acceptodds
Under review as a conference paper at ICLR 2027

DualFuture: Event-Triggered Dual-Future Residual Correction for Vision-Language-Action Policies

Abstract

Action-chunked vision-language-action (VLA) policies execute a predicted chunk before observing its consequences, so a small prediction error at a critical action event goes uncorrected, accumulates over the executed commands, and fails the episode. We introduce Dual-Future Residual Correction, an event-triggered framework that wraps a pretrained VLA without fine-tuning its weights. At a detected event, it predicts two futures—one consistent with task success, one induced by the proposed action—and decodes their discrepancy into a bounded residual added to the action before execution. Both futures and the residual are learned from realized rollout futures and known action differences, so no manual correction labels are needed. Instantiating critical events as persistent gripper transitions, we attach DualFuture to the baseline VLA policy on RoboCasa365 and report three further frozen policies to situate its base success. Success rises from 73.8% to 78.7% on atomic tasks and from 45.4% to 50.9% on unseen composite tasks. On five tabletop tasks with a real Franka robot, DualFuture raises the frozen policy's pooled success from 40.0% to 52.0% under randomized object placements.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLAs, world models, correction

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_E3b7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a DualFuture approach for action chunk correction. Without fine-tuning the model weights, the proposed method corrects the predicted actions from a pretrained VLA model. The approach contains three stages. An even-trigger first detects the critical events from an imminent gripper transition. Then, the proposed DualFuture applies two types of future prediction, including a task-consistent future and the action-induced future from the predicted action chunk. Once these two types of future are obtained, the residual derived from these two futures is used to correct the original action chunk. The experiments are conducted on RoboCasa365 and the real-world Franka robot.

Critical strengths
  1. Action correction without fine-tuning the VLAs are practical for real-world deployment. It preserves the knowledge learned from large-scale pretraining while providing an efficient way to refine the proposed action chunk before execution.
  2. The proposed method is reasonable and technically sound. It leverages the latent future prediction as signals to guide the refinement of the action chunk.
  3. The overall paper is well structured.
Critical weaknesses
  1. The proposed DualFuture adds additional modules upon a frozen VLA model, which could increase inference costs. As real-time execution is critical for robot manipulation tasks, it would be more comprehensive that this paper could also provide the experiment to analysis and discussion about the trade-off between performance and the induced additional inference costs.
  2. For the future prediction, this paper applies V-JEPA-2 as the prediction target. The motivation behind the use of feature-space future prediction instead of in the pixel space are not clearly highlighted. In addition, for feature-level prediction, previous works also adopt other foundation models like DINO features, etc as the prediction target. Providing the clarification or comparisons among different future prediction targets would be beneficial.
  3. The Figure 2 is not easy to read due to the orientation of the text in the figure.
Questions for authors
  1. Is the proposed method model-agnostic? The current method is based on pi0.5. Whether the proposed approach can be applied to different foundation VLA models like GR00T N1.5 or RLDX?
3 comments
AuthorsOpenReview ↗

Response to Reviewer E3b7 - Part 1

[Weakness 1] Inference Costs

On RoboCasa365 Composite Unseen, DualFuture improves the success rate of the frozen π0.5 policy from 45.4% to 50.9%, a gain of 5.5 percentage points. On an RTX 5090, DualFuture evaluates its trigger once every 16 observation steps. Each evaluation takes 10.27 ms, equivalent to an amortized monitoring cost of 0.64 ms per observation step. When the trigger fires, future prediction and residual correction take a further 26.69 ms, bringing the total additional latency for that cycle to 36.96 ms. The average intervention rate was 20.96%.

The table compares DualFuture with two baselines under the same trigger schedule, π0.5 backbone, training data, and evaluation protocol. All latency figures represent overhead beyond ordinary base-policy planning. A single π0.5 planning call takes 77.88 ms on the same hardware.

