acceptodds
Under review as a conference paper at ICLR 2027

Utility-WAM: Correcting Future-Induced Action Changes

Abstract

World Action Models (WAMs) can condition robot actions on generated visual futures, but this additional context is neither free nor uniformly beneficial. We study what a predicted future actually changes: the action delta between the future-conditioned and future-free paths of the same model. Under a shared WAM checkpoint, this delta barely moves average success ( points on LIBERO, on RoboTwin 2.0), while generating the future increases inference latency by 1.7–1.9 on every query. The averages hide a concentrated failure mode that we call future-induced negative transfer: the delta helps on long-horizon tasks ( points on LIBERO-Long), is negligible on most (26 of 50 RoboTwin 2.0 tasks change by at most one point), and fails severely on a few. Press Stapler and Place Can Basket lose 54.0 and 34.0 points and are also among the three most harmed tasks for both future-generating FAST-WAM variants. We introduce Utility-WAM for in-place correction of WAMs that retain future-conditioned inference. Rather than improving a current-only student overall, it targets concentrated failures without discarding the long-horizon benefit of generated futures. Its main learning component is a residual adapter that repairs future-induced action changes while keeping the shared WAM frozen. It aligns induced velocity deltas with beneficial ones induced by ground-truth futures and penalizes local error beyond the future-free reference. As intended, the adapter acts selectively: the two severe-task losses shrink to 6.0 and 14.0 points, while the remaining 48 RoboTwin tasks improve by 0.35 points on average and the long-horizon gain is retained. Because this correction is soft and still pays the video cost on every query, a lightweight gate reuses the adapter's paired outcomes to skip futures that correction does not make useful. It generates video on only 24.5% of RoboTwin queries, reduces latency by 38%, and improves over always-on correction (93.3% versus 92.8%; 98.9% versus 98.7% on LIBERO). On three real-world bimanual tasks, Utility-WAM succeeds in 53 of 60 trials, compared with 43 for FAST-WAM.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Visual Future Prediction, Action Correction Alignment, Robotic Manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ADhQOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies when visual future prediction actually helps World Action Models (WAMs). The authors show that future conditioning has only a small effect on average performance, but can cause large drops on a few precision-sensitive tasks. They call this future-induced negative transfer. To address this, the paper proposes Utility-WAM: a lightweight residual adapter corrects harmful future-induced action changes, and a gate decides when future generation is worth the extra computation. Experiments on LIBERO, RoboTwin 2.0, and real-world bimanual tasks show improved success rates while reducing unnecessary video generation.

Critical strengths
  1. Clear and useful problem formulation. The comparison between future-conditioned and future-free paths under the same checkpoint is well controlled, and the paper goes beyond benchmark averages to analyze task-level failures. This makes the negative-transfer observation quite convincing within the evaluated model.

  2. The method is simple and practical. The adapter is lightweight, keeps the base WAM frozen, and the ablations clearly separate the effects of alignment, error regularization, and routing. The real-world experiments are also a nice addition.

Critical weaknesses
  1. The main negative-transfer result is still based on one shared Stage-1 checkpoint. The strongest failures on RoboTwin are also heavily concentrated on two tasks. The FAST-WAM comparisons help, but those models are trained separately, so they are not a fully controlled replication. It would be useful to see whether the same failure pattern appears across multiple independently trained checkpoints.

  2. The gate supervision could be better justified. The gate is trained from an action-MSE-based utility label, while the actual goal is closed-loop task success. I would like to see more evidence that this proxy reliably predicts when future generation is actually useful.

Questions for authors
  1. How stable is the negative-transfer pattern across independently trained Stage-1 checkpoints? In particular, do Press Stapler and Place Can Basket remain the most strongly affected tasks?

  2. How well does the gate's action-error utility correlate with closed-loop success? Reporting AUROC/AUPRC or some correlation analysis would help clarify this.

  3. Could you compare against a gate trained on the original uncorrected future-conditioned branch? This would help isolate whether the adapter-derived supervision is actually important for routing.

2 comments

Thank you for the helpful review. Revisions incorporated in the manuscript are marked in red. The new checkpoint and retraining results below are reported only in this response.

