Test-Time Scaling for World Action Models via Zero-Shot Geometric Evaluation
1. Paper overview
In one sentence: Gated GeoBoN spends extra WAM inference compute when action and imagined motion disagree, then selects geometrically consistent futures, trading some always-on success gain for lower average latency. e-purposee-gatee-geometrye-gatede-failure
| At a glance | What to know |
|---|---|
| Research problem | Source description A robot must decide whether extra inference compute is worthwhile before executing its current action chunk. WAMs expose imagined multiview observations alongside actions, making internal consistency available as a selection signal. The problem is to allocate sampling selectively and choose useful rollouts without training a reward model or testing each candidate in the environment. e-purposee-rollout |
| Core mechanism | Source description GeoBoN ranks sampled visual–action rollouts by cross-view depth reprojection inconsistency; Gated GeoBoN adds a cheaper action–future test to decide when ranking is worth invoking. e-gatee-geometrye-inference |
| A key reported result | Selective sampling — RoboCasa with Cosmos Policy: 67.9% success; 76.2% gain recovery; 24.7% triggering; 1.29 s. Success, gain recovery, trigger frequency and average latency. Same evaluation setting; Gated GeoBoN Nmax=8 versus always-on N=8. Baseline: 66.3%, 0.90 s; always-on: 68.4%, 3.65 s. Gating trades some success gain for lower reported latency. Across five settings, mean recovery is 74.8% and mean triggering 26.2%. e-setupe-gated |
| Reading caution | Reader analysis Geometric consistency does not certify instruction fulfillment. The gate compares motion directions, so agreement alone does not establish correct displacement magnitude or object interaction. These are reader deductions from the scoring definitions. e-gatee-geometry |
Core contributions
- Source description
GeoBoN ranks sampled visual–action rollouts by cross-view depth reprojection inconsistency; Gated GeoBoN adds a cheaper action–future test to decide when ranking is worth invoking. e-gatee-geometrye-inference
- Source description
The experiments separate closed-loop success, matched-budget offline selection, gate informativeness and the growth of spurious low-score selections with candidate count. e-fixed-budgete-selectore-gate-diagnostice-failure
Figure 1. The first stage allocates sampling; the second stage chooses the action chunk. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the primary and wrist observations and follow the WAM output into the blue gate. Its two arrows compare capsule-averaged visual motion \(\bar f\) with projected end-effector motion \(\Delta u\). The orange branch uses \(c<\tau_{\mathrm{gate}}\) to request more candidates; the green branch keeps the initial rollout when agreement passes. In the yellow evaluator, “Primary ← Wrist” means wrist-view geometry is reprojected into the primary camera, matching Eq. (5) on page 3. The lowest-error candidate supplies the executed action chunk. The red dashed path shows ordinary direct execution. The diagram omits proprioception, which Section 3 explicitly includes as an input. e-rolloute-gatee-geometrye-inferencee-native-depthe-purpose
What it supports. The figure makes two distinct roles visible: the gate decides whether to sample, and the geometry score ranks the resulting candidates. Sections 3.1–3.3 clarify that the initial rollout stays in the enlarged pool and that any moving arm can trigger the gate. No candidate is physically tested to obtain its selection score.
Where the evidence stops. The inequalities and reprojection direction agree with the caption and equations. The drawing simplifies the idle-arm rule and multi-arm aggregation. X-WAM experiments replace external depth prediction with native depth; this exception is stated on page 3.
