective supplies the training signal and the projection supplies the evidence that the signal is pointed in the right direction.

Core claim. The central claim is the principle of action sufficiency: for single-step cascade WAM prediction, the adequacy of a predicted future is determined by the action it enables, not by how faithfully it reproduces the teacher's generative trajectory or the reference future representation. Concretely, DirectWAM first trains a single-step action model $G_\psi$ under teacher forcing on demonstrated futures, freezes it, and then trains the future predictor $F_\theta$ by decoding actions from $F_\theta$'s own predicted futures and minimizing demonstration action error $\mathcal{L}_{\mathrm{ASD}} = \mathbb{E}_{\mathcal{D}}[\ell_{\mathrm{act}}(G_\psi(F_\theta(h_t), h_t), a^\star_{t:t+K-1})]$, with an optional future-latent reconstruction term kept only as an anchor. Because supervision is applied to the student's own predictions through the deployment-time interface, the student never has to reproduce the teacher's denoising trajectory. The paper reports that this objective outperforms Flash-WAM's modality-aware consistency distillation at the same one-evaluation-per-stream budget (85.4% versus 78.2% on RoboTwin; 97.1% versus 95.1% average on LIBERO, the latter quoted rather than strictly matched, as the paper acknowledges), that the gain persists on three real-world manipulation tasks, and that inference latency drops from 5042 ms to 180 ms. Supporting analyses show why: representation-space error magnitude is a poor proxy for action consequences, since selectively correcting a component holding only 0.63–0.85% of teacher-referenced discrepancy energy recovers 19.8–34.0% of the achievable offline action-error reduction.

The premise it rests on. Everything rests on the frozen single-step action model $G_\psi$ being a faithful, smooth conduit: if its gradients with respect to student-generated futures are miscalibrated, saturated, or directionally insensitive to the future variations that matter under closed-loop execution, then minimizing action error through it optimizes a hollow target and action-sufficient distillation would not transfer to task success.

If the paper is right (Pith's inferences, not the authors')

Explicit future prediction and iterative diffusion inference are separable: DirectWAM keeps the future-to-action interface while cutting model-side latency from 5042 ms to 180 ms, a 28.0x reduction, without leaving the efficiency–performance frontier.

Action sufficiency becomes a design criterion for cascade WAM conversion: future predictors should be judged by the actions their own predictions induce, with representation matching demoted to an optional anchor (as on LIBERO-Long, where $\lambda_v = 0$ suffices).

The controlled comparison implies that video-space supervision alone is insufficient: matching teacher or ground-truth futures reaches only 76.0–78.2% on RoboTwin versus 87.0% for action-error supervision under identical initialization, data, budget, and frozen action expert.

Representation-space error magnitude is a poor proxy for control relevance: selectively transferring 1.4% of the prediction-difference energy recovers essentially the full closed-loop gain (87.3% versus 87.0% success), arguing for control-weighted future errors.

Video-side stochastic sampling is not load-bearing for control in the evaluated regime: fixing video noise or averaging four sampled futures does not systematically change task success, so a deterministic future predictor costs nothing on these benchmarks.

My reading: the recipe should transfer wherever the consuming action model is differentiable, so the same two-stage conversion could accelerate other released cascade WAMs or planner-based world models; applying DirectWAM to such architectures is a testable extension that would confirm or bound the generality the paper leaves open.

My reading: because success depends on $G_\psi$'s gradients being informative, the method's margin should track the local conditioning of the frozen action expert; reporting the Jacobian conditioning of $G_\psi$ on student-generated futures would act as a canary for when action-sufficient distillation begins optimizing a hollow target.

My reading: the funneling diagnostic suggests a cheaper surrogate — weighting or projecting future errors by local action sensitivity — could approximate full action-sufficient distillation during training; the paper uses the projection only as a diagnostic, so this is an extrapolated extension, not a claim.