MethodComposite Unseen success rateAdded latency per 16-step cycle (trigger inactive)Added latency per 16-step cycle (trigger active)
Frozen π0.545.4%0 ms—
Triggered Replan47.9%10.27 ms88.15 ms
Sample-and-Rerank (4 candidates)49.5%10.27 ms345.31 ms
DualFuture50.9%10.27 ms36.96 ms

All three correction methods incur the same monitoring cost. Once the trigger fires, Triggered Replan makes an additional π0.5 planning call, Sample-and-Rerank evaluates four candidate action chunks, and DualFuture directly corrects the existing chunk. Among these methods, DualFuture achieves the highest task success rate and incurs the lowest additional latency on triggered cycles.

We also benchmarked the open-source VLA-Corrector on the same RTX 5090. It achieves 46.25% task success and runs detection at every observation step, taking 38.81 ms per step, compared with DualFuture’s amortized 0.64 ms per step. After triggering, VLA-Corrector’s gradient-guided recovery takes 145.75 ms, compared with 26.69 ms for DualFuture’s future prediction and residual correction.

For context, FPC-VLA reports a correction latency of 1.766 s per keyframe using Qwen2.5-VL-7B as its supervisor. VLA-in-the-Loop reports 21 s per correction using a 7B discriminator and a 14B video generation model. These figures come from the respective papers.

[Weakness 2] Choice of Latent Representation

DualFuture predicts both the expected outcome of the current action chunk and the task-desired state in a shared feature space, enabling the corrector to directly correct actions based on the difference between the two, without reconstructing future images. We prioritized V-JEPA 2 because its features have been validated in real-robot planning: V-JEPA 2-AC predicts the consequences of candidate actions in this feature space and uses them to perform pick-and-place tasks on a Franka robot arm. This aligns directly with DualFuture's need to correct actions based on predicted consequences.

We further replace the prediction target with DINOv3 ViT-L/16 features while keeping the same training and evaluation setup. Surprisingly, we observe that although DINOv3 ViT-L/16 has only about 30% of the parameters of V-JEPA 2 ViT-g/16 (300M vs. ~1,000M), DualFuture with DINOv3 features achieves a higher success rate on RoboCasa 365 Composite Unseen (51.75% vs. 50.9%). Both feature targets improve on the frozen π0.5 baseline (45.4%), with gains of 6.35 percentage points for DINOv3 and 5.5 points for V-JEPA 2. These results show that pixel space DINOv3 features can also serve as an effective prediction target for action correction in DualFuture.

[Weakness 3] Size of Figure 2 Text

Thank you for pointing this out. We will follow your advices to revise Figure 2 so that all labels can be read horizontally without rotating the page .

AuthorsOpenReview ↗

Response to Reviewer E3b7 - Part 2

[Question 1] Transfer Across VLA Backbones

DualFuture is designed to work with different VLA backbones. It takes the action chunk proposed by a frozen policy and predicts a residual correction. To test whether a corrector learned on one backbone remains effective on another, we evaluated DualFuture with GR00T N1.5 on RoboCasa365 Composite Unseen under the same protocol used for π0.5.

Applying the π0.5-trained DualFuture to GR00T N1.5 without retraining raised task success from 44.0% to 48.875%, a gain of 4.875 percentage points. When we trained DualFuture specifically for GR00T N1.5 while keeping the backbone frozen, success reached 49.38%, a gain of 5.38 points. This is close to the 5.5-point gain observed on π0.5.

Direct transfer thus retained approximately 91% of the gain achieved with backbone-specific training. These results demonstrate transfer from π0.5 to GR00T N1.5.

Thanks to the authors for the response. My concerns have been addressed. I will retain my positive score.

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

This paper proposes DualFuture, an event-triggered action correction framework for VLA policies. The method predicts two future representations: a Goal Future representing the desired task outcome and an Action Future representing the expected consequence of the VLA's proposed actions. Their discrepancy is used to generate action residuals that correct the original action chunk before critical gripper events. The framework is trained using successful trajectories and alternative action samples without manual correction labels or fine-tuning the underlying VLA. Experiments on RoboCasa365 and a real Franka robot demonstrate improvements in task success rates.