2. Motivation
2.1 The problem and the proposed response
A robot must decide whether extra inference compute is worthwhile before executing its current action chunk. WAMs expose imagined multiview observations alongside actions, making internal consistency available as a selection signal. The problem is to allocate sampling selectively and choose useful rollouts without training a reward model or testing each candidate in the environment. e-purposee-rollout
2.2 What this reading follows
A world action model offers a useful inspection point before a robot acts: it predicts both an action chunk and images of the expected future. Gated GeoBoN turns that interface into two decisions. First, a motion check asks whether the initial action agrees with its imagined consequence. Only a failed check opens a larger candidate pool, where depth reprojection ranks the imagined futures. The experiments show modest success improvements and substantial reductions in reported latency relative to always-on selection. They also expose an important limit: a very low geometric score can be a misleading outlier, so sampling more futures need not improve control. e-purposee-gatee-geometrye-gatede-failure
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 dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Not applicable |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Conflicts with the recorded classification
The snapshot’s WAM relevance, Not applicable architecture, Other mechanisms and Outside quadrants fit an inference selector attached to existing WAMs: the wrapper is neither a new joint-prediction backbone nor inverse dynamics. Multiview geometry and efficient inference are supported. However, Policy post-training & WM-RL is unsupported as a literal mechanism label: the experiments explicitly use no fine-tuning and introduce no reinforcement-learning update. Preserve the recorded snapshot while flagging that subcategory for editorial review. e-rolloute-geometrye-inferencee-setup
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
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Use the action as a prediction of motion before executing it
The initial WAM rollout couples an action chunk with an imagined future, so the gate can compare two descriptions of motion without moving the robot first. Proprioception supplies the current end-effector position. The generated action implies an endpoint through forward kinematics or accumulated displacement, depending on the action representation. Project both into the primary camera, and compare the resulting direction with average optical flow inside the capsule around the projected trajectory. The gate ignores arms whose displacement is below 1 cm; any remaining arm with cosine agreement below −0.2 requests more candidates. If every arm is idle or every moving arm passes, execute the initial chunk. This is a selective-sampling rule rather than a learned success predictor, and the paper adds no training stage to obtain it. e-rolloute-gatee-setupe-inference
5.2 Ask whether two imagined views can share coherent depth
After a trigger, the candidate set includes the original rollout plus seven new samples in the main gated setting. Each candidate supplies primary and wrist future frames to the geometric evaluator. VGGT-Ω predicts geometry, allowing wrist-view points to be projected into the primary camera and compared with primary-view depth. Equation (5) averages the absolute logarithm of their depth ratio over valid pixels; the selected candidate minimizes this inconsistency. The validity mask requires in-image projections, positive depth and sufficient confidence, while an empty mask receives a large score. X-WAM experiments use native predicted depth instead of an external depth foundation model, with the adaptation incompletely specified. The output of selection is the candidate’s existing action chunk: the evaluator neither synthesizes a replacement action nor observes the actual future to choose it. e-geometrye-inferencee-native-depthe-gatede-reporting
5.3 Separate allocation quality, ranking quality and task completion
Reader interpretation: the strongest reading of the experiments requires keeping three questions separate. Table 4 asks whether the gate directs sampling toward states where offline reranking helps, compared with a random trigger at the same rate. Table 3 asks how well different selectors perform, and shows that higher offline error recovery need not imply higher closed-loop success. Tables 1 and 2 then measure completed tasks under fixed or gated inference, including the cost of invoking the larger pool. Finally, Figure 2 warns that a larger pool amplifies exposure to extreme evaluator errors; it does not say that 32% of tasks fail at N=16. Together, these results support selective use of an imperfect consistency signal. They leave open whether matched-latency random allocation, calibrated duplicate handling or different task distributions would change the observed tradeoff. e-gate-diagnostice-selectore-fixed-budgete-gatede-failure
5.4 Training and inference
During training
This is an inference wrapper around pretrained models. Experiments use public WAM checkpoints and default evaluation setups without fine-tuning; the external geometry model is frozen. No new objective, training stage or policy reinforcement learning is introduced. e-geometrye-setup
During inference
Baseline draws one rollout; fixed-budget GeoBoN always draws N. Gated GeoBoN draws one first and, only when triggered, adds Nmax−1 candidates. Selection precedes action execution at each control step; it does not evaluate candidates using actual future observations. e-purposee-inference
5.5 Implementation flow
- Test the initial action against its imagined motion
Compute Farneback flow from the current primary image to the predicted future. Obtain the action-implied end-effector endpoint using forward kinematics for joint-position actions or accumulated delta displacement. Project endpoints, average flow in a capsule around the projected trajectory, and compare motion directions. Ignore arms moving less than 1 cm; any remaining arm below cosine −0.2 triggers more samples. e-gatee-setup
- Score cross-view depth agreement
For each candidate, frozen VGGT-Ω estimates depth and camera geometry from the future-image pair. Reproject wrist-view points into the primary camera and compare their depths against directly estimated primary depth. Average only valid, positive-depth, sufficiently confident pixels; empty masks receive a large score. X-WAM instead supplies native predicted depth. e-geometrye-native-depth
- Select and execute
Keep the first rollout among the candidates and execute the chunk with minimum reprojection inconsistency. The gate passes the initial chunk directly if no moving arm disagrees, including when all arms are filtered as idle. e-gatee-inference
6. Experiments & results
Gated GeoBoN uses a WAM’s predicted futures twice: inexpensive action–future agreement determines whether to draw more candidates, then geometric consistency ranks those candidates. It requires no new training. Fixed-budget selection improves all five reported benchmark–backbone averages at N=8; gating retains much of that benefit with fewer sampling decisions, while evaluator outliers limit larger budgets.