1. Weakness 1: Single-checkpoint evidence and task concentration.

We evaluated an additional independently trained Stage-1 checkpoint. Within each checkpoint, both branches share weights, with adapter and gate disabled. Changes below are future-conditioned minus future-free success, in percentage points.

Stage-1 checkpointRoboTwin 2.0 averagePress StaplerPlace Can Basket
Original-0.8-54.0-34.0
Independently trained replication-0.7-59.0-32.5

Both checkpoints reproduce small aggregate effects and severe losses on Press Stapler and Place Can Basket, still ranked first and second. This controlled replication complements the FAST-WAM comparisons.

2. Weakness 2: Justification of action-MSE-based gate supervision.

Section 3.3 explains why we use paired action-error utility: a query's episode-level effect is difficult to observe offline because subsequent routing and execution also affect success. The adapter is the core component correcting future-induced Δa\Delta a; the gate uses whether its corrected action reduces expert-action MSE relative to the future-free reference (Eq. 5).

Table 3 reports that gating raises LIBERO success from 98.6% to 99.0% and RoboTwin success from 92.7% to 93.7%, while reducing NeffN_{\mathrm{eff}} from 10 to 8.70 and 2.45, respectively. We also retrained the adapter and gate with two additional seeds, keeping Stage-1 fixed. Across three training seeds (mean ± sample standard deviation, %):

ConfigurationLIBERORoboTwin 2.0
Correction98.57 ± 0.1592.67 ± 0.23
Selective99.03 ± 0.1593.67 ± 0.32

Selective improves over Correction in every seed on both benchmarks. Separately, Fig. 2 shows better success–compute curves than random routing and static truncation at intermediate budgets.

On Place Can Basket, gating increases the deficit from 14 to 17 points (Table 10). Chunk-averaged MSE may not fully reflect precise final placement, potentially explaining this decrease. Appendix C.3 discusses the task-level proxy gap. Supervising the gate with episode-level rollout outcomes remains future work (Section 6).

3. Question 1: Negative-transfer stability across independent Stage-1 checkpoints.

Aggregate success rates of the two independently trained Stage-1 checkpoints differ by less than 1 percentage point; Press Stapler and Place Can Basket remain the two most harmed tasks in the same order (W1).

4. Question 2: Discrimination and closed-loop evidence.

AP exceeds prevalence on both benchmarks, supporting offline utility-label ranking.

Appendix B, Table 7 reports offline discrimination using existing gate scores on cached validation/model-selection inputs, without retraining. Each branch's cached action MSE is averaged over two paired inference draws before applying y=1[uact>0]y=\mathbf{1}[u_{\mathrm{act}}>0] (Eq. 5; Section 3.3):

BenchmarkAUROCAPPositive-class prevalence
LIBERO0.59760.62720.5447
RoboTwin 2.00.58280.48080.4099

This cached-label diagnostic is reported separately from threshold calibration and uses no closed-loop-return labels. In closed-loop deployment, gating improves success while reducing video computation (W2). A single Δa\Delta a's effect depends on later routing and execution; random routing adds stochastic branch choices. Supervising the gate with episode-level rollout outcomes remains future work (W2).

5. Question 3: Comparison with a gate trained on the uncorrected branch.

Adapter-only achieves higher success than gate-only on both benchmarks. The full framework further improves success while reducing computation relative to adapter-only.

Gate-only is included in Table 3 and Fig. 2; Appendix C.8 (Tables 14–15) reports aggregate success/computation and the two severe-task results. We train the same gate architecture using utility labels from the uncorrected future-conditioned and future-free branches, keeping Stage-1 frozen and omitting the adapter during label generation, training, and inference.

ConfigurationLIBERO (%)RoboTwin 2.0 (%)NeffN_{\mathrm{eff}} LIB./RT
Gate-only98.392.29.21/0.83
Adapter-only (Correction)98.692.710/10
Full framework99.093.78.70/2.45

This supports the combined design of correction and its corresponding gate supervision; the different branches and budgets prevent isolating label-source effects.

We hope these clarifications address your concerns. We welcome further discussion.

Thank you for your answer. My concerns have been resolved, and I will keep my original score.

