Awesome-WAMs
GitHub
PAPER REPORTENAll readings ↗

World-Coherent Decoding: Self-Verifying Test-Time Planning for World Action Models

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Chuhan Zhang, Seiji Ito, Kenta Hoshino, Satoshi Ikehata, Ikuro Sato

Affiliations: Department of Computer Science, Institute of Science Tokyo; DENSO IT Lab, Japan; National Institute of Informatics, Japan

Source: 2609.02159 ↗ · Catalog record

Reading: 2 / 558 · 6 original figures & tables · ~21 min ·

1. Paper overview

In one sentence: WCD spends extra test-time compute choosing among frozen-model rollouts, then learns from delayed prediction errors to improve future choices. E03E04E05E07E08E12E14

At a glanceWhat to know
Research problem
Source description

A causal WAM can generate different visual futures from identical history and instruction, then decode different actions from those futures. The control problem is to choose a reliable candidate before execution reveals its consequences. The authors identify within-prefix variation in hindsight prediction error and an association between early mismatch and eventual failure; neither observation alone supplies a usable pre-execution selector. E02E03E12

Core mechanism
Source description

WCD combines video flow surprisal and action denoising effort to rank sampled future–action candidates without changing the backbone weights. E04E05E06

A key reported resultRoboTwin 2.0 Horizon-3 manipulation: 46.43%

Task success rate. Limited-randomization training; Hard evaluation; Table 2 H3 column

LingBot-VA: 30.00%; reported gain: 16.43 percentage points. The strongest horizon-specific gain supports usefulness on longer tasks, but does not isolate error accumulation as its causal mechanism. E14

Reading caution
Reader analysis

Higher model density and lower latent prediction error need not imply task progress. The early-error diagnostic supports relevance to failure, but supplies neither causal validation nor a success-probability calibration. E04E07E12

Core contributions

  • Source description

    WCD combines video flow surprisal and action denoising effort to rank sampled future–action candidates without changing the backbone weights. E04E05E06

  • Source description

    Executed observations provide delayed supervision for a small predictor of future imagination–reality error. After warm-up, this predictor replaces expensive divergence-based video scoring. E07E08E09

  • Reader analysis

    The experiments support the usefulness of structured selection on this backbone, especially under limited randomization. They do not establish that latent reliability is a universally sufficient control objective. E12E14E15

Figure 3. The execution boundary determines when a prediction can be verified. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow the diagram from left to right. In Phase 1, the frozen WAM samples N imagined-future/action pairs from one observation history. A video score and action path effort produce a fused ranking, and only the selected action chunk crosses the execution boundary. In Phase 2, observed latents can finally be compared with the retained imagination, producing backward error b_m. The blocked arrow is essential: this error did not exist when the current candidate was selected. Phase 3 stores executed-candidate features and error labels, trains f_psi, and uses its predictions to rank candidates at later decisions. E03E06E07E08E09E20

What it supports. WCD learns how to anticipate its model’s prediction errors while leaving the WAM weights frozen. Online adaptation resides in the lightweight reliability predictor and fusion rule. This is a temporal calibration loop: evidence from one executed chunk informs subsequent choices rather than retroactively verifying the current one.

Where the evidence stops. Replay labels come from selected candidates only; the actual outcomes of rejected candidates remain unknown. The figure does not establish unbiased coverage of every possible action or guarantee that a low-error future completes the task.

2. Motivation

2.1 The problem and the proposed response

Source description

A causal WAM can generate different visual futures from identical history and instruction, then decode different actions from those futures. The control problem is to choose a reliable candidate before execution reveals its consequences. The authors identify within-prefix variation in hindsight prediction error and an association between early mismatch and eventual failure; neither observation alone supplies a usable pre-execution selector. E02E03E12

2.2 What this reading follows

A world-action model can imagine several futures from the same observation, yet only one action chunk will be executed. WCD makes that choice explicit. It initially ranks candidates with video-flow surprisal and action-denoising effort, then uses the mismatch between its chosen imagination and subsequent observations to train a small reliability predictor. The central constraint is temporal: the world reveals the best supervision after the choice has already been made. The figures below trace that boundary before examining the strongest Hard RoboTwin result, the selector ablations and the evidence that prediction mismatch is behaviorally meaningful. The backbone remains frozen throughout this process. E03E04E05E07E08E12E14

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryWAMs
ArchitectureNot applicable
Prediction paradigmOther mechanisms
QuadrantOutside quadrants