My reading: the results imply that world-action-model benchmarks should report control-relevant future error, not distribution-level visual metrics alone, since FID-class metrics rank DirectWAM's futures below the full teacher's even though its closed-loop success is comparable.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Cascade world-action models (WAMs) use predicted visual futures to condition robot actions, but inherit costly iterative inference from diffusion-pretrained experts. We ask what a single-step predicted future must preserve to support accurate single-step action prediction. In the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both streams to one step substantially reduces task success. Video-space supervision only partially closes this gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. These findings motivate DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling through action-sufficient distillation, while retaining the explicit future-to-action cascade. After learning a one-step action model, we train the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert, with an optional future-latent reconstruction loss. The resulting objective evaluates the student-generated future through its downstream control consequences rather than relying on representation matching alone. DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget. This advantage has also extended to three real-world tasks that we designed. Compared with the full iterative teacher, it reduces inference latency by 28.0 ×.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-21

The authors' main claim, as the classifier picked it out 0.81

“DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.22 · human participants 0.05 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Cascade world-action models (WAMs) use predicted visual futures to condition robot actions, but inherit costly iterative inference from diffusion-pretrained experts. background 0.99 0.06 0.03 asserted 1.00
We ask what a single-step predicted future must preserve to support accurate single-step action prediction. gap 0.81 0.27 0.04 unclear 0.76
In the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both streams to one step substantially reduces task success. result 0.98 0.91 0.06 asserted 1.00
Video-space supervision only partially closes this gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. result 0.78 0.90 0.05 asserted 0.99
These findings motivate DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling through action-sufficient distillation, while retaining the explicit future-to-action cascade. contribution 0.85 0.79 0.38 asserted 0.94
After learning a one-step action model, we train the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert, with an optional future-latent reconstruction loss. method 1.00 0.93 0.06 asserted 0.99
The resulting objective evaluates the student-generated future through its downstream control consequences rather than relying on representation matching alone. method 0.94 0.84 0.09 asserted 1.00
DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget. result 1.00 0.96 0.46 asserted 1.00
This advantage has also extended to three real-world tasks that we designed. result 0.98 0.92 0.08 asserted 0.97
Compared with the full iterative teacher, it reduces inference latency by $28.0\times$. result 1.00 0.93 0.14 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.94

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-21: 1 request, 7,575 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. Overview of DirectWAM. (a) Video randomness produces similar downstream actions. (b) Naive step truncation and visual supervision induce a noticeable performance gap. (c) Direct￾WAM is a deterministic one-step cascade WAM trained through action-sufficient distillation. can therefore be removed, yielding a deterministic future predictor that nonetheless still relies on the teacher’s multi-step computation. Removing this iteration is non-trivial: naively reducing the steps degrades control, and learning a one-step predictor with video-space supervision recovers only part of the gap.…
Figure 2 of the paper
Figure 2. Closed-loop effects of video-side stochasticity. (a) RoboTwin success rates of the teacher, fixed video noise, and averaging four future predictions. (b) Eight rollouts per scene under different video-noise seeds. Successful trajectories remain closely aligned. action mapping. For two samples from the same context, let DV denote their distance in future￾representation space and DA the corresponding distance between their induced actions. We evalu￾ate variation propagation through three complementary statistics: (i) the Pearson correlation between DV and DA across same-context sample…
Figure 3 of the paper
Figure 3. Inference-step reduction and action-sensitive error recovery. (a) Success rates (%) across video/action step budgets. Performance degrades as the number of steps decreases. (b) The action-sensitive component holds only 0.63–0.85% of the discrepancy energy yet recovers 19.8– 34.0% of the offline action-error improvement. motion range R, giving Dtraj = meant[σ(t)]/R. The median normalized spread is 0.020, ac￾counting for a median absolute spread of 0.86 cm compared to a median motion range is 43 cm. Normalization details and outcome-specific summaries appear in Appendix A.3. Video…
Figure 4 of the paper
Figure 4. Efficiency and qualitative comparison on RoboTwin. (a) Success–latency trade-off; labels show speedups over full LingBot-VA (5042 ms). (b) Six examples of predicted futures from the full teacher, the one-step teacher, Flash-WAM, and DirectWAM. tational cost of multi-step diffusion. DirectWAM preserves this interface while replacing iterative inference with direct prediction, without moving off the observed efficiency–performance frontier. Action-supervised futures remain visually decodable. Using downstream action error as the pri￾mary supervision raises a natural concern: could the…
Figure 5 of the paper
Figure 5. Real-world task trajectories. Six keyframes per task are ordered from left to right. Each keyframe pairs the external view (left) with the wrist view (right). 5 DISCUSSION Separating pretrained capability from inference procedure. DirectWAM highlights a distinc￾tion between inheriting a diffusion-pretrained model and retaining its generative inference proce￾dure. A pretrained future-to-action cascade can be converted into deterministic direct prediction without removing the intermediate future. This does not imply that iterative computation is redun￾dant: naive step reduction causes…
Figure 6 of the paper
Figure 6. reports the phase in which the action expert is held fixed, and the video expert is updated. We first consider LIBERO-Long, where λv = 0. Despite receiving no reconstruction supervision, the future-prediction pathway learns to reduce the action loss over training. The reduction therefore does not require updating the action expert or explicitly minimizing a video reconstruction objective. RoboTwin provides a complementary setting with λv = 0.01. Here, action loss decreases substan￾tially while the weighted anchor loss remains approximately stable. Improved action regression is…

Figures are the authors', reproduced from the paper.