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

Utility-WAM studies the action changes induced by generated visual futures, including cases where they hurt a World Action Model. The same checkpoint supports future-conditioned and future-free execution, so the comparison holds the trained backbone fixed. A residual adapter learns from useful ground-truth-future velocity changes, with a penalty for error exceeding the future-free reference. Before generating video, a separate gate predicts whether the corrected future-conditioned path will reduce action error. Experiments on LIBERO, RoboTwin 2.0, and three physical manipulation tasks examine severe task-level failures that are obscured by benchmark averages.

Critical strengths

I found the task-level diagnosis the most useful part. The small average effect in Table 2 hides losses of 54 and 34 points on Press Stapler and Place Can Basket in Table 9. Correction brings these down to 6 and 14 points while retaining the LIBERO-Long benefit. These tasks also suffer severe losses in independently reported FAST-WAM variants, which gives the finding more weight than choosing failures from this checkpoint alone. The authors are careful to distinguish this corroboration from the controlled comparisons using a shared checkpoint.

Table 3 uses the same adapter with action loss alone, which barely improves results, and separately tests alignment and excess-error regularization. Figure 2 then compares learned routing against random routing and static video truncation at matched video-step budgets. These controls address the method's components fairly directly. The measured RoboTwin latency drops from 528.7 to 328.2 ms, although future-free inference is still faster. For the physical experiments, Table 5 includes shared-backbone and correction-only baselines. This makes it possible to assess more than the headline gain over FAST-WAM.

Critical weaknesses

I would most like to see a learned gate on the uncorrected shared WAM. The experiments show that correction helps always-video inference and that routing helps the corrected model. How much could learned routing achieve on its own? The random-gate control does not answer this. Avoiding harmful future conditioning is a natural alternative, especially for the two severe tasks, and this comparison would establish how much correction contributes to the final selective system.

One paired stochastic sample per input supplies the gate's label: the sign of action-MSE improvement. Appendix B explains threshold calibration on held-out data, but I am still missing an analysis of how well that proxy predicts closed-loop benefit. Gating actually worsens the remaining losses on the two severe tasks, from 6 to 6.5 points and 14 to 17 points, while improving the benchmark average. There is no retraining uncertainty estimate for the smaller aggregate gate improvements, and the paper acknowledges the single-backbone limitation. My recommendation is weak accept for the task-level diagnosis and recovery. The smaller routing gains need more evidence.

Questions for authors
  1. Could you train the same gate architecture to choose between the uncorrected future-conditioned and future-free paths? Please report both the two severe tasks and the aggregate accuracy/compute comparison.

  2. Where does predicted action-error utility disagree with rollout benefit? I am particularly interested in Place Can Basket, where selective deployment still has a 17-point deficit.

  3. Are the smaller gate and loss-ablation gains stable when the adapter and gate are retrained, keeping the expensive Stage-1 checkpoint fixed?

1 comment

We appreciate your detailed assessment and positive evaluation of the task-level diagnosis and recovery results. Revisions incorporated in the manuscript are marked in red.

Weakness 1: Learned Gating without the Adapter

We add a gate-only baseline that learns to route between the uncorrected future-conditioned and future-free branches, keeping the shared Stage-1 checkpoint frozen and omitting the adapter.

ConfigurationRoutingAdapter objectiveLIBERO (%)RoboTwin 2.0 (%)NeffN_{\mathrm{eff}} LIB./RT
Shared WAM + GateLearned--98.392.29.21/0.83

At the reported operating points, the full method achieves 99.0% and 93.7%, improving over gating alone by 0.7 and 1.5 percentage points. This baseline is included in Table 3 and Appendix C.8 (Table 14), with success–compute curves in Fig. 2.

Weakness 2: Utility Proxy and Reliability of Routing Gains

The adapter corrects future-induced Δa\Delta a, and the gate reuses the corrected action-error utility. A query's effect on episode success is difficult to observe in offline training because it also depends on later routing and execution. We therefore use paired action-error utility as supervision. Each comparison shares the input and initial action noise, measuring corrected future-conditioned versus future-free expert-action MSE (Eq. 5; Sec. 3.3); its label is a sampled utility estimate.

