acceptodds
Under review as a conference paper at ICLR 2027

From Physical Change to Action: Causal Attribution in World Action Models

Abstract

World Action Models (WAMs) couple future prediction with robot action generation, yet how prediction causally contributes to action and whether this contribution survives reduced predictive computation remain unclear. We use matched physical counterfactuals to define action responses and representation interventions to trace their transmission across four frozen WAM configurations. Model-specific interfaces strongly transmit these responses, although Fast-WAM-IDM shows that strong transmission need not reconstruct the complete counterfactual action. In Fast-WAM-Joint, fixing the action decoder’s future inputs sharply reduces the effect of replacing representations associated with the current observation, revealing substantial propagation through downstream future representations. Under blocking, restoring future inputs in a predefined late window of decoder layers recovers substantially more of the response than an equally sized early window. Targeted RoboTwin experiments likewise find larger propagation-allowed effects than future-fixed current effects, with the additional propagated effect opposing the latter in signed mean. Reducing Fast-WAM-Joint’s world-branch computation lowers cumulative policy-computation time by a median of 23.7% while all evaluated LIBERO continuations still complete. However, matched action responses change, with little held-out improvement from a shared scalar rescaling. Together, these findings show that prediction’s causal contribution to action can change with the inference procedure even when task completion is preserved.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
World Action Models, Causal Attribution, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9nK5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper audits how action responses to matched physical perturbations travel inside four frozen World Action Model (WAM) configurations, using activation patching to track their transmission. In Fast-WAM-Joint, with the decoder's future inputs fixed, the current observation accounts for under one-tenth of the full action change for the robot-displacement factor, so most of its effect travels through the predicted future. Reducing world-denoising steps from ten to five changes these responses while cutting cumulative policy-computation time by a median of 23.7%, with all 50 evaluated LIBERO continuations still completing.

Critical strengths

The response-level criterion complements task success: reducing world-denoising steps from ten to five keeps all 50 evaluated LIBERO continuations complete while changing the responses to the same physical changes, in ways a uniform rescaling does not explain.

The current/future source interventions separate the controlled direct effect of the current observation from the additional effect admitted by the recomputed future, isolating a direct pathway under one-tenth of the full effect for the robot-displacement factor.

Critical weaknesses

The source-attribution, layer-localization, and computation-budget findings all come from Fast-WAM-Joint, and three of the four audited configurations are variants of the same prior work. Breadth is narrow where the depth is claimed: the four-model interface audit rests on 544 cases per model drawn from 20 held-out states in two tasks, and the restoration cohort comes from two tasks. The interaction and budget analyses span five tasks, but all of these depth findings remain within Fast-WAM-Joint, so whether the central phenomena hold for WAMs generally is untested.

The headline ratio, the future-fixed current effect under one-tenth of the joint radial effect for the robot-displacement factor, is measured in a cell whose input combination need not occur during native inference; the ratio under the donor-future background is unreported, and the appendix magnitudes leave the one-tenth unguaranteed there. The practical import of the efficiency result is likewise open: the authors note that a response difference alone is not evidence of worse control.

Questions for authors

What is the headline ratio under the donor-future background? The four-cell outputs in Appendix C.1 should already contain it, and the reported magnitudes suggest an upper bound near a seventh of the joint effect. For F1, does the ratio change if the donor's proprioception is used in those cells?

Could you assemble the audited Fast-WAM-Joint architecture in one description, covering how the predicted-future inputs are produced, what the current-state node corresponds to, and how the command-state token and proprioception enter, so that a reader can judge whether the intervention sites are natural divisions of the model?

Could you document which checkpoint versions were audited and how they were obtained?

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer 9nK5 (Summary)

We thank the reviewer for the constructive feedback. We organize our responses into three parts addressing configuration coverage, source-background attribution, and reproducibility.

  1. Configuration coverage: The three Fast-WAM configurations provide a within-family comparison of prediction–action couplings, complemented by ImageWAM as a second model family. The four-configuration audit shows directional transfer coefficients of 0.951–0.999. We identify Joint as the configuration used for the detailed source-attribution, layer-restoration, and computation-budget analyses.

  2. Donor-future attribution: Under the donor-future background, dose-wise mean current-effect projection coefficients range from 0.04578 to 0.07330, all below one-seventh. We also report the completed donor-proprioception intervention as a single-case paired check.

  3. Architecture and checkpoint provenance: We explain future generation, current/future intervention boundaries, and instruction/state conditioning in one description. We report checkpoint download sources and file identities, together with the recorded parameter digests and code version.

AuthorsOpenReview ↗

Author Response to Reviewer 9nK5 (Part 1)

"The source-attribution, layer-localization, and computation-budget findings all come from Fast-WAM-Joint, and three of the four audited configurations are variants of the same prior work."

"Whether the central phenomena hold for WAMs generally is untested."

Cross-configuration evaluation. We acknowledge that Direct, Joint, and IDM belong to the same Fast-WAM family, rather than representing three independent model families. Our four-configuration evaluation is intended to compare distinct action-facing configurations, not to claim broad architectural diversity. It covers 7,318 valid discovery donors, 544 primary held-out donors per configuration across 20 shared states from two tasks, and 374 prospective cases per configuration. Transfer coefficients are consistently high across the four configurations (0.951–0.999), supporting action-response transmission through their respective interfaces.

Joint-specific causal analysis. We selected Fast-WAM-Joint for deeper attribution because its video and action experts expose identifiable current-origin and future-origin K/V inputs at the actual action-consumer interface. This enables controlled source replacement, downstream propagation analysis, and layer-specific restoration. The analysis includes 174 F1 records and 526 F3-G records across five tasks, 103 restoration donors from 20 source trajectories, and 200 computation-budget donors from 50 source trajectories.

We further observe the distinction between propagation-allowed node effects and future-fixed current effects on RoboTwin: their mean radial RMS magnitudes are 7.707 mm and 2.666 mm, respectively, with a matched signed radial difference of 3.304 mm (95% CI: [0.285, 6.933]).

Together, these results establish cross-configuration interface-level response transmission and provide deeper causal evidence for Fast-WAM-Joint on LIBERO and RoboTwin. We have clarified that the source-attribution and layer-localization findings are specific to the Joint configuration, rather than claiming that these mechanisms hold universally across WAM architectures.

AuthorsOpenReview ↗

Author Response to Reviewer 9nK5 (Part 2)

"What is the headline ratio under the donor-future background? The four-cell outputs in Appendix C.1 should already contain it, and the reported magnitudes suggest an upper bound near a seventh of the joint effect."

"For F1, does the ratio change if the donor's proprioception is used in those cells?"

Donor-future background. The small current-side effect persists under the donor-future background: the six dose-wise mean projection coefficients range from 0.04578 to 0.07330, all below one-seventh.

We recomputed these coefficients from all 174 valid F1 records, covering 35 registered states and five tasks, using the saved four-cell action arrays and original bootstrap weights. For each record, the conditional current-effect vector is projected onto the matched strict joint radial response and normalized by its squared norm.

F1 doseRecipient-future coefficientDonor-future coefficient [95% CI]
−4 cm0.070400.06614 [0.04277, 0.11068]
−2 cm0.062260.05688 [0.02317, 0.10973]
−1 cm0.073990.07330 [0.03616, 0.11987]
+1 cm0.073050.06930 [0.04176, 0.10372]
+2 cm0.052450.05082 [0.03361, 0.07955]
+4 cm0.045030.04578 [0.03688, 0.05412]

The donor-future range closely matches the recipient-future range of 0.04503–0.07399. All six paired background-difference intervals include zero, and all six pointwise 95% confidence intervals have upper bounds below one-seventh. Thus, the small current-side effect is observed under both future-input backgrounds.

Proprioception sensitivity. The original four-cell evaluation retained recipient action-side proprioception. We additionally conducted a matched intervention using donor proprioception. In the one fully completed paired case (task 0, state 00, −1 cm), the UC∣FRUC|FR coefficient changes from 0.03786 to 0.04076, while UC∣FDUC|FD changes from 0.03667 to 0.03852. Using the original recipient-proprioception joint radial response as the fixed reference vector, including its projection direction and squared-norm denominator, gives corresponding donor-proprioception coefficients of 0.04081 and 0.03859.

In this case, donor proprioception increases the two coefficients by 0.00290 and 0.00185, respectively, while both remain below 0.05. The pilot completed all four donor-proprioception conditions for one of five selected states; the remaining 16 outputs were unavailable following a GPU resource interruption.

Reviewer_E7ssOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how predicted futures influence robot actions in World Action Models, using controlled physical changes in simulation to define reference action responses. Across four frozen model configurations, it replaces internal representations with those from matched counterfactual scenes to measure how these representations transmit the resulting action changes. Its main analysis of Fast-WAM-Joint separates current-observation effects from effects propagated through future representations, finding substantially weaker current-side influence when future inputs are held fixed. Restoring future inputs in the final 5 decoder layers recovers more of the blocked response than restoring earlier layers, while RoboTwin experiments show that different input pathways can exert opposing effects. Halving world-branch denoising steps saves a median 23.7% of policy-computation time while preserving completion of all 50 evaluated LIBERO continuations

Critical strengths

Matching physical states, instructions, weights, and inference randomness gives the internal interventions an interpretable external reference. Particularly interesting is the comparison between replacing a current-state representation with downstream propagation and replacing it while fixing future inputs at every decoder read. This demonstrates strong upstream influence should not automatically be interpreted as strong direct dependence.

In the restoration experiment, radial response error falls from approximately 0.248 under blocking to 0.038 with late-layer restoration, whereas equally sized early-layer restoration leaves it essentially unchanged at 0.247; the subsequent six-window scan supports this localization. Prespecified windows, paired comparisons, and source-level uncertainty estimates strengthen this evidence.

Critical weaknesses

The experiments identify dependence on future-origin representations more convincingly than dependence on their predictive content. For example, F3-G shifts the target basket, and future representations may transmit its changed current location without the decoder needing information about future scene evolution. The donor-current/recipient-future combination can be inconsistent with native inference, a limitation explicitly acknowledged. This is consequential because the central motivation concerns the causal contribution of prediction.

Changing the denoising procedure can reasonably change action vectors, and failure of one shared scalar to explain those differences establishes only that they are not uniform rescalings. The LIBERO completion experiment evaluates recipient continuations, while the counterfactual responses are measured separately; it therefore does not establish how the changed responses affect execution after the physical perturbations. The authors appropriately acknowledge that response differences alone do not imply worse control.

Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer E7ss (Summary)

We thank the reviewer for the constructive feedback and for recognizing the value of our physical counterfactuals, source-control experiments, and late-layer restoration analysis. We provide additional controlled-intervention and paired-execution evidence addressing the two concerns.

  1. Controlled future-input dependence: In a same-current alternative-future pilot, changing future-origin K/V produces different action outputs while current K/V and action randomness remain fixed. This provides additional controlled evidence of dependence on future inputs.

  2. Reduced computation and perturbed-state execution: From 80 matched F3-G perturbed states derived from 20 source trajectories, both K=10 and K=5 complete 80/80 continuations. Across five fixed inference seeds, the budget-induced response discrepancy is smaller than native seed-pair variation, with an aggregated ratio of 0.683 [0.604, 0.777].

AuthorsOpenReview ↗

Author Response to Reviewer E7ss (Part 1)

“The experiments identify dependence on future-origin representations more convincingly than dependence on their predictive content.”

“For example, F3-G shifts the target basket, and future representations may transmit its changed current location without the decoder needing information about future scene evolution.”

“The donor-current/recipient-future combination can be inconsistent with native inference.”

Same-current alternative-future intervention. Alternative future-origin representations change action outputs while current conditions remain fixed. We tested this in a controlled pilot on five LIBERO source states, one from each of tasks 5–9. For each state, we compared three alternative future samples while keeping the current physical state, instruction, Current K/V, and action-decoder randomness fixed.

At each registered consumer event across the 30 layers and ten action-denoising steps, Current K/V were bit-identical across conditions. All 15 native replay checks reproduced the saved normalized action arrays element-wise. Separately, the three alternative future samples yielded three pairwise action comparisons per source, giving 15 within-source comparisons across five sources. The mean radial-command RMS difference, averaged first across the three comparisons within each source and then equally across the five sources, was 0.033714 in de-normalized translation-command units.

Because all alternative futures are generated from the same current state and instruction, this pilot controls the between-scene physical variation raised in the basket example and provides a more controlled comparison of future inputs.

Interpretation. The pilot strengthens the evidence for a causal influence of generated-future inputs on action generation. It does not separately identify static scene geometry and future scene evolution within those representations. We therefore interpret the result as evidence of action dependence on future-origin representations, rather than identification of the specific predictive content used by the policy.

AuthorsOpenReview ↗

Author Response to Reviewer E7ss (Part 2)

“Changing the denoising procedure can reasonably change action vectors, and failure of one shared scalar to explain those differences establishes only that they are not uniform rescalings.”

“The LIBERO completion experiment evaluates recipient continuations, while the counterfactual responses are measured separately.”

Paired closed-loop evaluation under physical perturbations. We conducted an additional post-hoc evaluation using 80 registered F3-G perturbed states from 20 source trajectories, selected from the original 50-source confirmation cohort. Both K=10 and K=5 were executed from each perturbed state under matched initial conditions, resulting in 160 closed-loop episodes.

Computation budgetSuccessful continuationsSuccess rate
K=1080/80100%
K=580/80100%

Both budgets completed all 80 perturbed-state continuations, with no discordant completion outcomes. This extends the original recipient-state evaluation to the physical states used in the counterfactual-response analysis.

Action-response variation under reduced computation. We compare donor-minus-recipient predicted action responses to the same physical perturbation under matched inference randomness. Across five fixed inference seeds, the mean RMS discrepancy between K=5 and K=10 de-normalized radial-command response vectors is 0.008186, compared with 0.011993 for native K=10 seed-pair variation. Their ratio is 0.683 (95% CI: [0.604, 0.777]). Thus, the budget-induced response discrepancy is smaller than the measured native seed-to-seed variation.

The original source-level cross-fitted shared-gain analysis tests a single uniform multiplicative scaling. It yields an estimated 3.29% reduction in held-out squared discrepancy (95% CI: [−9.05%, 13.63%]), providing little held-out improvement. Numerical approximation introduced by reducing denoising steps remains a possible contributor to the observed response differences.

Computation efficiency. Separately, the original 50-recipient continuation evaluation shows a median 23.7% reduction in cumulative policy-computation time when reducing world-branch denoising steps from K=10 to K=5.

Reduced world-branch computation changes the measured physical-response vectors while preserving task completion in this perturbed-state cohort.

I appreciate the authors' response. However, my main concern remains: whether these effects reflect the use of predicted future events, rather than current-scene information carried through those representations. I think the idea and the attempt is interesting, but I am not convinced by the approach, as other reviewers have pointed out.

Perhaps more importantly, I want to point out the writing is very difficult to read - I was unsure if this is because I was not very familiar with causal/counterfactual inference, but other reviewers have expressed the same things. Before we discuss about novelty, I think the paper needs to spend a lot of time reorganizing for clarity for a real discussion. Just to give some examples (to prevent this being thought of as AI generated), the abstract is filled with terms that are not defined. Fast-WAM-IDM is not defined but directly referred to, similarly Fast-WAM-Joint. How is predictive computation defined? Similar phenomena is happening in the rest of the paper as well. The main text requires readers to reconstruct the architecture and experimental sequence across sections. Even on the writing ground, I think this paper deserves another revision.

I will maintain my score and confidence.

Reviewer_dyGsOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper uses physical counterfactuals and activation patching to trace action responses through four frozen world-action model (WAM) configurations. In Fast-WAM-Joint, a substantial part of the response passes through future keys and values (K/V). Halving world-denoising steps changes these responses while reducing cumulative policy-computation time by 23.7%, with all 50 tested LIBERO continuations still completing.

Critical strengths
  • The experiments use careful controls, including matched inference noise, frozen weights, held-out states, and uncertainty estimates that account for repeated measurements from the same source.
  • The current/future interventions and layer-restoration experiments provide useful evidence about where action-relevant information flows.
Critical weaknesses
  • The paper is difficult to follow. The baseline configurations, conditioning inputs, and evaluation tasks need clearer explanations. Definitions and experimental details are scattered across the main text and appendix, making it hard to reconstruct the setup and compare results.

  • Patching future K/V establishes dependence on that pathway, but does not distinguish predicted physical information from current-state information carried through it. This limits what the experiments establish about the role of prediction itself.

  • Donor current K/V and recipient future K/V need not co-occur during normal inference. Decoder sensitivity to this mismatch remains a possible explanation for part of the measured effect.

  • Detailed source attribution is restricted to one architecture. The four-model held-out audit includes 544 perturbations per model, but these come from only 20 base states across two tasks. The restoration experiments also cover only two tasks.

  • The response differences are not linked to worse control or reduced robustness. The 29-pair RoboTwin evaluation has wide uncertainty and does not establish comparable success rates between compute budgets.

  • The failure of a shared scalar correction only rules out uniform rescaling. It does not distinguish a change in the use of predictive information from approximation error introduced by fewer denoising steps.

Questions for authors
  • Could you summarize each model’s conditioning inputs, patched tensors, and information available to the action decoder in one table? Why were these four configurations selected over other WAM baselines?

  • Why were LIBERO-Object and RoboTwin chosen? What complementary properties do they test, and why were the other LIBERO suites omitted?

  • Could you consolidate the tasks, state-selection procedures, and independent state or trajectory counts for each experiment?

  • Can you design a control that separates predicted future information from current-state information carried through future K/V?

  • How could you assess whether incompatible current/future inputs contribute to the measured intervention effects?

  • Do larger response discrepancies predict failures or reduced robustness in closed-loop rollouts from the perturbed states?

  • Does the late-layer restoration result generalize to additional tasks and another architecture?

  • How large are the reduced-compute response differences relative to variation across inference seeds at the native compute budget?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer dyGs (Summary)

We thank the reviewer for the careful assessment and for recognizing the value of our controlled interventions. We address the concerns in three parts, combining clearer experimental descriptions with additional intervention evidence and paired closed-loop results.

  1. Experimental setup and coverage: We summarize the conditioning inputs, patched tensors, configuration and benchmark choices, and cohort definitions. We explicitly distinguish source trajectories from repeated donor conditions and identify shared cohorts.

  2. Causal interpretation: A same-current alternative-future pilot shows that changing future-origin K/V changes actions while current K/V and action randomness remain fixed. We distinguish this controlled future-input dependence from identification of the predictive content used by the policy.

  3. Execution and validation: K=10 and K=5 each complete 80/80 continuations from perturbed states derived from 20 source trajectories. The budget-induced response discrepancy is smaller than measured native seed-pair variation. We also clarify the execution-outcome evidence and the task and architecture coverage of late-layer restoration.

AuthorsOpenReview ↗

Author Response to Reviewer dyGs (Part 1)

“Could you summarize each model’s conditioning inputs, patched tensors, and information available to the action decoder in one table? Why were these four configurations selected over other WAM baselines?”

“Why were LIBERO-Object and RoboTwin chosen? What complementary properties do they test, and why were the other LIBERO suites omitted?”

“Could you consolidate the tasks, state-selection procedures, and independent state or trajectory counts for each experiment?”

Configurations and inputs. The four configurations cover current-image conditioning, joint current/future processing, future-first action decoding, and cached image conditioning. Direct, Joint, and IDM provide comparisons within the Fast-WAM family; ImageWAM extends the comparison to a second model family.

All four LIBERO configurations receive current agent-view and wrist RGB images, the instruction, and an 8-D measured state: EEF position (3), axis-angle orientation (3), and gripper positions (2).

ConfigurationAction-decoder conditioningMain intervention
DirectCurrent-image K/V and direct language/state contextCurrent-image K/V
JointCurrent/future K/V and direct language/state contextCurrent-node K/V, allowing future propagation
IDMK/V derived from completed video latents and direct language/state contextFuture-latent slice; current slice retained and downstream K/V recomputed
ImageWAMImage K/V and fused text/state-prefix K/VImage K/V; recipient prefix retained

The Joint row describes the propagation-allowed main intervention in the 544-case audit. Its strict source-control experiments separately assign current/future K/V at every consumer event. ImageWAM’s image-plus-prefix replacement is reported as a full-interface control.

Benchmark choice. LIBERO-Object provides a shared pick-and-place structure, supporting consistent physical-factor interventions and phase definitions across the four configurations. We concentrated detailed matched attribution in this common setting. RoboTwin provides complementary checks on Direct and Joint through a different simulator, embodiment, three-camera input, and joint-target action interface. Its 14-D conditioning contains joint-drive targets and gripper commands, with FK providing geometric readouts of predicted targets. These checks complement LIBERO’s measured-state conditioning and end-effector action readouts.

Cohorts and selection. LIBERO task IDs below refer to LIBERO-Object.

ExperimentSource coverage and repeated observations
Discovery, tasks 0–450 model-specific successful native rollouts/model; 7,318 valid donor records across configurations
Held-out/prospective, tasks 5/9Shared 20 expert trajectories; 544 held-out conditions/configuration; 374 shared prospective conditions evaluated by all four configurations
Source controls, tasks 0–4Joint discovery trajectories reused: F1 has 35 retained states/174 records; F3-G has 47 trajectories/132 boundaries/526 radial donors
Restoration, tasks 0/120 new trajectories/103 donors; the six-window scan reuses this cohort
Budget confirmation, tasks 5–950 sources/200 donors; the supplement uses a 20-source/80-donor subset
RoboTwin mechanism/staticOriginal 34-source set nested within the expanded 42-source set
RoboTwin budget executionSeparately registered cohort: 29 preflight-eligible pairs from 50 candidates

Discovery uses fixed task/init IDs from successful native rollouts. Expert-derived cohorts use ordered init/demo IDs and phase/geometric eligibility. For each F1 state, a dose magnitude enters the factorial analysis when both signed perturbations pass geometric QC. Restoration states are captured at qualified native call boundaries before closing intent. RoboTwin uses fixed task/seed lists, with budget-execution eligibility assessed before either budget produces outputs.

Source units and repeated donor, phase, window, seed, and action-position observations are reported separately. The revised appendix will present the full 25-row cohort table and task catalog, including task names, selection and phase rules, registered and valid counts, exclusions, and experiment-specific clustering units.

AuthorsOpenReview ↗

Author Response to Reviewer dyGs (Part 2)

“Can you design a control that separates predicted future information from current-state information carried through future K/V?”

“How could you assess whether incompatible current/future inputs contribute to the measured intervention effects?”

Completed fixed-current intervention. Our Fast-WAM-Joint controls show that changing future-origin K/V changes actions under fixed current conditions. We evaluated five LIBERO-Object source states, one from each task 5–9, using three alternative future banks generated from the same current state and instruction. The current physical state, language/state context, and action randomness remained fixed. Current K/V were bit-identical across conditions at each corresponding consumer event.

All three pairwise action comparisons within each source showed differences, yielding 15 within-source comparisons across five sources. Averaging the radial-command RMS differences within each source and then equally across sources gives 0.033714 in de-normalized command coordinates. All 15 native replay checks reproduced their corresponding reference action arrays element-wise.

Paired source-background comparisons. We also evaluated current-side effects under both recipient and donor future backgrounds. In strict F1, dose-wise mean projection coefficients range from 0.0450–0.0740 with recipient future and 0.0458–0.0733 with donor future. All six paired background-difference intervals include zero.

F3-G retains the future-over-current ordering across both backgrounds and shows measurable background sensitivity. At −0.5 cm, the donor-minus-recipient future-background change in the current-effect coefficient is −0.0245 [−0.0472, −0.0044] (pointwise 95% CI). These paired comparisons quantify how the conditional effects vary with the source combination.

Conclusion. The fixed-current intervention establishes an action effect from changing future-origin inputs while the observed current conditions remain unchanged. The paired background analyses further show that the future-over-current ordering persists across the tested source backgrounds.

AuthorsOpenReview ↗

Author Response to Reviewer dyGs (Part 3)

“Do larger response discrepancies predict failures or reduced robustness in closed-loop rollouts from the perturbed states?”

“Does the late-layer restoration result generalize to additional tasks and another architecture?”

“How large are the reduced-compute response differences relative to variation across inference seeds at the native compute budget?”

Execution from perturbed states. We evaluated K=10 and K=5 from the same 80 registered F3-G perturbed states, drawn from a post-hoc 20-source subset of the original 50-source confirmation cohort. Response scores were frozen before rollout, and the 80 paired initial states and 160 first-call input, noise, and action records were verified.

Computation budgetSuccessful continuationsSuccess rate
K=1080/80100%
K=580/80100%

Both budgets completed all 80 continuations. Since no completion failures occurred, failure prediction cannot be assessed in this cohort. After task-and-signed-dose centering, the correlations with paired terminal object–goal distance differences and control-step differences are r=−0.024 [−0.307, 0.236] and r=0.119 [−0.201, 0.414], respectively. Both 95% confidence intervals include zero.

The separate RoboTwin evaluation contains 29/50 preflight-eligible episode pairs, with 18/29 native and 19/29 K5 successes. The task-equal K5−native difference is −0.027 [−0.267, 0.160]. The 21 shared exclusions—12 unstable resets and nine plan/expert-eligibility failures—occurred before either budget was evaluated. This conditional result does not establish comparable success rates.

Late-layer restoration coverage. The prespecified comparison uses 20 new Joint trajectories from LIBERO-Object tasks 0 and 1, with 103 valid donors. Radial response error is 0.248 under blocking, 0.247 with early restoration (layers 0–4), and 0.038 with late restoration (layers 25–29). The paired early-minus-late difference is 0.210 [0.184, 0.238]. The six-window scan refines localization using the same cohort. This strict restoration protocol has not yet been validated on additional tasks or another architecture.

Comparison with native seed variation. Across five fixed inference seeds, the mean RMS discrepancy between K5 and K10 de-normalized radial-command response vectors is 0.008186, compared with 0.011993 for native K10 seed-pair variation. Their aggregated ratio is 0.683 [0.604, 0.777]. Thus, the budget-induced discrepancy is smaller than the measured native seed variation.

The original source-level cross-fitted shared-gain analysis yields an estimated 3.29% reduction in held-out squared discrepancy [−9.05%, 13.63%], providing little improvement from uniform scaling. Numerical approximation remains a possible contributor to the response changes.

Separately, the original 50-recipient evaluation reports a median 23.7% reduction in policy-computation time. All bracketed intervals are 95% confidence intervals.

Conclusion. Reduced world-branch computation changes the measured physical-response vectors while preserving task completion in the evaluated perturbed-state cohort.

Comments on the paper

1 comment

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage in response periods so that we have our issues discussed thoroughly.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.