Critical strengths
  1. The paper proposes an action correction method based on dual-future prediction. By predicting the Goal Future and Action Future, the method uses their discrepancy to generate action residuals and correct VLA actions before critical actions are executed.
  2. The method uses actions from successful trajectories as anchors, samples alternative actions from the same state, and uses the action differences as supervision to train the Corrector, without requiring manually annotated correction labels.
Critical weaknesses
  • DualFuture is only attached to π0.5. Results for other frozen policies do not show that the correction method transfers across backbones.
  • On composite tasks, the method rescues 19.1% of all evaluated episodes but also turns base-policy successes into failures in 13.6%. The improvement is useful, but preserving successful behavior remains a concern.
  • The paper does not clearly report correction latency or end-to-end overhead on the real robot.
Questions for authors
  1. Authors emphasize that the event-triggered mechanism can avoid unnecessary computation, but the paper does not explicitly report the inference latency of each correction or the end-to-end execution overhead on real robots.
  2. What happens if the Goal Future prediction is incorrect? If the model predicts an unreasonable Goal Future, could the Corrector modify an originally correct action and cause it to fail?
  3. When the Goal Future and Action Future are inconsistent, is it possible that the Action Future represents another successful trajectory with a slightly different position or approach (e.g., grasping an object from either the left or the right)? How does the system avoid unnecessary corrections in such cases? Is there a mechanism to determine whether the original action is already valid?
5 comments
AuthorsOpenReview ↗

Response to Reviewer faBi - Part 1

[Weakness 1] Transfer to other VLA backbones

To examine whether DualFuture can transfer to other VLA backbones, we compare two settings on RoboCasa 365 Composite Unseen with a frozen GR00T N1.5 backbone under the same evaluation protocol:

A. Direct transfer without retraining. We directly apply DualFuture trained on π0.5 to GR00T N1.5 at inference time. The task success rate improves from the frozen baseline of 44.0% to 48.875%, an absolute improvement of 4.875 percentage points.

B. Target-backbone training. Following the training setup of the π0.5 experiments, we train DualFuture for the frozen GR00T N1.5 backbone. The task success rate reaches 49.38%, an improvement of 5.38 percentage points over the same baseline, which is close to the gain on π0.5 (5.5 percentage points).

Direct transfer captures approximately 91% (4.875 / 5.38) of the gain from target-backbone training. This result indicates that the correction mechanism learned on π0.5 can directly improve the performance of GR00T N1.5 at inference time, while target-backbone training brings further improvements.

[Weakness 2] Corrections that turn successes into failures

DualFuture incorporates two mechanisms to preserve actions that would succeed under the base policy. First, DualFuture invokes the Corrector only when the Trigger predicts that the gripper is about to open or close. At all other times, it directly executes the actions proposed by the base policy. Second, even when the Trigger fires, the Corrector can preserve the original action by outputting a zero residual. During training, the success no-op term encourages the Corrector to output zero residuals on successful examples, and the residual magnitude regularizer further limits the size of the correction (Eq. 15).

To help DualFuture preserve different actions that can accomplish the same task, we incorporated the rollout outcome of each alternative action chunk into residual supervision. Specifically, we sampled alternative action chunks from the same state and evaluated whether each rollout completed the current task. For successful alternatives, we set the target residual to zero even when the approach direction or grasp location differed from that of the reference action. For unsuccessful alternatives, we retained the residual target in Eq. (19), computed as the successful reference action minus the alternative action. This change reduced the number of episodes in which DualFuture turned a base-policy success into a failure from 109 to 97, an 11.0% relative reduction.

In a manual review of all 109 original harm episodes on RoboCasa365 Composite Unseen, 23 (21.1%) appeared to involve unnecessary interventions, while the remaining 86 (78.9%) involved inappropriate corrections after the Trigger fired. These observations suggest that more selective triggering and more accurate residual corrections could further reduce harm.