This signal has positive closed-loop deployment results. On RoboTwin, adding the gate raises average success from 92.7% to 93.7% while reducing NeffN_{\mathrm{eff}} from 10 to 2.45 (Sec. 4.3, Table 3). The success–compute curves also favor learned routing over random routing and static video truncation across intermediate budgets (Fig. 2).

For this response, we retrained the adapter and gate with two additional seeds, keeping Stage-1 fixed. This response-only table reports success rates (%) for all three seeds and mean ± sample standard deviation.

ConfigurationLIBERO (three seeds)LIBERO (mean ± SD)RoboTwin 2.0 (three seeds)RoboTwin 2.0 (mean ± SD)
Correction98.7 / 98.4 / 98.698.57 ± 0.1592.8 / 92.4 / 92.892.67 ± 0.23
Selective98.9 / 99.2 / 99.099.03 ± 0.1593.3 / 93.8 / 93.993.67 ± 0.32

Across all three training seeds, Selective consistently improves over Correction on both benchmarks, with mean gains of 0.47 and 1.00 percentage points on LIBERO and RoboTwin 2.0, respectively.

As an offline proxy, action-error utility can differ from episode success. Place Can Basket decreases by 3 points after gating (14- versus 17-point deficits; Appendix C.3, Table 10). Chunk-averaged MSE may not fully capture final placement, potentially explaining this gap. Supervising the gate with episode-level rollout outcomes remains future work (Sec. 6).

Question 1: Gate-Only Accuracy and Compute

Using the same gate architecture, we train on the frozen shared WAM with the adapter omitted from label generation, training, and inference (Appendix C.8). Combining Table 15 with the Correction deficits in Table 10 gives success rates (%):

TaskFuture-freeUncorrected future-conditionedGate-onlyAdapter-only (Correction)
Press Stapler95.541.587.089.5
Place Can Basket76.542.554.062.5

Adapter-only enables future generation and disables gating. Relative to uncorrected future conditioning, it recovers 48.0/20.0 points, versus 45.5/11.5 for gate-only: adapter-only yields larger gains on both tasks.

On LIBERO/RoboTwin, adapter-only achieves 98.6%/92.7% at Neff=10/10N_{\mathrm{eff}}=10/10, versus 98.3%/92.2% at 9.21/0.83 for gate-only. The full method reaches 99.0%/93.7% at 8.70/2.45. Aggregate results are in Table 3 and Appendix C.8 (Table 14).

Question 2: Action Utility versus Rollout Benefit

Place Can Basket provides a task-level disagreement: gating increases the deficit from 14 to 17 points (Appendix C.3, Table 10). As discussed in W2 and Secs. 3.3 and 4.3, offline action-error utility enables practical supervision and yields positive aggregate gains. It measures chunk-level action error, whereas placement success depends on subsequent execution and the final criterion. The paired Place A2B Left visualization (Fig. 6; Appendix C.7) illustrates placement sensitivity despite similar TCP paths. This proxy/episode gap may explain the decrease; episode-level rollout supervision remains future work (Sec. 6).

Question 3: Stability across Adapter and Gate Retraining

With Stage-1 fixed, Selective outperforms Correction in all three retraining seeds on both benchmarks, with mean gains of 0.47/1.00 percentage points (W2).

We hope these responses clarify the method and its evaluation. We welcome further discussion.

Reviewer_dRA8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper asks when generating visual futures helps robot control and when it makes things worse. By comparing action predictions with and without future information under the same WAM, the authors find that small average changes hide large drops on some precision tasks. They train a residual adapter using ground-truth futures to correct these action changes, then use a gate to skip video generation when it is unlikely to help. Simulation and real-world results show that the method recovers much of the lost performance and offers a better trade-off between success rate and computation.

Critical strengths
  1. The paper asks a meaningful question: does a generated future actually help the robot act? From a physically consistent video generation perspective, a plausible-looking video does not necessarily capture contact, spatial relationships, or motion accurately enough for control. Evaluating its effect on actions provides a useful perspective beyond visual quality.

  2. The component analysis helps make the problem clearer. Comparing both paths under the same model, then separating the effects of correction and routing, makes it easier to see where future information hurts and which parts of the method help.