6.1 Read the original evidence
Table 1. Fixed-budget GeoBoN improves all five aggregate settings at N=8, while individual task groups can regress. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across a row before comparing backbones: the baseline executes one rollout, whereas the four GeoBoN columns vary the candidate count. Bold marks the best fixed-budget entry within that row. The last column is the paired N=8 improvement in percentage points, with a reported 95% confidence interval. The shaded RoboCasa averages are weighted by task count, according to Section 4.2 on page 3; they are not an unweighted mean of the four printed categories. Section 4.1 specifies four seeds, 10 rollouts per task per seed on RoboCasa and RoboTwin, and 50 on LIBERO Long. Exact task lists and split identifiers are not supplied. e-fixed-budgete-setupe-native-depthe-reporting
What it supports. At N=8, RoboCasa improves from 66.3 to 68.4% for Cosmos Policy and from 80.8 to 82.5% for X-WAM. The X-WAM Door/Drawer row nevertheless falls from 96.7 to 92.1%, with a reported paired interval entirely below zero. Aggregate improvement therefore cannot be read as uniform improvement across tasks.
Where the evidence stops. The table gives confidence intervals for N=8 minus baseline, not for comparisons between budgets. Its CI construction is unspecified. Several settings decrease at N=16; X-WAM also uses a different depth source, so the table is not a controlled comparison of geometry estimators.
Table 2. Gating keeps much of the success improvement while invoking the larger pool at a minority of decision points. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each benchmark block contains three modes, so compare the shaded gated row with both neighboring rows. Success rate is task completion; full-gain recovery measures the fraction of the always-on improvement retained, not the percentage of tasks completed. BoN Trigger counts decision points that request extra samples, not whole episodes. The caption specifies a maximum pool of eight candidates: Section 3.3 says a trigger adds seven to the initial rollout. Finally, read latency in seconds alongside success, rather than treating triggering as a direct measurement of compute savings. The source names H200 and RTX Pro 6000 GPU nodes on page 3 without assigning hardware to individual rows. e-gatede-inferencee-setupe-reporting
What it supports. RoboCasa/Cosmos reaches 67.9% success at 1.29 seconds average latency, between baseline’s 66.3% at 0.90 seconds and always-on selection’s 68.4% at 3.65 seconds. Across the five settings, the paper reports mean full-gain recovery of 74.8% with mean triggering of 26.2%. The benefit is an explicit success–latency tradeoff.