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

WAM is supported by inference-time visual-future generation followed by future-conditioned action decoding. Other mechanisms / Outside quadrants and architecture Not applicable are defensible for WCD as a selection framework applied to an existing backbone. The underlying generation order has an inverse-dynamics interpretation, however, and should not be mistaken for simultaneous future/action prediction or evidence of a One Model architecture. The subcategory labels are less secure: WCD trains a reliability regressor without updating the policy or using rewards, so WM-RL is not established. Replacing divergence replay supports an inference-efficiency motivation, but multiple samples remain necessary and real-time control is not quantitatively demonstrated. The supplied snapshot is preserved rather than revised. E02E03E08E09E18

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • Language instruction and shared task conditioning
  • Causal cache C_m containing realized visual latents and previously executed actions
  • Stochastic video/action generation traces
  • After execution: realized observation chunk encoded by E
  • One selected low-level action chunk for execution
  • Selected imagined visual future retained for subsequent auditing
  • Updated reliability predictor, replay buffer, and video/action fusion weight

4.2 Equations and their role

z^m,1:K∼qθv(⋅∣Cm),am,1:K∼qθa(⋅∣z^m,1:K,Cm),zm,1:Kreal=E(om,1:K).\begin{aligned}\hat z_{m,1:K}&\sim q^v_\theta(\cdot\mid C_m),\\ a_{m,1:K}&\sim q^a_\theta(\cdot\mid\hat z_{m,1:K},C_m),\\ z^{\mathrm{real}}_{m,1:K}&=E(o_{m,1:K}).\end{aligned}
Equation (1) separates learned visual prediction, future-conditioned action decoding, and encoding of observations obtained after actual execution. K indexes frames within the chunk. E03
ddτlog⁡pτ(xτ∣Cm)=−∇ ⁣⋅uθv(τ,xτ∣Cm),sm,nv,surp=−log⁡pbase(xm,n,0v)+∑r=0Rv−1∇ ⁣⋅uθv^(τrv,xm,n,rv∣Cm) Δτrv.\begin{aligned}\frac{d}{d\tau}\log p_\tau(x_\tau\mid C_m)&=-\nabla\!\cdot u^v_\theta(\tau,x_\tau\mid C_m),\\ s^{v,\mathrm{surp}}_{m,n}&=-\log p_{\mathrm{base}}(x^v_{m,n,0})+\sum_{r=0}^{R_v-1}\widehat{\nabla\!\cdot u^v_\theta}(\tau^v_r,x^v_{m,n,r}\mid C_m)\,\Delta\tau^v_r.\end{aligned}
Equations (2–3) use the source-to-data coordinate τ = 1 − σ. Positive accumulated divergence raises surprisal. The base-density term is part of the score; divergence alone is not the stated likelihood estimate. R_v counts saved video solver steps, and the divergence estimate uses the Hutchinson estimator. E04
sm,na=∑q=0Ra−1mean⁡ ⁣[(Δam,n,q)2]∣Δσqa∣+ϵs^a_{m,n}=\sum_{q=0}^{R_a-1}\frac{\operatorname{mean}\!\left[(\Delta a_{m,n,q})^2\right]}{|\Delta\sigma^a_q|+\epsilon}
Equation (4) measures squared motion along the action denoising trajectory, normalized by scheduler spacing. It is not the physical energy or smoothness of the executed robot trajectory. E05
s~m,nj=sm,nj−μ(sm,1:Nj)std⁡(sm,1:Nj)+ϵs,j∈{v,a},nm⋆=arg min⁡n[λms~m,nv+(1−λm)s~m,na].\begin{aligned}\tilde s^j_{m,n}&=\frac{s^j_{m,n}-\mu(s^j_{m,1:N})}{\operatorname{std}(s^j_{m,1:N})+\epsilon_s},\qquad j\in\{v,a\},\\ n_m^\star&=\operatorname*{arg\,min}_n\left[\lambda_m\tilde s^v_{m,n}+(1-\lambda_m)\tilde s^a_{m,n}\right].\end{aligned}
Equations (5–6) make the video and action scores comparable within a pool. Lower fused score determines which action chunk is executed. E06
em,kvb=MSE⁡ ⁣(z^m,nmexec,k,zm,kreal),bm=1K∑k=1Kem,kvb,λm+1=λbasemin⁡ ⁣(1,bˉmhistmax⁡(bm,ϵb)).\begin{aligned}e^{\mathrm{vb}}_{m,k}&=\operatorname{MSE}\!\left(\hat z_{m,n_m^{\mathrm{exec}},k},z^{\mathrm{real}}_{m,k}\right),\\ b_m&=\frac1K\sum_{k=1}^K e^{\mathrm{vb}}_{m,k},\\ \lambda_{m+1}&=\lambda_{\mathrm{base}}\min\!\left(1,\frac{\bar b_m^{\mathrm{hist}}}{\max(b_m,\epsilon_b)}\right).\end{aligned}
Equations (7–8) turn observed prediction error into delayed feedback. Above-baseline mismatch reduces subsequent video weighting; the rule caps that weight at λ_base. The index n_m^exec identifies the candidate actually executed; the historical mean uses prior backward errors. E07
e^m,n,kvb=fψ(ϕm,n)k,sm,nv,pred=1K∑k=1Ke^m,n,kvb,L(ψ)=E(ϕ,y)∼B ⁣[∥fψ(ϕ)−y∥22].\begin{aligned}\hat e^{\mathrm{vb}}_{m,n,k}&=f_\psi(\phi_{m,n})_k,\\ s^{v,\mathrm{pred}}_{m,n}&=\frac1K\sum_{k=1}^K\hat e^{\mathrm{vb}}_{m,n,k},\\ \mathcal L(\psi)&=\mathbb E_{(\phi,y)\sim\mathcal B}\!\left[\left\|f_\psi(\phi)-y\right\|_2^2\right].\end{aligned}
Equations (9–10) regress the observed framewise mismatch vector y. The resulting video score estimates prediction error, not a calibrated probability of task success. E08

5. Method in detail

5.1 Turn one stochastic rollout into a controlled comparison

Source description

The backbone first predicts a visual latent future and then generates an action chunk conditioned on that future. WCD repeats this process N times from the same causal cache. Holding the history and instruction fixed is what makes candidate scores comparable: differences come from sampling, not from different observations. Before any outcome is known, video surprisal and action-denoising effort are standardized within the pool and combined. The chosen action is executed, while its imagined future is retained for later comparison. This decoding layer should be distinguished from the underlying future-conditioned action model. WCD adds a selection mechanism; it does not replace the backbone with a newly trained unified video-action network, and its online learning does not update the backbone weights. E02E03E04E05E06

Figure 2. Cold-start selection uses generation traces before any physical feedback is available. Original paper, p. 4 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Panel (a) keeps the visual and action caches fixed while stochastic flow trajectories produce different candidates. Panels (b) and (c) explain the video-side signal in the source-to-data coordinate tau = 1 - sigma: positive divergence expands local volume, lowers density and raises surprisal. The score includes the base negative log density as well as accumulated divergence, not divergence alone. Panel (d) switches to the action branch. Its path effort sums squared denoising updates normalized by solver-step spacing. These are internal generation trajectories. After separate standardization across the current candidate pool, both scores contribute to the final ranking. E04E05E06E20

What it supports. The two proxies ask complementary questions: how likely is this imagined future under the learned video distribution, and how much corrective motion did generating its action require? Neither needs a task reward. Their scale normalization makes them relative comparisons within a shared context, rather than universal reliability thresholds.

Where the evidence stops. Action denoising effort is not physical energy, torque or smoothness of the executed robot motion. Video density likewise reflects the learned distribution rather than guaranteed physical feasibility; an unfamiliar but valid future could receive high surprisal.

5.2 Use delayed labels without crossing the causal boundary

Reader analysis

After the selected action executes, the encoder turns real observations into the same latent space as the prediction. Their framewise mean-squared difference supplies a backward-error vector. One use is immediate but still delayed: if the new mean error exceeds the historical baseline, the next decision reduces its trust in the video-side score. The other use is supervised regression. The system pairs saved pre-execution features with the revealed error vector in a replay buffer and trains a small predictor. My interpretation is that this trades an expensive generative-density proxy for an estimate more directly aligned with experienced mismatch. It still learns only from what was actually tried. Rejected candidates do not acquire counterfactual labels simply because they were sampled. E07E08E09

5.3 Read the warm-up switch as a change in scoring, not generation

Reader analysis

Before the replay buffer reaches its threshold, the video score uses flow surprisal. Computing that score requires saved solver traces and gradient-based divergence replay. After warm-up, the predictor consumes compact generation features and replaces surprisal; the action-effort branch remains, and the WAM still generates the candidate pool. Predictor weights and replay persist across episodes of the same task, whereas episode-local cache and fusion state reset. This persistence is part of the evaluation protocol, not an interchangeable implementation detail. My reading is that a replication should report both cold-start and post-warm performance and latency. Otherwise gains from accumulated task experience can be mixed with one-episode decoding quality, and reduced scoring cost can be mistaken for removal of candidate-generation cost. E08E09E18E19

5.4 Training and inference

During training

Source description

The standard backbone uses 2.5k clean and 25k heavily randomized demonstrations. The limited-randomization backbone receives 18k clean-scene training steps followed by 18k steps with a 70% clean / 30% randomized mixture. These are distinct backbone-training protocols. E10

Source description

During WCD deployment, backbone weights remain frozen. Only the reliability predictor is trained online, using squared regression error on outcomes of executed candidates. This is supervised calibration from interaction, without task rewards or policy-gradient updates. E03E07E08

During inference

Source description

Selection is genuinely pre-execution: current backward error is unavailable when choosing the current candidate. It influences subsequent decisions through the fusion update and learned predictor. E06E07E08

Source description

The inference budget includes multiple backbone samples. Cold-start scoring additionally stores traces and replays the velocity field with gradients. Switching to the predictor removes that scoring operation, but retains candidate-generation cost. E08E09E18

5.5 Implementation flow

  1. Generate coupled candidates

    From the same causal cache, sample N visual futures and decode an action chunk conditioned on each future. The modality order is visual prediction followed by action generation; the paper explicitly distinguishes this from simultaneous video/action denoising. E03

  2. Score video density and action-generation effort

    Initially, estimate each visual future's negative log density using base-distribution surprisal plus accumulated flow divergence, estimated with Hutchinson's method. Separately, sum squared action denoising updates normalized by solver-step size. These measure model density and generation-path effort, respectively; neither directly measures task success. E04E05

  3. Fuse scores and execute

    Standardize both scores within the current candidate pool, combine them using weight λ_m, and execute the minimum-score candidate's action chunk. Save its imagined future and generation features. Only the chosen action is physically executed. E06

  4. Audit the selected prediction

    Encode the resulting observations and compute framewise latent MSE against the selected imagination. Average this error into b_m. If it exceeds the historical baseline, reduce the video-score weight for the next decision. E07

  5. Learn a planning-time mismatch estimate

    Store the executed candidate's features and error vector in a replay buffer. Regress framewise error from video velocity norms, pooled future latents, and pooled frame differences using separate feature encoders and a shared prediction head. E08

  6. Switch after sufficient feedback

    Once the buffer reaches a threshold, replace video surprisal with predicted mean mismatch and disable divergence replay. Action effort remains active. Predictor weights and the buffer persist across episodes of the same task; the fusion weight and autoregressive cache reset. E09

6. Experiments & results

World-Coherent Decoding selects among a frozen WAM's imagined futures and associated action chunks, then uses observed prediction errors to improve later selections. It raises simulated Hard success from 55.80% to 60.90% under limited randomized-scene supervision. Its central contribution is a feedback-calibrated decoding procedure; physical-robot evidence is qualitative, and missing implementation appendices constrain reproducibility.

6.1 Read the original evidence

Table 2. The strongest reported gain appears in the longest-horizon group. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each column vertically: WCD is a decoding intervention on the same frozen base WAM. H1, H2 and H3 separate horizon groups, while Avg reports the benchmark aggregate. The delta row contains percentage-point changes. The experimental context matters: the base policy trains for 18,000 clean-scene steps and then 18,000 steps with a 70% clean and 30% randomized mixture. This differs from the standard pretrained regime that uses much more randomized data. The table therefore tests candidate selection where the base generator retains room to improve. Do not average the displayed horizon columns to reconstruct the published aggregate. E10E13E14

What it supports. Overall Hard success rises from 55.80% to 60.90%, a 5.10-point improvement. H3 rises from 30.00% to 46.43%, a 16.43-point improvement. These are simulated executed-task outcomes, demonstrating that the selection rule can improve control beyond merely generating more visually plausible samples.

Where the evidence stops. The paper leaves the trial count unspecified and reports no uncertainty interval here. This limited-randomization gain should not be substituted for the much smaller improvement in the standard pretrained evaluation, whose starting performance is already high.

Figure 5. Observed mismatch is associated with failure, motivating the online learning target. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. The left panel groups episodes by eventual outcome, with 4,565 successful and 529 failed episodes labeled beneath the distributions. The error metric aggregates the first three executed chunks after task- and depth-dependent normalization. The right panel sorts episodes into four early-error quartiles and reports the fraction that later fail; counts under each bar make the denominators visible. The sequence runs from 62 failures among 1,274 episodes in Q1 to 270 among 1,274 in Q4. Because the error is measured after execution, this is evidence about a useful delayed label, not a score available before those same chunks were executed. E12

What it supports. Failure frequency rises from 4.9% in the lowest-error quartile to 21.2% in the highest. That association makes backward error a meaningful supervision target for a predictor of later candidate reliability. The overlap between outcome distributions also shows why the score cannot certify success or failure for an individual rollout.

Where the evidence stops. The plot is correlational and conditions on early executed behavior. It does not establish that lowering latent error causes success, nor that the observed relationship transfers unchanged to a different task, camera or physical robot.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
RoboTwin 2.0 manipulation

Standard training protocol; frozen pretrained LingBot-VA backbone; overall Hard evaluation

92.28%

Average task success rate

Base WAM: 91.55%; reader-computed improvement: 0.73 percentage points.

A modest improvement in the stronger training regime. Precise values come from Section 4.2 prose; the visually checked Table 1 displays rounded values of 91.6% and 92.3%. No horizon-specific Table 1 values are reported. E10E13

RoboTwin 2.0 manipulation

Limited-randomization training; Hard evaluation

60.90%

Average task success rate

LingBot-VA: 55.80%; reported gain: 5.10 percentage points.

This is simulated task execution success, not a video-quality metric. It measures WCD's selection benefit with the backbone fixed within this protocol. E10E14

RoboTwin 2.0 Horizon-3 manipulation

Limited-randomization training; Hard evaluation; Table 2 H3 column

46.43%

Task success rate

LingBot-VA: 30.00%; reported gain: 16.43 percentage points.

The strongest horizon-specific gain supports usefulness on longer tasks, but does not isolate error accumulation as its causal mechanism. E14

Candidate-selection module ablation

RoboTwin 2.0 hard generalization protocol; Table 3 Hard Avg. column

WCD: 60.9%; Predictor-only: 59.3%; Surprisal-only: 55.4%; Action-only: 55.1%; Random-N: 53.5%.

Average task success rate

Reader-computed WCD margins: 7.4 percentage points over Random-N and 1.6 over Predictor-only.

Structured selection beats random selection in this experiment. The prose's claimed 5.50-point margin over the strongest single branch contradicts the table; 5.5 instead matches the margin over Surprisal-only. E15E16

Early prediction-error diagnostic

Online rollouts; task- and depth-normalized log error over the first three executed chunks. Figure 5 labels 4,565 successful and 529 failed episodes (5,094 total).

4.9% (62/1,274) in the lowest quartile; 21.2% (270/1,274) in the highest.

Episode failure rate by early-error quartile

The authors report a monotonic increase across quartiles.

This is an association between latent mismatch and failure. It does not demonstrate causal identification or establish a calibrated failure predictor. E12

Physical Franka tabletop hammering

Clean-only demonstration training; unseen colored illumination, camera blur, and clutter

The authors report an executable WCD trajectory under colored light; no numerical success rate is supplied.

Qualitative execution outcome

The text reports failure of the frozen base WAM and π0.5 in the colored-light example.

This is author-reported qualitative physical-robot evidence. Figure 4 static setup images were inspected; supplementary execution videos were not. The source does not establish aggregate robustness across the three shifts. E11

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Table 3. Sampling a pool and selecting well are separate contributions. Original paper, p. 9 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start with the Random-N row: it generates a candidate pool but chooses uniformly, so it controls for sampling without a useful ranking signal. The next three rows isolate action effort, video surprisal and the learned predictor. Read the Online Pred. column together with Video Branch, because Predictor-only already learns from deployment feedback. WCD changes the video score from surprisal to predictor after warm-up while retaining action effort. The Hard Avg column favors the combination, but the horizon breakdown is less uniform: Predictor-only leads H1, Surprisal-only leads H2, and WCD leads H3. Clean performance also does not peak at WCD. E15E16

What it supports. WCD reaches 60.9% Hard average against 53.5% for Random-N and 59.3% for Predictor-only. Those differences are 7.4 and 1.6 percentage points. The combined method’s largest horizon-specific advantage is in H3, where 46.4% exceeds Predictor-only’s 35.4% by 11.0 points. These values follow the displayed table.

Where the evidence stops. The surrounding prose claims a 5.50-point improvement over the strongest single branch, inconsistent with the displayed 59.3% and 60.9% cells. This edition follows the table and identifies the actual gap as 1.6 points.

Figure 6(a)-(b). A larger pool helps only when the ranking signal can identify better candidates. Original paper, p. 9 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. The left panel plots average success against candidate count N for three selectors, not the complete WCD pipeline. Random-N stays near its baseline as the pool grows. Action-only improves at a smaller pool and then varies, whereas Predictor-only improves more clearly at larger counts. The right panel compares the change from N = 2 to N = 8 for individual tasks. Read the colored gain bars against the left axis and the green task-success diamonds against the right axis. The shaded difficulty ordering is a contextual guide, not a controlled intervention on task difficulty or an independent uncertainty estimate. E17E18E21

What it supports. At N = 8, the labels show 61.2% for Predictor-only, 60.0% for Action-only and 54.1% for Random-N. The task panel shows positive and negative changes rather than universal monotonic improvement. These diagnostics support selective use of extra generation compute, conditional on the ranking signal and remaining task headroom.

Where the evidence stops. The panel does not provide a full-WCD latency curve or matched wall-clock budget. Larger N generates more candidates even after the cheaper predictor replaces divergence replay, so these success curves do not alone establish efficient real-time scaling.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Higher model density and lower latent prediction error need not imply task progress. The early-error diagnostic supports relevance to failure, but supplies neither causal validation nor a success-probability calibration. E04E07E12

Reader analysis

The predictor receives realized labels only for executed candidates. Its accuracy on rejected candidates is therefore unestablished, and deployment-time selection can change the feature distribution on which it must rank candidates. E07E08

Reader analysis

Table 3 bundles warm-up, learned video scoring, action fusion, and online feedback in the full method. It does not cleanly isolate every component, and the prose overstates the gain over the strongest single-branch variant. E09E15E16

Reader analysis

Trial counts for the main benchmark comparisons, confidence intervals, and seed variation are not specified in the supplied evaluation account. Figure 5 does report episode counts for its separate error diagnostic. Persistent cross-episode calibration also makes task episode order and buffer initialization relevant to fair comparison. E09E10E13E14E15E12

Author claim

Additional parallel candidate generation remains a compute trade-off. The authors describe post-warm scoring overhead as negligible, but the supplied body contains no quantitative latency or high-frequency control benchmark. E18

Reader analysis

Repeated unresolved appendix references leave scheduler equivalence, estimator settings, exploration, predictor architecture details, and real-robot protocol details unavailable. Franka conclusions consequently remain preliminary. E11E19

7.2 Questions for discussion

  1. Does predicted mismatch rank useful task progress, or can it favor predictable but unproductive actions? [E07, E08, E12]
  2. How much does cross-episode predictor persistence contribute relative to within-pool selection? [E09, E15]
  3. What happens when visual shifts raise latent error without changing physical controllability? [E07, E11]
  4. At matched wall-clock budgets, does WCD outperform fewer candidates with more frequent feedback? [E17, E18]

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

A reproduction needs the corresponding LingBot-VA checkpoints and training regime, RoboTwin 2.0 task configurations, access to video/action solver traces, and synchronized imagined/real latent chunks. Standard and limited-randomization results should remain separate. E03E04E05E10

Reader analysis

Missing settings include the main evaluation candidate count and chunk horizon, solver grids, Hutchinson probe configuration, score stabilizers, fusion initialization, warm-up threshold, exact feature dimensions, predictor optimizer, replay sampling, and evaluation episode ordering. These should be recovered before claiming numerical replication. E04E06E07E08E09E19

Reader analysis

A small mechanism test would compare Random-N, surprisal-only, predictor-only, and full WCD at matched candidate counts on an unsaturated task, recording task success, selected-candidate mismatch, calibration history, and end-to-end latency. Equal feedback budgets and explicit buffer resets would test selection gains without confounding adaptation exposure. E09E15E17E18

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Audit the cold-start to predictor transition

Reader-proposed: use identical frozen checkpoints and initial scenes, then compare persistent replay, per-episode reset replay and a frozen predictor after warm-up. Log every buffer threshold crossing, score source, sample count and latency. Report first-episode and later-episode success separately. If gains require persistent same-task experience, this would narrow the claim from immediate decoding improvement to online adaptation over repeated deployments. E08E09E10E19

Check 2: Compare selectors under a measured compute budget

Reader-proposed: sweep N for Random-N, Action-only, Predictor-only and full WCD using shared candidate seeds and identical observations. Record generation, scoring and total latency, alongside success and selected-candidate error. Compare both fixed N and matched wall-clock budgets. A success advantage that disappears under matched latency would expose a compute tradeoff hidden by candidate-count-only curves. E15E17E18E21

8.3 Reading coverage

Visual audit: Original PDF pages 4, 6, 8 and 9 were visually inspected. The six final crops reproduce Figures 2, 3, 5 and 6 plus Tables 2 and 3; all were re-opened after 200-DPI cropping, with axis labels, legends and headers retained. Method text and the caption qualifications were read separately. Physical rollout videos and code were not inspected. The source contains unresolved appendix references and no supplied appendix text.

PDF pages inspected for this edition: 4, 6, 8, 9. Appendix coverage: not present.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Title, authors, and arXiv version identifier, p. 1
  • Abstract
  • 1 Introduction
  • 2 Related Work
  • 3.1 Preliminary: Chunked World Model and the Falsifiability Opportunity
  • 3.2 Test-Time Candidate Selection via Internal Generative Signals
  • 3.3 Delayed Self-Verification via Backward Error
  • 3.4 Online Calibration: Amortizing Delayed Error into a Learned Proxy
  • 4.1 Experiment Setup
  • 4.2 Main Results
  • 4.3 Ablation Study
  • 5 Conclusion
  • References, pp. 11–13
  • Supplied figure captions as text only
  • Independent PDF visual audit: pp. 5–10, including Equations (2–10), Tables 1–3 and static Figures 3–6

Outside the original text pass

  • The complete main text was reviewed; independent visual verification covered PDF pp. 5–10. Figures 1–2 were read as extracted text and captions, without visual inspection.
  • No appendix appears in the supplied 13-page artifact, despite repeated unresolved Appendix?? references. Separate supplements and rollout videos were not supplied or verified.
  • Merged horizon-specific cells in the Table 1 text extraction were resolved by inspecting the retained PDF, but no horizon-specific Table 1 values are reported. Standard-protocol results retain the precise Section 4.2 prose values.
  • No code was inspected and no experiments were reproduced.
  • Separate supplemental material availability has not been fully verified.
  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

E01PDF p. 1, title/author block and arXiv identifierInspect

Title and all five authors match the catalog. The supplied artifact identifies arXiv:2609.02159v1, dated 2 September 2026.

Go to primary source ↓
E02PDF pp. 2–3, Section 1, future-conditioned decoding discussion and pilot motivationInspect

The authors interpret future-conditioned action decoding through inverse dynamics and describe differing hindsight errors among samples sharing an observation prefix.

Go to primary source ↓
E03PDF p. 4, Sections 3.1–3.2, Eq. (1) and footnote 1Inspect

The frozen causal WAM generates visual latents before conditioned action chunks. Executed observations enter the causal history; candidates share history and task conditioning.

Go to primary source ↓
E04PDF p. 5, Section 3.2, Video-side flow surprisal, Eqs. (2–3)Inspect

Video surprisal combines base negative log density with integrated velocity-field divergence in τ = 1 − σ coordinates; divergence uses a Hutchinson estimator.

Go to primary source ↓
E05PDF p. 5, Section 3.2, Action-side path effort, Eq. (4)Inspect

Action scoring sums mean squared denoising updates divided by absolute scheduler increments plus a stabilizer and is described as a relative reliability proxy.

Go to primary source ↓
E06PDF p. 5, Section 3.2, Fusion and selection, Eqs. (5–6)Inspect

Scores are standardized across candidates and fused with λ_m. The minimum-score action is executed, while its future and generation features are retained.

Go to primary source ↓
E07PDF pp. 5–6, Section 3.3, Eqs. (7–8) and continuation paragraphInspect

Selected imagined and realized latents yield framewise MSE and mean backward error. This updates video weighting and supplies labels for executed-candidate replay entries without rewards. Equations (7–8) were checked against the rendered PDF; damaged text-extraction parentheses were resolved visually.

Go to primary source ↓
E08PDF p. 6, Section 3.4, feature description and Eqs. (9–10)Inspect

A predictor regresses framewise error from velocity norms and pooled latent/motion features. Surprisal requires saved traces and gradient-based replay; the predictor estimates mean future mismatch.

Go to primary source ↓
E09PDF pp. 6–7, Section 3.4, Warm-up curriculumInspect

A buffer-size threshold triggers replacement of surprisal by predictor scoring. Predictor and buffer persist across same-task episodes; fusion weight and cache reset.

Go to primary source ↓
E10PDF p. 7, Section 4.1, RoboTwin 2.0 simulationInspect

The benchmark has 50 tasks. Standard training uses 2.5k clean and 25k randomized demonstrations; limited training uses 18k clean steps then 18k mixture steps at 70%/30%. The text mentions multiple trials without a count.

Go to primary source ↓
E11PDF p. 7, Section 4.1, Real-robot Franka evaluation; p. 8, Section 4.2, Real-robot Franka stress testInspect

Clean-trained hammering is tested under colored light, blur, and clutter. The qualitative colored-light account reports baseline failures and an executable WCD trajectory, without success-rate statistics. Figure 4 was visually inspected as static setup imagery; it does not show aggregate execution outcomes.

Go to primary source ↓
E12PDF pp. 7–8, Section 4.2, Imagination error is behaviorally meaningful; rendered Figure 5, group counts and quartile barsInspect

Early error aggregates the first three decision chunks with task/depth normalization. Figure 5 labels 4,565 successes and 529 failures. Quartile failure counts are 62/1,274, 75/1,273, 122/1,273 and 270/1,274; displayed endpoint rates are 4.9% and 21.2%.

Go to primary source ↓
E13PDF p. 8, Section 4.2, Standard setting with a strong pretrained WAMInspect

The prose explicitly reports overall Hard success increasing from 91.55% to 92.28% with the pretrained backbone frozen. Rendered Table 1 rounds these overall Hard values to 91.6% and 92.3%.

Go to primary source ↓
E14PDF p. 8, Table 2, LingBot-VA/WCD/Δ rows, H3 and Avg. columnsInspect

Hard averages are 55.80% and 60.90%, with a 5.10-point gain. H3 values are 30.00% and 46.43%, with a 16.43-point gain. The cited cells were checked visually in the retained PDF.

Go to primary source ↓
E15PDF p. 9, Table 3, module labels and Hard Avg. columnInspect

Hard averages are Random-N 53.5%, Action-only 55.1%, Surprisal-only 55.4%, Predictor-only 59.3%, and WCD 60.9%. The cited cells were checked visually in the retained PDF.

Go to primary source ↓
E16PDF p. 9, Section 4.3, Module ablation paragraph; Table 3, Predictor-only/WCD Hard Avg. cellsInspect

The prose claims a 5.50-point gain over the strongest single branch, while the table lists Predictor-only at 59.3% and WCD at 60.9%. The cited cells were checked visually in the retained PDF.

Go to primary source ↓
E17PDF p. 9, Section 4.3, Effect of candidate count NInspect

The prose reports no consistent scaling trend for random selection and greater benefits from structured selectors on tasks with remaining headroom.

Go to primary source ↓
E18PDF p. 10, Section 5, final paragraphInspect

The authors acknowledge additional parallel generation cost, describe post-warm scoring overhead as negligible, and identify faster generation and high-frequency deployment as future work.

Go to primary source ↓
E19PDF pp. 4–7, Sections 3.1–3.4 and 4.1, unresolved appendix references; pp. 10–13, conclusion and referencesInspect

Implementation and protocol details are repeatedly deferred to unresolved Appendix?? references. The supplied artifact ends with references and contains no appendix text.

Go to primary source ↓
E20PDF p. 4, Figure 2(a)-(d); p. 6, Figure 3 (visually inspected)Inspect

Figure 2 distinguishes cache-conditioned flow samples, divergence/density accumulation and action-denoising effort. Figure 3 places selected-action execution before backward-error measurement, then uses replay and an online predictor for subsequent selection; legends identify the frozen WAM and warm-up switch.

Go to primary source ↓
E21PDF p. 9, Figure 6(a)-(b), axes, legends and endpoint labels (visually inspected)Inspect

Panel (a) compares Random-N, Action-only and Predictor-only, with N=8 endpoint labels of 54.1%, 60.0% and 61.2%. It does not plot the complete WCD selector. Panel (b) shows task-specific success changes from N=2 to N=8 with task-success proxies on a separate axis; gains vary by task.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.