Critical weaknesses
  1. It is still unclear why future prediction causes failures. The generated video might show the wrong motion or contact, or the action model might use the video features poorly. The paper shows that future information can hurt performance, but more analysis is needed to explain the cause.

  2. The method treats an action as better when it is closer to the expert demonstration, but this does not always mean better task performance. Even with the gate, the method still performs worse than skipping future generation on the two most affected tasks. It is therefore unclear how reliably the gate identifies when future prediction helps.

Questions for authors
  1. What kinds of video errors lead to action failures? Comparing actions guided by real future video, generated video, and video with controlled errors would help explain whether the problem comes from the video or from how the action model uses it.

  2. Does the adapter make better use of the predicted future, or mostly move the action back toward what the model would do without it? Also, why does the gate not choose the better no-future option more often on the two most affected tasks?

  3. Could more varied visual examples be included, especially failure cases? The supplementary videos reviewed show only three fairly similar tabletop scenarios. Side-by-side views of the predicted video and robot execution, with and without correction, would make the results easier to understand.

2 comments

Thank you for your constructive feedback. Revisions incorporated in the manuscript are marked in red.

1. Weakness 1: Causes of future-induced failures.

Future conditioning can provide useful guidance, but its induced Δa\Delta a is not inherently beneficial. We examine how a harmful action change manifests in execution.

We added a paired Press Stapler analysis with the adapter and gate disabled, sharing the frozen WAM, initial scene, and sampling seed (Sec. 4.2; Appendix C.6, Fig. 5). At the final layer and denoising step, current visual tokens receive 71.0% attention without future conditioning and 52.2% with it; future tokens receive 25.2%. During denoising, the RMS difference between branch action predictions grows from 0.45 to 0.67 for the left arm and 0.26 to 0.74 for the right arm, in normalized action units.

The future-free branch uses the right arm, establishes contact, and succeeds; the future-conditioned branch uses the left arm and reaches the action limit without gripper–stapler contact. This case illustrates how a harmful future-induced action change can prevent the contact needed for task completion.

2. Weakness 2: Action-error proxy and gate reliability.

Our adapter learns to correct the direction and magnitude of future-induced Δa\Delta a. The paired Place A2B Left comparison (Sec. 4.3; Appendix C.7, Fig. 6) shares the initial scene and WAM, with future generation enabled and gating disabled. Only the adapter differs: the uncorrected policy places the object outside the target region and fails; the adapter places it inside and succeeds, despite similar TCP paths.

The gate reuses action-error utility, computable offline. A query's effect on episode success is difficult to observe during training because it depends on later routing and execution (Sec. 3.3).

On RoboTwin, gating raises success from 92.7% to 93.7% while reducing NeffN_{\mathrm{eff}} from 10 to 2.45 (Table 3). In rebuttal-only retraining experiments, Selective outperforms Correction in all three training seeds on both benchmarks. Its success–compute curves also outperform random routing and static truncation at intermediate budgets (Sec. 4.3, Fig. 2).

This offline proxy is distinct from episode success. Place Can Basket loses 3 points after gating (14- versus 17-point deficits; Table 10); chunk-averaged MSE may not fully reflect precise final placement, potentially explaining this decrease. Supervising the gate with episode-level rollout outcomes remains future work (Sec. 6).

3. Question 1: Video errors behind action failures.

We compared real, generated, time-reversed, and frozen futures on eight prespecified states from four demonstrations across Press Stapler and Place Can Basket, using three paired seeds. All conditions share the same frozen WAM, current inputs, and initial action noise; future conditions use the same conditioning protocol, with the adapter and gate disabled. Real futures reduce mean 32-step action MSE by 66.6% relative to matched-generated futures and 76.2% relative to the future-free branch, with lower MSE in 8/8 states for both comparisons after seed averaging. Reversing the future frames or repeating the current frame increases MSE in 8/8 states relative to real futures. These controls show that accurate future content can provide useful action guidance, while temporal perturbations can degrade action prediction. Appendix C.5 (Fig. 4, Table 13) reports these controls; Appendix C.6 (Fig. 5) gives complementary contact and execution analysis.

4. Question 2: Adapter behavior and gate choices on the two severe tasks.

The adapter learns a residual to correct the direction and magnitude of Δa\Delta a. Its alignment target follows beneficial ground-truth-future velocity changes; otherwise, it anchors the corrected velocity to the future-free reference (Eqs. 7–8). This conditional target encourages useful future guidance and constrains harmful shifts.

Correction achieves 98.2% on LIBERO-Long versus 95.4% for future-free execution (Table 2). In the paired Place A2B Left comparison (Fig. 6), adding the adapter changes failure to success with future generation enabled and gating disabled.

The gate predicts query-level action-error utility, not episode success, so it can invoke futures on these tasks despite higher future-free success. It improves the overall success--compute trade-off (W2), while chunk MSE may miss precise final placement in Place Can Basket.

5. Question 3: Diverse visualizations and failure cases.

Future-conditioned Press Stapler branches, with and without the adapter, succeed in Fig. 7 and fail in Fig. 8, where future-free execution succeeds. The panels compare predicted futures and execution (Appendix C.9). Rollouts share initial scenes but later diverge. The adapter turns placement failure into success on Place A2B Left (Fig. 6; Appendix C.7).

We hope this addresses your concerns. We welcome further discussion.

Thank you for the rebuttal. I will keep the weak accept score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank the Area Chair for their time and consideration, and all reviewers for their careful evaluation and constructive feedback. Their comments helped us clarify Utility-WAM’s central contribution, strengthen the analysis of future-induced failures, and expand the evaluation of correction and selective deployment.

Summary of revisions

1. Failure causes and the adapter’s role

(Reviewer dRA8: W1, W2, Q1–Q3)

(Sections 3.2, 4.2, and 4.3; Appendices C.5–C.7 and C.9) We have clarified that future conditioning can provide useful guidance, but its induced Δa\Delta a is not inherently beneficial. The adapter learns to correct its direction and magnitude through training targets that encourage useful future guidance and constrain harmful action shifts.

We added offline comparisons of real, generated, time-reversed, and frozen futures (Fig. 4, Table 13), showing lower action error with real futures and higher error after temporal perturbations. The paired Press Stapler analysis illustrates failure to establish effective contact (Fig. 5). The paired Place A2B Left comparison shows that adding the adapter changes placement failure to success with future generation enabled and gating disabled (Fig. 6). We also added predicted-future/execution comparisons with and without the adapter, including success and failure cases (Figs. 7–8).

2. Reliability of the proxy gate

(Reviewer dRA8: W2, Q2; Reviewer qSHk: W1, W2, Q1, Q2; Reviewer ADhQ: W2, Q2, Q3)

(Sections 3.3 and 4.3; Appendix B; Appendices C.3 and C.8) We clarified why action-error utility provides practical offline supervision and reported the gate’s ability to distinguish offline utility labels (Table 7). We added a learned gate-only baseline under the frozen shared WAM (Table 3, Fig. 2, Tables 14–15). At the reported operating points, adapter-only achieves higher success than gate-only on both benchmarks, and the full framework further improves success while reducing computation relative to adapter-only.

On RoboTwin, gating raises success from 92.7% to 93.7% while reducing effective video generation steps from 10 to 2.45. The intermediate-budget curves also improve over random routing and static truncation. We clarified the gap between action-error utility and episode success, including the Place Can Basket decrease, in Appendix C.3 and the Limitations discussion in Section 6.

3. Robustness across training seeds

(Reviewer qSHk: W2, Q3; Reviewer ADhQ: W1, Q1)

(Additional results reported in the individual responses) We evaluated an additional independently trained Stage-1 checkpoint on RoboTwin. Press Stapler and Place Can Basket remain the two most harmed tasks in the same order. Separately, keeping Stage-1 fixed, we retrained the adapter and gate with two additional seeds. Selective outperforms Correction in all three training seeds on both benchmarks, with mean gains of 0.47 and 1.00 percentage points on LIBERO and RoboTwin, respectively.

Changes incorporated in the revised manuscript are marked in red. We believe these revisions and additional results address the reviewers’ main concerns. We appreciate Reviewer ADhQ’s confirmation that their concerns have been resolved. We also thank Reviewer dRA8 for their follow-up confirming that they will maintain their weak accept score. We welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.