Where the evidence stops. Average latency does not establish a worst-case control deadline. The paper leaves timing boundaries, batching, GPU counts and per-setting hardware assignments unspecified. It supplies no uncertainty estimates for these gated comparisons, so precise deployment speedups cannot be assumed.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboCasa task-group average — Cosmos Policy Default benchmark setup; four seeds, 10 rollouts/task/seed; task-count-weighted groups; N=8. | 68.4%; +2.1 percentage points [0.6, 7.3]. Closed-loop success; paired gain (95% CI) | Single rollout: 66.3%; N=16: 68.2%. Positive aggregate gain, with no monotonic guarantee as budget increases. e-setupe-fixed-budget |
| RoboCasa task-group average — X-WAM Same RoboCasa protocol; native-depth scoring; N=8. | 82.5%; +1.7 percentage points [0.3, 2.6]. Closed-loop success; paired gain (95% CI) | Single rollout: 80.8%. Door/Drawer falls 96.7→92.1%, while Coffee rises 73.3→83.7%; the average hides opposing effects. e-setupe-native-depthe-fixed-budget |
| LIBERO Long — Cosmos Policy Default evaluation; four seeds, 50 rollouts/task/seed; N=8. | 99.3%; +1.8 percentage points [0.4, 2.2]. Closed-loop success; paired gain (95% CI) | Single rollout: 97.5%. A near-ceiling benchmark result, not evidence of physical deployment. e-setupe-fixed-budget |
| LIBERO Long — LingBotVA Default evaluation; four seeds, 50 rollouts/task/seed; N=8. | 98.3%; +1.1 percentage points [0.2, 1.5]. Closed-loop success; paired gain (95% CI) | Single rollout: 97.2%; N=16: 99.1%. This setting continues improving at the larger tested budget. e-setupe-fixed-budget |
| RoboTwin 2.0 — Motus Default evaluation; four seeds, 10 rollouts/task/seed; N=8. | 89.9%; +2.1 percentage points [0.2, 3.4]. Closed-loop success; paired gain (95% CI) | Single rollout: 87.8%; N=16: 89.5%. The N=8 improvement extends to this benchmark but declines at N=16. e-setupe-fixed-budget |
| Selective sampling — RoboCasa with Cosmos Policy Same evaluation setting; Gated GeoBoN Nmax=8 versus always-on N=8. | 67.9% success; 76.2% gain recovery; 24.7% triggering; 1.29 s. Success, gain recovery, trigger frequency and average latency | Baseline: 66.3%, 0.90 s; always-on: 68.4%, 3.65 s. Gating trades some success gain for lower reported latency. Across five settings, mean recovery is 74.8% and mean triggering 26.2%. e-setupe-gated |
| Selector and gate diagnostics Fixed N=8 candidate dumps; four settings; matched trigger rates for the gate comparison. | X-WAM: GeoBoN ER 6.9%, online success 82.5%; gate help rate 67.2%. Offline error recovery and help rate; separate online success | Consensus: ER 8.6%, online 77.4%; random-trigger help: 43.2%. Offline rankings can disagree with closed-loop outcomes. Motus consensus also beats GeoBoN online, 90.5 versus 89.9%. e-selectore-gate-diagnostic |
| Cosmos Policy value-head comparison Identical N=8 candidate budgets; RoboCasa and LIBERO Long. | GeoBoN: 68.4% and 99.3%, respectively. Closed-loop success | Value head: 66.1% and 98.1%. Supports this selector comparison on two benchmarks, not superiority to every learned verifier. e-value-head |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Offline selection quality and closed-loop success can rank the same methods differently. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the left and right blocks as different measurements. Offline error recovery uses fixed pools of eight candidates; Section 4.4 on page 4 defines it as the selected candidate’s reduction in ADE relative to the first candidate, normalized by the reduction available to a ground-truth-aware oracle. The zero baseline is defined by this normalization. The oracle appears only in offline analysis. The right block measures executed closed-loop task success. Compare selector columns within each block, then look for ranking reversals across blocks. The source does not define ADE’s representation or units sufficiently to reconstruct it, and the consensus method is a paper-description reimplementation. e-selectore-reportinge-value-head
What it supports. On RoboCasa/X-WAM, consensus scores higher offline ER than GeoBoN, 8.6 versus 6.9%, yet has lower online success, 77.4 versus 82.5%. GeoBoN improves online success over baseline in all four rows, but consensus wins the Motus online comparison at 90.5 versus 89.9%. These are consistency advantages, not universal dominance.
Where the evidence stops. These diagnostics omit LIBERO Long/Cosmos, despite its inclusion in the main results. No uncertainty is shown. The separate value-head comparison on page 6, Table 5 supports another matched-budget comparison, but it does not remove the limits of the offline proxy.
Table 4. The gate targets decision points where reranking helps more often than a matched-rate random trigger. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First read Trigger: within each row, the random control invokes selection at the same rate as the action–future gate. Then compare Random and AF Gate under GeoBoN Help Rate. Section 4.4 on page 4 defines help as a positive difference between first-candidate ADE and GeoBoN-selected ADE; it does not define help as successful task completion. The diagnostic uses the same fixed eight-candidate dumps, isolating the choice of which decision points receive extra computation. The final delta column is a percentage-point difference between help rates. Read it separately from the gate’s trigger percentage and from Table 2’s full-gain recovery. e-gate-diagnostice-selectore-gatede-reporting
What it supports. At a 25.2% trigger rate on RoboCasa/X-WAM, the gate’s help rate is 67.2%, versus random’s 43.2%, a 24.0-point increase. The Motus difference is much smaller: 75.8 versus 71.7%, or 4.1 points. The signal carries information, but its strength varies substantially by setting.
Where the evidence stops. The matched-rate control is offline. It does not establish an online success or latency advantage over random triggering after the policies visit different states. Diagnostic sample counts, uncertainty and the exact ADE definition remain missing.
Figure 2. Larger pools increase exposure to spuriously low geometric scores. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the candidate count along the horizontal axis and the false-selection percentage vertically; each plotted point carries its reported percentage. The operational definition is on page 6, Section 4.5: near-duplicate futures have LPIPS distance below 0.02, and a selected score is flagged when its separation from a near-duplicate exceeds the evaluator’s 95th-percentile duplicate-pair variation. Figure 3 on that page provides the qualitative companion: similar future-image pairs receive widely different reprojection scores, and the first is selected at 0.016. Thus the plot audits score stability around similar images. It does not directly plot action error, task success or physical failure. e-failuree-fixed-budgete-reporting
What it supports. The reported rate increases from 7% at N=2 to 12% at N=8 and 32% at N=16. The authors interpret this as a multiple-comparisons effect: additional candidates offer both better rollouts and more opportunities for a misleading score minimum. Table 1’s nonmonotonic results are consistent with that explanation.
Where the evidence stops. The diagnostic is suggestive rather than a causal decomposition of task failures. The source omits sample counts, uncertainty and the calibration population for duplicate-pair variation. Figure 3 labels two candidates infinity/abstain without fully specifying that displayed abstention convention.
7. Analysis & limitations
7.1 What the evidence leaves open
Geometric consistency does not certify instruction fulfillment. The gate compares motion directions, so agreement alone does not establish correct displacement magnitude or object interaction. These are reader deductions from the scoring definitions. e-gatee-geometry
False low-score selection rises from 7% at N=2 to 32% at N=16. The authors attribute saturation partly to outlier selection among near-duplicate futures; this diagnostic is not a measured task-failure probability. e-failure
Tables 2–5 and Figure 2 omit uncertainty estimates; diagnostic sample counts and ADE representation/units are unspecified. Benchmark evidence does not establish physical-robot reliability or a transferable real-time latency guarantee. e-setupe-reporting
7.2 Questions for discussion
- Would the gate still outperform random allocation on matched-latency closed-loop trials? [e-gate-diagnostic, e-gated]
- Can duplicate-aware scoring preserve useful rollout diversity while reducing the large-budget outlier effect? [e-failure]
8. Reproducibility audit
8.1 Requirements and known gaps
Use the stated benchmark/checkpoint defaults and four-seed rollout counts. Reimplement projection, action conversion, Farneback flow and the depth mask; fix capsule radius at 30 pixels and confidence threshold at 0.5, alongside the stated idle/cosine thresholds. GPU families are H200 and RTX Pro 6000; counts and assignment are unspecified. e-gatee-geometrye-setupe-reporting
Exact reproduction still requires checkpoint revisions, task/split lists, software versions, sampling/frame horizons, X-WAM geometry adaptation, empty-mask penalty, ADE definition, CI construction and timing/batching rules. The proposed checks in the illustrated edition test matched-budget gate allocation and duplicate-sensitive ranking; neither has been run here. e-inferencee-native-depthe-reportinge-gate-diagnostice-failure
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: Compare the gate with random allocation offline and in closed loop
Reader-proposed check, not run: begin with fixed N=8 candidate dumps for RoboCasa/X-WAM and RoboTwin/Motus. Implement the published gate thresholds and compare its positive-help rate with repeated random triggers matched to its trigger frequency. Resolve and document ADE before computing this diagnostic. Then run paired-seed closed-loop trials using the same checkpoint, task set, GPU and candidate budget, matching average candidate counts and reporting measured latency as well as success. Include baseline and always-on selection as controls. The allocation claim predicts higher help rates offline and better success at comparable cost online; an offline advantage that disappears online would narrow the claim. e-gatee-setupe-gate-diagnostice-gatede-reporting
Check 2: Test whether duplicate-stable scoring reduces the large-budget penalty
Reader-proposed check, not run: use nested candidate pools of sizes 2, 4, 8 and 16 from common sampled rollouts, preserving image pairs, actions, masks and raw scores. Implement the paper’s LPIPS<0.02 duplicate rule and calibrate the duplicate-score variation on a disjoint, documented set. Compare raw minimum-score selection with a proposed control that gives each near-duplicate group its median score and breaks within-group ties by fixed candidate order. Keep generation budgets unchanged and record false-selection rate, offline error and paired-seed closed-loop success. If stabilizing duplicate scores reduces the N=16 diagnostic spike without improving control, score instability alone cannot explain the task-level saturation. e-native-depthe-geometrye-failuree-fixed-budgete-reporting
8.3 Reading coverage
Visual audit: All three source chunks were read individually. All eight PDF pages were rendered at 200 DPI and actually viewed. Page 1 verified title, seven authors, affiliations, version and date; page 2 supplied the architecture and rollout definition; page 3 supplied equations, inference, hardware families, evaluation counts, fixed thresholds and the X-WAM exception. Page 4 supplied Table 1 and diagnostic definitions; page 5 supplied Tables 2–4 and Figure 2; page 6 supplied Table 5, the duplicate criterion, Figure 3 and conclusion. Reference pages 7–8 were also visually inspected; no appendix is present. All six final crops were individually viewed. The four full-width crops use 200 DPI and the two narrow diagnostics use 400 DPI original renders. Figure 1 branch directions and wrist-to-primary geometry agree with the equations; its omitted proprioception and idle-arm details are disclosed in the guide. Separate supplements, code and experiments remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. 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
- PDF p. 1: title, authors, Abstract and Section 1, Introduction
- PDF p. 2: Section 2, Related Work; Section 3 and start of 3.1
- PDF p. 3: Sections 3.1–3.3, Method; Sections 4–4.2, setup and fixed-budget results
- PDF pp. 4–5: Sections 4.2–4.4, fixed-budget, gated and selector/gate evaluations
- PDF p. 6: value-head comparison, Section 4.5, failure analysis; Section 5, Conclusion
- PDF pp. 7–8: References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Primary identity verification on 2026-09-08: arXiv abstract metadata names Geometric Evaluation, while its directly linked same-version 2607.17454v1 PDF and HTML title read Geometric Verification. All seven authors and the version/date agree; freshly downloaded official PDF matches the supplied PDF hash exactly. This is a disclosed same-version metadata/manuscript inconsistency, not an inferred historical title change; no cross-version revision history is claimed.
- Identity notes: the preceding acquisition-verification statement is supplied provenance, not a network action in this reading. It resolves the catalog title Test-Time Scaling for World Action Models via Zero-Shot Geometric Evaluation to the inspected manuscript title Test-Time Scaling for World Action Models via Zero-Shot Geometric Verification. This review covers 2607.17454v1, whose PDF stamp is 20 Jul 2026; the supplied catalog submittedDate is 2026-07-19. Both dates are preserved without inferring a revision or date-conversion explanation. All seven authors agree after name-order/capitalization normalization.
- All three supplied text chunks were read individually and every PDF page was rendered and visually inspected. The extraction’s figure-image omission was addressed by inspecting Figures 1–3, Tables 1–5 and six final original crops.
- No appendix is present in this eight-page PDF. No separate supplement was supplied, and its availability remains unverified.
- Code, external links and checkpoint files were not inspected; experiments were not reproduced. No later edition or external source was acquired.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author/affiliation block and arXiv margin stamp
The manuscript title ends in Geometric Verification. The seven authors are Zesen Zhao, Minkyoung Cho, Hui Shen, Boyuan Zheng, Kunxiao Gao, Yulong Cao and Z. Morley Mao. Yulong Cao is affiliated with NVIDIA; the other six authors with University of Michigan. The stamp reads arXiv:2607.17454v1 [cs.RO], 20 Jul 2026.
Go to primary source ↓e-purposePDF p. 1, Abstract and Section 1, paragraphs describing the two consistency checks
Gated GeoBoN audits action–future agreement before invoking cross-view geometric selection. The checks use predicted images, actions and proprioception without task-success labels, ground-truth futures or trial environment rollouts for candidate scoring.
Go to primary source ↓e-rolloutPDF p. 2, Figure 1, caption and Section 3, Eq. (1)
Each stochastic WAM rollout contains predicted primary-view and wrist-view future frames plus an action chunk over horizon H. Inputs include multiview observations, proprioception and a language instruction. Figure 1 routes low cosine agreement to more sampling and high agreement to direct execution; reprojection runs from wrist into primary view.
Go to primary source ↓e-gatePDF p. 2, Section 3.1 opening; PDF p. 3, Section 3.1, Eqs. (2)–(4) and following paragraph
Current and action-implied end-effector positions are projected into the primary camera. Optical flow from current to predicted primary image is averaged within a capsule around the projected trajectory. Idle arms are ignored. Any moving arm below the cosine threshold triggers additional sampling; otherwise the initial chunk is executed, including when all arms are idle.
Go to primary source ↓e-geometryPDF p. 3, Section 3.2, Eq. (5) and validity-mask paragraph
Frozen VGGT-Ω estimates geometry from the predicted frame pair. Wrist-view points are reprojected into the primary view, and absolute log depth ratios are averaged over valid positive-depth, in-image, sufficiently confident pixels. Empty valid sets receive a large score. Lowest inconsistency wins.
Go to primary source ↓e-inferencePDF p. 3, Section 3.3, all three inference modes
Baseline executes the first action chunk. Fixed-budget GeoBoN always draws N rollouts and minimizes reprojection inconsistency. Gated GeoBoN draws Nmax−1 extra rollouts only when triggered, retains the initial rollout, and uses the same selector.
Go to primary source ↓e-setupPDF p. 3, Section 4.1, Experimental Setup
Public checkpoints and default evaluation setups are used without fine-tuning on H200 and RTX Pro 6000 GPU nodes. Each setting uses four seeds, with 50 rollouts per task per seed for LIBERO Long and 10 for RoboCasa and RoboTwin 2.0. Farneback flow, 1 cm idle threshold, −0.2 cosine threshold, 30-pixel capsule radius and 0.5 confidence threshold are fixed across experiments.
Go to primary source ↓e-native-depthPDF p. 3, Section 4.2, first paragraph
Fixed budgets are N in {2,4,8,16}. X-WAM uses its native depth prediction head for reprojection scoring instead of an external depth foundation model.
Go to primary source ↓e-fixed-budgetPDF p. 3, Section 4.2, group-weighting statement; PDF p. 4, Table 1, baseline, N=8, N=16 and paired-improvement columns
At N=8, task-count-weighted RoboCasa averages are 66.3→68.4 for Cosmos Policy and 80.8→82.5 for X-WAM, with paired gains +2.1 [0.6,7.3] and +1.7 [0.3,2.6] percentage points (95% CIs). LIBERO Long is 97.5→99.3 for Cosmos Policy and 97.2→98.3 for LingBotVA, with gains +1.8 [0.4,2.2] and +1.1 [0.2,1.5]. RoboTwin/Motus is 87.8→89.9, with gain +2.1 [0.2,3.4]. X-WAM Door/Drawer falls 96.7→92.1 (−4.6 [−6.9,−2.9]), whereas Coffee rises 73.3→83.7 (+10.4 [−2.8,24.5]). Several aggregate rows decrease at N=16.