References (24)
  1. [1] Causal World Modeling for Robot Control , author=. arXiv preprint arXiv:2601.21998 , year= on PithPith's reviewarXiv
  2. [2] Du, Yilun and Yang, Sherry and Dai, Bo and Dai, Hanjun and Nachum, Ofir and Tenenbaum, Josh and Schuurmans, Dale and Abbeel, Pieter , booktitle =. Learning Universal Policies via Text-Guided Video Generation , url =. doi:10.52202/075280-0403 , editor = on Pithdoi:10.52202/075280-0403
  3. [3] World Action Models are Zero-shot Policies , author=. 2026 , eprint= on Pith
  4. [4] Bi, Hongzhe and Tan, Hengkai and Xie, Shenghao and Wang, Zeyuan and Huang, Shuhe and Liu, Haitian and Zhao, Ruowen and Feng, Yao and Xiang, Chendong and Rong, Yinze and Zhao, Hongyan and Liu, Hanyu and Su, Zhizhong and Ma, Lei and Su, Hang and Zhu, Jun , title =. Proceedings o... on Pith
  5. [5] Fast-WAM: Do World Action Models Need Test-time Future Imagination? , author=. arXiv preprint arXiv:2603.16666 , year= on PithPith's reviewarXiv
  6. [6] Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models , author =. arXiv preprint arXiv:2608.04404 , year = on PithPith's reviewarXiv
  7. [7] Flash-WAM: Modality-Aware Distillation for World Action Models , author=. 2026 , eprint= on Pith
  8. [8] DIDO: Distilling Interaction-Centric Dynamics into One-Step Denoising for World Action Models , author=. 2026 , eprint= on Pith
  9. [9] Objective Mismatch in Model-based Reinforcement Learning , author =. Proceedings of the 2nd Conference on Learning for Dynamics and Control , pages =. 2020 , editor = on Pith
  10. [10] Farahmand, Amir-Massoud and Barreto, Andre and Nikovski, Daniel , booktitle =. 2017 , editor = on Pith
  11. [11] Grimm, Christopher and Barreto, Andre and Singh, Satinder and Silver, David , booktitle =. The Value Equivalence Principle for Model-Based Reinforcement Learning , url = on Pith
  12. [12] Learning 4D Geometric Priors for Inference-Efficient World Action Models , author=. 2026 , eprint= on Pith
  13. [13] VGGT: Visual Geometry Grounded Transformer , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year= on Pith
  14. [14] Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models , author =. 2026 , eprint = on Pith
  15. [15] GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch , author=. arXiv preprint arXiv:2607.13960 , year= on PithPith's reviewarXiv
  16. [16] LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning , author=. arXiv preprint arXiv:2306.03310 , year= on PithPith's reviewarXiv
  17. [17] Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation , author=. arXiv preprint arXiv:2506.18088 , year= on PithPith's reviewarXiv
  18. [18] Hartigan, J. A. and Hartigan, P. M. , year=. The Dip Test of Unimodality , volume=. The Annals of Statistics , publisher=. doi:10.1214/aos/1176346577 , number= on Pithdoi:10.1214/aos/1176346577
  19. [19] Decoupled Weight Decay Regularization , author=. 2019 , eprint= on Pith
  20. [20] Consistency Models , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , editor = on Pith
show 4 more
  1. [21] Yin, Tianwei and Gharbi, Micha\"el and Zhang, Richard and Shechtman, Eli and Durand, Fr\'edo and Freeman, William T. and Park, Taesung , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages = on Pith
  2. [22] Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author=. Proceedings of Robotics: Science and Systems (RSS) , year= on Pith
  3. [23] Aaditya Prasad AND Kevin Lin AND Jimmy Wu AND Linqi Zhou AND Jeannette Bohg , TITLE =. Proceedings of Robotics: Science and Systems , YEAR = on Pith
  4. [24] Much Ado About Noising: Dispelling the Myths of Generative Robotic Control , author =. International Conference on Learning Representations (ICLR) , year = on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment