acceptodds
Under review as a conference paper at ICLR 2027

CausalWAM: Learning Action-Conditioned Visual Dynamics for Efficient World Action Models

14 views

Abstract

World Action Models (WAMs) jointly model robot actions and future visual observations, leveraging future visual dynamics as dense supervisory signals to learn physically grounded action and alleviate the sparse action-supervision bottleneck in Vision-Language-Action (VLA) models. However, this paradigm suffers from : although WAMs may generate plausible future observations, their predictions remain dominated by the current visual context and learned video priors, exhibiting only weak dependence to the conditioning action. To address this issue, we present , a causally grounded Mixture-of-Transformers (MoT) framework with a mixed training recipe. The recipe combines Action-Conditioned World Modeling (AC-WM), which uses clean action tokens to provide direct supervision for visual dynamics, with World Action Modeling (WAM) training, which preserves the joint action–video modeling pathway required for action generation. A progressive schedule is developed to emphasize AC-WM early to establish reliable action dependence and gradually shifts toward WAM to integrate the learned dynamics into action generation. By treating actions as conditioning variables rather than imitation targets, CausalWAM effectively exploits failed and suboptimal trajectories as transition-level supervision, thereby enhancing data efficiency. Experiments in simulation and on real robots demonstrate the effectiveness of CausalWAM in improving action dependence and policy performance, while system-optimized action-only inference reaches on an NVIDIA RTX . The code will be publicly released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Vision Language Models

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_mvvBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CausalWAM addresses action neglect in joint video-action models. It combines clean-action-conditioned video prediction with joint video-action generation, gradually shifting the training emphasis toward the latter. Failed and suboptimal trajectories are used for dynamics learning without treating their actions as imitation targets. The method is evaluated on RoboTwin and real-world instruction-following, long-horizon, and out-of-distribution tasks.

Critical strengths
  1. The use of suboptimal trajectories is well motivated. An unsuccessful action can still provide valid transition supervision, even when it is unsuitable as an imitation target. Assigning these trajectories to dynamics learning is a practical way to use otherwise discarded data.

  2. The combined objectives improve downstream control. Joint training reaches 75% success, compared with 71% for WAM-only and 59% for AC-WM-only. The real-world evaluations also go beyond short, in-distribution tasks and provide evidence of practical benefit.

Critical weaknesses
  1. The paper does not directly show that CausalWAM fixes causal neglect. Figure 1 illustrates weak action sensitivity, but there is no systematic comparison against the actual outcomes of different actions from the same state. Better policy performance alone does not verify the proposed mechanism.

  2. The novelty lies mainly in the training recipe, since the MoT architecture and attention mask follow prior work. However, the key schedule ablation is confounded: the fixed AC-WM ratio is 0.25, whereas the cosine schedule averages approximately 0.5. The gain may therefore come from more AC-WM training rather than the progressive schedule.

  3. The real-world success metric is unclear. The text-following score of 0.89 measures partial completion across four stages, rather than full-task success. Table 3 averages this score with binary success rates, making the aggregate result difficult to interpret. Tables 1 and 4 also disagree on randomized Scan Object for CausalWAM and randomized Turn Switch for FastWAM.

  4. The contribution beyond existing unified training frameworks remains unclear. UWM[1] and Motus[2] already combine conditional dynamics modeling and action generation. The main additions here appear to be the curriculum and quality-aware data routing. The former is confounded by the mixing ratio, while the latter improves success from 73% to 75% without clearly establishing matched data exposure.

References

[1]. Bi, H., Tan, H., Xie, S., Wang, Z., Huang, S., Liu, H., Zhao, R., Feng, Y., Xiang, C., Rong, Y., Zhao, H., Liu, H., Su, Z., Ma, L., Su, H., & Zhu, J. (2026). Motus: A unified latent action world model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 35101–35113).

[2]. Zhu, C., Yu, R., Feng, S., Burchfiel, B., Shah, P., & Gupta, A. (2025). Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets. In Robotics: Science and Systems XXI.

Questions for authors
  1. Does the progressive schedule outperform a fixed AC-WM ratio of 0.5 under matched total updates and data exposure?
  2. Can you quantify whether action changes produce the correct changes in predicted outcomes, rather than merely increasing action sensitivity?
  3. Does the suboptimal-data ablation hold total data exposure constant? What proportion of training data is suboptimal, and do the baselines have access to the same data?
  4. Were the latency baselines evaluated with the same caching, compilation, precision, and deployment optimizations?
5 comments
AuthorsOpenReview ↗

Response to Reviewer mvvB (1/2)

W1&Q2: Quantitative Validation of Causal Neglect. We conducted a controlled counterfactual experiment to directly evaluate whether predicted changes agree with the observed consequences of different actions from the same state. For each of Open Microwave and Turn Switch in RoboTwin, we evaluated 50 matched no-op/expert trajectory pairs, with each pair starting from the same initial state. CausalWAM, WAM-only, and FastWAM were evaluated at the shared horizons t={12,24}t=\{12,24\} using four diffusion sampling seeds per state, with identical initial noise for the paired no-op and expert predictions within each model.

To isolate action-induced changes from the static background, we subtract the predicted no-op video from the predicted expert-action video and perform the same subtraction on the corresponding observed videos. We compare the predicted and observed differences using Normalized Effect Error (NEE) and Effect Cosine Similarity (ECS). As shown in the table bew, CausalWAM achieves lower mean NEE and higher mean ECS than WAM-only and FastWAM on both tasks. This directly shows that its predicted action-induced changes agree more closely with the observed outcomes, rather than merely producing different videos when the action changes.

These results provide quantitative evidence that CausalWAM mitigates causal neglect on the evaluated tasks. In the revised manuscript, we provide systematic visualizations of the observed outcomes and the predictions from CausalWAM, WAM-only, and FastWAM, together with the complete experimental protocol and quantitative analysis in Section 4.4.

TaskModelNEE ↓ECS ↑
Open MicrowaveCausalWAM0.860.37
WAM-only0.890.28
FastWAM0.990.07
Turn SwitchCausalWAM0.630.76
WAM-only0.730.71
FastWAM1.06−0.05

W2&Q1: Exposure-Matched Schedule Control. To separate the effect of the progressive schedule from total AC-WM exposure, we added a fixed-ρAC=0.5\rho_{\mathrm{AC}}=0.5 control under the same training setup as the cosine 1-to-0 schedule. The two settings use the same total number of updates and have matched expected AC-WM/WAM exposure. The fixed-0.5 control achieves a 60% average binary success rate, compared with 72% for the cosine schedule, indicating that the improvement cannot be explained solely by greater AC-WM exposure and supporting the benefit of progressively shifting the training emphasis from AC-WM to WAM.

W3: Real-World Metric Clarification. We agree that labeling the original aggregate as “SR” was imprecise. In the revised table, we additionally report the strict full-task success rate for Text-following, while retaining the stage-completion score in parentheses as a secondary metric. We also recompute all aggregate results using the macro-average binary success rate across the three real-world task categories: Text-following, Long-horizon, and OOD.

W3: Table Consistency. The values in Table 4 are correct: the randomized success rates are 96% for CausalWAM on Scan Object and 59% for FastWAM on Turn Switch. The corresponding entries in Table 1 are transcription errors. We thank the reviewer for identifying these inconsistencies. The 50-task averages were computed using the correct per-task values in Table 4; therefore, correcting the two entries in Table 1 does not change the reported averages or numerical rankings. We will correct both entries in Table 1 and rigorously cross-check all reported values in the revised manuscript.

W4: Distinction from Existing Unified Training Frameworks. We agree that UWM and Motus already unify visual dynamics modeling and action generation, and we do not claim the MoT backbone, attention-mask structure, or unified video-action formulation as our novelty. Our contribution specifically targets causal neglect in action-centered WAM training. We introduce a training recipe that combines clean-action-conditioned future prediction with standard joint WAM training and progressively shifts the training emphasis from the former to the latter. This design is intended to first establish action-dependent visual dynamics and then integrate this dependence into the action-generation pathway.

AuthorsOpenReview ↗

Response to Reviewer mvvB (2/2)

W4&Q1&Q3: Progressive Schedule and Quality-Aware Data Routing. As described in our response to W2&Q1, the exposure-matched fixed-ρAC=0.5\rho_{\mathrm{AC}}=0.5 control supports the benefit of the progressive schedule beyond average objective exposure. To clarify the data exposure in the quality-aware data-routing ablation, we included approximately 26.9 robot-hours of suboptimal data (about 1% of the pretraining data by duration) and kept the total number of training steps identical. As described in Appendix A.3, these trajectories arise from failed or unrecoverable executions during data collection. Because the suboptimal data are used during pretraining, and the other baselines lack a mechanism for learning from failed trajectories without treating their actions as imitation targets, we did not expose these data to the baselines.

Q4: Runtime Optimization and Deployment Efficiency. The baselines were evaluated using their released inference implementations rather than being retrofitted with an identical optimization stack. For example, FastWAM already incorporates video KV caching, whereas graph compilation and C++ deployment are additional optimizations applied to CausalWAM. We use these latency comparisons to show that CausalWAM achieves competitive latency among WAMs. With additional engineering optimizations, its latency becomes comparable to that of the evaluated VLA baseline, highlighting the potential of WAMs for responsive real-world deployment.

AuthorsOpenReview ↗

Response to Reviewer mvvB

Thank you again for your time and thoughtful review. As the discussion period draws to a close, we would appreciate hearing whether our responses and revisions have addressed your main concerns. If any questions remain, we would be happy to provide further clarification before the discussion closes.

Thank you for considering our responses.

Best regards,
The Authors

Reviewer_qvR7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies weak action dependence in action-conditioned world action models and proposes CausalWAM, a Mixture-of-Transformers architecture trained with a mixture of two objectives. In AC-WM updates, clean ground-truth actions condition future-video prediction, and only the video prediction loss is optimized. In WAM updates, the model jointly learns action and video denoising. A progressive schedule reduces the fraction of AC-WM updates from 1 to 0. This separation allows failed or suboptimal trajectories to provide transition supervision through AC-WM without using their actions as policy imitation targets. At inference, the model generates actions without generating future video. Evaluation covers 50 RoboTwin tasks and real-world text-following, long-horizon, and out-of-distribution settings.

Critical strengths
  1. Separating transition supervision from imitation targets is well motivated. AC-WM provides a coherent way to use failed or recovery trajectories for visual-dynamics learning without directly training the policy to imitate their actions. The mixed formulation retains an action-generation objective and an inference path that does not require future-video generation. This offers a useful approach to exploiting transition information that would otherwise be discarded.

  2. The pretraining comparison provides initial evidence of complementarity between the objectives. Figure 4(a) reports a downstream success rate of 75% for mixed pretraining, compared with 71% for WAM-only and 59% for AC-WM-only, although the subsequent training protocol needs clarification. Evaluation spans 50 simulated tasks and several real-world settings. The latency breakdown also distinguishes the contributions of caching, compilation, and C++ deployment, making the reported efficiency improvement more interpretable.

Critical weaknesses
  1. The explanation in terms of repairing “causal neglect” lacks direct quantitative validation. Figure 1(b) provides a qualitative failure example, but does not identify the tested model, task, sample count, or sampling-noise controls, and does not show a corresponding CausalWAM result. The paper does not quantitatively compare WAM-only and CausalWAM under controlled changes to the conditioning action, or verify that prediction changes agree with observed consequences. The attention mask permits action use but does not demonstrate learned action dependence. Improved policy success therefore supports the recipe’s practical value more directly than it supports the proposed mechanism.

  2. The evidence for sample efficiency and consistent superiority needs qualification and verification. The suboptimal-data experiment reports an increase from 73% to 75% without specifying the amount of suboptimal data, whether update budgets were matched, or uncertainty across training runs. This alone does not establish substantial sample-efficiency gains or a benefit specific to the routing mechanism. In simulation, the reported advantages over FastWAM are 0.48 percentage points under clean conditions and 0.06 under randomized conditions, without uncertainty estimates. Tables 1 and 4 also disagree on randomized CausalWAM/Scan Object, reported as 72 versus 96, and FastWAM/Turn Switch, reported as 95 versus 59. The original values and resulting 50-task averages therefore need confirmation before drawing conclusions from these small margins.

Questions for authors
  1. Is quantitative evidence available comparing WAM-only and CausalWAM under controlled changes to the conditioning action, including whether predicted changes agree with observed outcomes? Please identify the model, task, sample count, action interventions, and sampling-noise protocol used in Figure 1(b).

  2. For Figure 4(b,c), please specify the evaluation tasks, trial counts, training runs, suboptimal-data quantity, and whether update budgets were matched. Is an exposure-matched comparison available? What evidence beyond the 73% to 75% improvement supports the sample-efficiency claim?

3 comments
AuthorsOpenReview ↗

Response to Reviewer qvR7

W1&Q1: Quantitative Evaluation of Causal Neglect. We designed a controlled experiment to quantitatively evaluate counterfactual prediction accuracy. For each of Open Microwave and Turn Switch in RoboTwin, we evaluated 50 matched no-op/expert trajectory pairs (100 trajectories per task), with each pair starting from the same initial state. Predictions were evaluated at t={12,24}t=\{12,24\} using four diffusion sampling seeds, with identical initial noise for expert and no-op predictions within each model.

To focus on action-induced visual changes, we subtract the predicted no-op video from the predicted expert-action video and perform the same subtraction on the observed videos. We compare these differences using Normalized Effect Error (NEE), which measures pixel-level prediction error normalized by the magnitude of the observed difference, and Effect Cosine Similarity (ECS), which measures agreement in the pattern of pixel changes. As shown in the table below, CausalWAM achieves lower NEE and higher ECS than WAM-only and FastWAM on both tasks. These results provide quantitative evidence that CausalWAM mitigates causal neglect on the evaluated tasks. Its predictions better match the observed consequences of action changes, rather than merely exhibiting greater action sensitivity.

Figure 1(b) presents a qualitative example from GigaWorld-Policy on the Adjust Bottle task in RoboTwin. We provided a new, systematic visualization of observed outcomes and predictions from CausalWAM, WAM-only, and FastWAM under matched interventions in the revision.

TaskModelNEE ↓ECS ↑
Open MicrowaveCausalWAM0.860.37
WAM-only0.890.28
FastWAM0.990.07
Turn SwitchCausalWAM0.630.76
WAM-only0.730.71
FastWAM1.06−0.05

W2&Q2: Use of Suboptimal Data. Our pretraining dataset includes 26.9h of suboptimal trajectories (~1% by duration), used exclusively for AC-WM. With the same number of training updates, including this subset improves the revised average binary success rate from 69% to 72%. These results suggest that otherwise discarded transitions can provide useful supervision, supporting improved data utilization.

W2: Simulation Results and Table Consistency. The values in Table 4 are correct: the randomized success rates are 96% for CausalWAM on Scan Object and 59% for FastWAM on Turn Switch. The corresponding entries in Table 1 are transcription errors. The corrections leave the reported averages and numerical rankings unchanged. Across the full 50-task benchmark, CausalWAM achieves the highest reported mean performance under both clean and randomized settings.

Q2: Ablation Protocol. Figure 4(b,c) uses the real-world evaluation suite reported in Tables 2–3: six text-following instructions, one long-horizon task, and three OOD conditions, with 10 trials per individual setting. Each ablation configuration uses one training run. In the revised manuscript, we consistently report the macro-average binary full-task success rate across the three evaluation categories.

Q2: Exposure-Matched Schedule Ablation. We added a fixed-ρAC=0.5\rho_{\mathrm{AC}}=0.5 control under the same training setup, matching the cosine schedule’s average objective exposure. It achieves 60% average binary success, compared with 72% for the cosine schedule. This comparison supports a benefit from the progressive schedule beyond average objective exposure.

AuthorsOpenReview ↗

Response to Reviewer qvR7

Thank you again for your time and thoughtful review. As the discussion period draws to a close, we would appreciate hearing whether our responses and revisions have addressed your main concerns. If any questions remain, we would be happy to provide further clarification before the discussion closes.

Thank you for considering our responses.

Best regards,
The Authors

Thanks authors for the response. After considering your clarifications in light of my own assessment and the key points raised by the other reviewers, I have decided to keep my original score.

Reviewer_dtwuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CausalWAM combines action-conditioned world modeling and joint world-action modeling through a scheduled mixture. It uses suboptimal trajectories as transition data rather than imitation targets, and decodes actions without generating future video at inference. Evaluation covers RoboTwin 2.0 and real-robot instruction following, long-horizon tasks and distribution shifts. A separate runtime-optimization sequence reduces reported deployment latency to 85 ms.

Critical strengths
  1. Learning action consequences from suboptimal data without imitating those actions is a sensible distinction. Retaining future supervision in training while avoiding future-video generation at inference is also a useful deployment direction.
  2. The full 50-task simulation table is available, and all ten mean columns in Table 4 reproduce. The latency reduction is decomposed into caching, compilation and C++ deployment rather than presented solely as an unexplained final speedup. These details provide useful empirical evidence despite the concerns below.
Critical weaknesses
  1. The controls do not isolate the central causal/scheduling claim. Qualitative sensitivity to action changes is not evidence of quantitatively correct counterfactual prediction. The fixed mixture uses rho_AC=0.25, while the 1-to-0 schedules have approximately 0.5 average exposure, changing objective budget as well as order. The 75% versus 73% suboptimal-data effect also lacks independent-training uncertainty and complete data-volume accounting. As a result, the evidence does not yet identify which component causes the benefit.
  2. Several reported comparisons need reconciliation. Randomized Scan Object for CausalWAM is 72 in Table 1 but 96 in Table 4; randomized Turn Switch for FastWAM is 95 versus 59. These are confirmed table conflicts even though Table 4's means are correct. The real-robot aggregate labeled SR mixes partial-credit instruction-following scores with binary task success, and the fastest latency uses a runtime optimization level not clearly matched to the baselines. Both issues obscure the meaning of headline comparisons.
Questions for authors
  1. Which values are correct for the two conflicting cells? Please provide the corresponding run-level outputs and aggregation script, and state whether any conclusions or aggregate comparisons change.
  2. Does a fixed rho_AC=0.5 control with matched objective exposure retain a gap to the schedule? Please also specify the suboptimal-data counts and distinguish action sensitivity from counterfactual prediction accuracy. Available controlled results or appropriately narrower claims would materially affect my assessment.
5 comments
AuthorsOpenReview ↗

Response to Reviewer dtwu

W1: Quantitative Counterfactual Evaluation. We designed a controlled experiment to quantitatively evaluate counterfactual prediction accuracy across different WAMs. For each of Open Microwave and Turn Switch in RoboTwin, we collected 50 matched no-op/expert trajectory pairs (100 trajectories per task). The two trajectories in each pair start from the same initial state: in the no-op trajectory, the robot is commanded to hold its current pose; in the expert trajectory, it executes the original expert actions. We evaluated CausalWAM, WAM-only, and FastWAM at the shared horizons t={12,24}t=\{12,24\}, using four matched diffusion seeds.

To focus on action-induced visual changes and reduce the influence of the static background, we subtract the predicted no-op video from the predicted expert-action video for each matched initial state and perform the same subtraction on the corresponding observed videos. We compare these predicted and observed visual differences using two complementary metrics: Normalized Effect Error (NEE), which measures their pixel-level root mean squared error normalized by the magnitude of the observed difference; and Effect Cosine Similarity (ECS), which measures whether they exhibit the same pattern of increases and decreases across pixels, independently of their overall scale. This evaluation tests whether changing the action produces the correct predicted visual changes, rather than merely different videos.

TaskModelNEE ↓ECS ↑
Open MicrowaveCausalWAM0.860.37
WAM-only0.890.28
FastWAM0.990.07
Turn SwitchCausalWAM0.630.76
WAM-only0.730.71
FastWAM1.06−0.05

As shown above, CausalWAM achieves lower NEE and higher ECS than both baselines on both tasks, indicating more accurate prediction of action-induced visual changes. FastWAM shows limited agreement with the observed intervention effects due to its attention mask. Compared with WAM-only, CausalWAM’s stronger results are consistent with the benefits of explicit clean-action-conditioned video supervision for learning action-dependent visual dynamics. These findings provide quantitative evidence of counterfactual prediction accuracy on these two tasks, beyond merely demonstrating action sensitivity. We added the complete metric definitions, experimental analysis, and qualitative visualizations to Section 4.4 of the revised manuscript.

W1&Q2: Exposure-Matched Schedule Ablation. We added a fixed ρAC=0.5\rho_{\mathrm{AC}}=0.5 control under the same training setup. It achieves 60% average success, compared with 72% for the cosine 1-to-0 schedule, suggesting that the progressive schedule benefits performance beyond matching the average objective exposure.

W1&Q2: Use of Suboptimal Data. Our pretraining data include 26.9h of suboptimal trajectories (~1% by duration), with the remainder being high-quality data. These trajectories are used exclusively for AC-WM. Under the same number of training updates, including this subset improves success from 69% to 72%, suggesting that otherwise discarded transitions can provide useful supervision without observed aggregate performance degradation. Both results use the corrected binary success metric.

W2&Q1: Correction of Inconsistent Table Entries. Thank you for identifying these inconsistencies. The values reported in Table 4 are correct. The conflicting entries in Table 1 are transcription errors: the correct randomized success rates are 96% for CausalWAM on Scan Object and 59% for FastWAM on Turn Switch. The reported averages, model comparisons, and conclusions remain unchanged. We will correct these entries in Table 1 in the revision and rigorously cross-check all reported results.

W2: Clarification of Real-Robot Success Metrics. We add the strict full-task success rate for the Text Following task in the revised table. In the revised manuscript, all overall performance comparisons and ablation studies consistently use the average binary full-task success rate.

W2: Runtime Optimization and Deployment Efficiency. The evaluated implementations use different runtime optimizations; for example, FastWAM already incorporates video KV caching. These reported latencies serve as practical, implementation-level reference points for the evaluated methods. CausalWAM achieves competitive inference latency among WAMs, while additional engineering optimizations reduce its latency to 85 ms on an RTX 4090, comparable to the evaluated VLA baseline. This highlights the potential of WAMs for responsive real-world deployment. We will explicitly specify the runtime optimization settings for each method in the revised table.

AuthorsOpenReview ↗

Response to Reviewer dtwu

Thank you again for your time and thoughtful review. As the discussion period draws to a close, we would appreciate hearing whether our responses and revisions have addressed your main concerns. If any questions remain, we would be happy to provide further clarification before the discussion closes.

Thank you for considering our responses.

Best regards,
The Authors

Thank you for the exposure-matched schedule control and quantitative counterfactual evaluation. These address important parts of my original concerns. However, the suboptimal-data gain still lacks independent-training uncertainty. I also share mvvB’s remaining concerns about unequal access to suboptimal data and unmatched deployment optimizations, which limit attribution of the routing and latency benefits. While I appreciate the additional experiments and corrected reporting, these unresolved issues prevent me from changing my overall assessment. I will maintain my rating of 2.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers and Summary of Revisions

We thank all three reviewers for their thoughtful and constructive feedback. We appreciate their recognition of strengths of our work: (i) separating transition supervision from action imitation to exploit suboptimal trajectories (dtwu, qvR7, mvvB); (ii) the practical benefits of combining dynamics learning and action generation (qvR7, mvvB); and (iii) retaining future-video supervision during training without requiring future-video generation at inference, together with the decomposed efficiency analysis (dtwu, qvR7).

The reviews also raise several recurring concerns, which we address as follows:

  1. Quantitative validation of action consequences.
    We added a quantitative counterfactual evaluation in Section 4.4 on Open Microwave and Turn Switch in RoboTwin, using 50 matched no-op/expert trajectory pairs per task and four diffusion seeds. NEE and ECS compare predicted action-induced visual changes against observed changes, with full metric definitions in Appendix A.1. CausalWAM improves both metrics over WAM-only and FastWAM on both tasks. We also added matched qualitative visualizations in Figure 5 and Appendix A.2, and scoped our conclusions to the evaluated tasks.

  2. Separating schedule benefits from objective exposure.
    We added a fixed-ρAC=0.5\rho_{\mathrm{AC}}=0.5 control in Section 4.3 and Figure 4(b), using the same training update budget and matching the cosine schedule’s expected objective exposure. It achieves 60% average binary success, compared with 72% for the cosine schedule. This comparison supports a benefit from progressive scheduling beyond the average mixing ratio.

  3. Clarifying the evidence from suboptimal data.
    We clarified in Section 4.3 and Appendix A.4 that the pretraining dataset includes 26.9 hours of suboptimal trajectories, approximately 1% by duration, used exclusively for AC-WM. With the same number of training updates, including this subset improves the revised average binary success rate from 69% to 72%, as reported in Figure 4(c). We also revised the corresponding interpretation to emphasize improved data utilization rather than a broad sample-efficiency claim.

  4. Reconciling table entries and success metrics.
    We corrected the two conflicting simulation entries in Table 1; the reported simulation averages and numerical rankings remain unchanged. We standardized all overall real-world comparisons and ablations in Tables 2–3 and Figure 4 to macro-average binary full-task success rates, retaining stage-completion scores only as auxiliary diagnostics. We updated the corresponding analyses in Section 4.3 and clarified the metric definitions in Appendix A.1. The revised real-world and ablation values reflect this metric correction.

  5. Clarifying runtime comparability.
    We clarified the interpretation of the latency results in Section 4.3 and the description of Table 3. Since the evaluated implementations use different runtime optimizations, these latencies serve as implementation-level reference points. The 85 ms result represents optimized deployment performance rather than an optimization-matched comparison of model architectures.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.