Go to primary source ↓e-gatedPDF p. 1, Abstract; PDF p. 4, Section 4.3; PDF p. 5, Table 2, all five benchmark blocks
At Nmax=8, gated success rates are 67.9, 82.1, 98.8, 97.9 and 89.6%; corresponding trigger rates are 24.7, 25.2, 14.2, 34.5 and 32.2%. Full-gain recovery is 76.2, 76.5, 72.2, 63.6 and 85.7%, averaging 74.8%; mean triggering is 26.2%. RoboCasa/Cosmos average latency is 0.90/1.29/3.65 seconds for baseline/gated/always-on; RoboCasa/X-WAM is 2.68/3.11/9.67 seconds.
Go to primary source ↓e-selectorPDF p. 4, Section 4.4, Selector ablation and ER definition; PDF p. 5, Table 3, all four rows
Offline ER uses the gap from first candidate to a ground-truth-aware oracle on fixed N=8 dumps. GeoBoN has the highest ER in three of four settings and improves online success over baseline in all four. X-WAM consensus has ER 8.6 versus GeoBoN 6.9 but online success 77.4 versus 82.5%. Motus consensus has 90.5% online success versus GeoBoN 89.9%. Consensus was reimplemented from its paper description.
Go to primary source ↓e-gate-diagnosticPDF p. 4, Section 4.4, Gate ablation and help definition; PDF p. 5, Table 4 and continuation paragraph
Help means first-candidate ADE minus GeoBoN ADE is positive. At matched trigger rates on fixed N=8 dumps, random/AF-gate help rates are 43.2/67.2 for RoboCasa/X-WAM, 43.2/63.5 for RoboCasa/Cosmos, 47.8/69.4 for LIBERO/LingBotVA and 71.7/75.8 for RoboTwin/Motus, all percentages.
Go to primary source ↓e-value-headPDF p. 6, Table 5, both benchmark blocks; Section 4.4 continuation, value-head comparison
With Cosmos Policy and equal candidate budgets, GeoBoN beats the value head at every listed N. At N=8, RoboCasa success is 68.4 versus 66.1%; LIBERO Long is 99.3 versus 98.1%.
Go to primary source ↓e-failurePDF p. 5, Figure 2 and caption; PDF p. 6, Section 4.5 and Figure 3 with caption
Near-duplicates have LPIPS distance below 0.02. A false low-score selection has a score gap exceeding the evaluator’s 95th-percentile duplicate-pair variation. Reported rates at N=2,4,8,16 are 7,10,12,32%. Figure 3 shows similar future-image pairs with varied reprojection scores; candidate 1 is selected at 0.016, and candidates 15 and 16 are marked infinity/abstain.
Go to primary source ↓e-reportingPDF p. 3, Sections 3.1–4.2; PDF p. 4, Table 1 caption and Section 4.4; PDF p. 5, Tables 2–4; PDF p. 6, Table 5 and Section 4.5
The source supplies thresholds, GPU model families and seed/rollout counts, but these sections do not specify checkpoint revisions, task lists/split identifiers, software versions, GPU counts or per-setting hardware assignments, latency timing/batching protocol, exact ADE representation/units, confidence-interval construction, or diagnostic sample counts. Tables 2–5 and Figure 2 provide no uncertainty estimates. The geometry sections do not fully specify the X-WAM reprojection adaptation or the numerical empty-mask penalty.
Go to primary source ↓8.5 Primary sources
Test-Time Scaling for World Action Models via Zero-Shot Geometric Verification ↗
PDF · 4,750 extracted words
Source fingerprint
1fa68da603965ee37506d0bb38f187bb55062132c442b42eaa8fcd226ac7959e