[Weakness 3 & Question 1] Correction Latency

We measure the inference latency of a single planning step on a workstation equipped with an NVIDIA RTX 5090 GPU. π0.5 generates an action chunk in 77.88 ms on average. Attaching DualFuture adds 10.27 ms for the trigger, yielding 88.15 ms when untriggered; when triggered, correction adds another 26.69 ms (25.35 ms for the prediction stage, including 11.56 ms for the two future latents, and 1.34 ms for the corrector), totaling 114.84 ms. We also measure the baselines. Triggered Replan requires an additional VLA call, resulting in 88.15 ms of additional overhead, while Sample-and-Rerank with four candidates has an additional overhead of 345.31 ms. And VLA-Corrector performs detection at every observation step, taking 38.81 ms per detection and a 145.75 ms gradient-guided recovery each time it is triggered. FPC-VLA reports a correction latency of 1.766 s per keyframe. VLA-in-the-Loop has a reported correction latency of 21 s.

AuthorsOpenReview ↗

Response to Reviewer faBi - Part 2

[Question 2] Robustness of the Corrector to Goal Future errors

To test how robust the Corrector is to an imperfect Goal Future, on RoboCasa 365 Composite Unseen, we fix the model parameters and all other inputs, and inject Gaussian noise only into the predicted Goal Future at inference time. Without additional perturbation, the task success rate is 50.9%; When the noise standard deviation is set to 10%, 20%, and 50% of the feature-channel standard deviation of each patch, the success rates are 51.00%, 50.63%, and 50.13%, respectively. Even under the strongest perturbation, where the average cosine similarity between the representations before and after perturbation drops to 0.8945, the overall success rate decreases by only 0.77 percentage points. These results indicate that the overall performance of DualFuture is reasonably robust to random representation perturbations within the tested range.

[Question 3] Avoiding unnecessary corrections to valid alternative actions

To avoid unnecessary corrections, our method first restricts when intervention occurs via the trigger: the corrector is invoked only when the system predicts that the gripper is about to open or close (Eq. 3). During the approach phase before triggering, the robot directly executes the actions proposed by the base policy. Even when the trigger is activated, the corrector can still output a near-zero residual. During training, the success no-op loss encourages the model to preserve the original actions on successful samples, while the residual magnitude regularization term further constrains the correction magnitude (Eq. 15). Moreover, the system operates in a short-horizon closed loop. At the standard 20 Hz control frequency of RoboCasa 365, inference compares local states 8 control steps (i.e., 0.4 s) ahead. Correction applies only to the current action chunk; after execution, the system replans based on new observations, so a single Goal Future prediction does not fix the entire approach path. These design choices help reduce unnecessary interventions.

We tested a variant that uses each alternative action chunk’s rollout outcome to set its residual target during training. From the same state as the successful anchor action at+a_t^+, we sampled an alternative action chunk atka_t^k and evaluated whether its rollout completed the current subtask. If the alternative succeeded, we kept each branch’s own future representation and set the target residual to zero. If it failed, we used at+−atka_t^+ - a_t^k as the target residual, following Eq. (19). For example, if the anchor grasps an object from the left and the alternative grasps it from the right, and both succeed, the right-side grasp receives a zero-residual target even though the two futures differ. These rollout outcomes provide supervision during training. At inference, the Corrector predicts the residual from the learned mapping. On RoboCasa365 Composite Unseen, this change reduced the number of episodes in which DualFuture turned a base-policy success into a failure from 109 to 97, an 11.0% relative reduction.

Authors only provided experiments on one new VLA backbone, i.e. GR00T N1.5, which cannot fully address my concerns. Hence, I keep my score.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5djf - Part 1

[Weakness 1 & Question 1] Necessity of the Dual-feature Design

All models in Table 3 were independently retrained with their respective input configurations. The variant trained without the Induced Future (the Action Future in the reviewer’s terminology) corresponds to the suggested Goal-Only Residual model. Its corrector receives the current state, the Goal Future, and the original action. On RoboCasa365 Composite-Unseen, this model achieves 45.4% success, compared with 50.9% for full DualFuture. Thus, even when the original action is available to the corrector, explicitly providing its predicted consequences improves closed-loop task success by 5.5 percentage points on this split. We also independently trained the suggested Direct Residual baseline using the same training data, residual supervision, and event trigger as DualFuture. It predicts residuals directly from the current visual representation, proprioceptive state, and original action, achieving 45.6% success. The variant in Table 3 that retains the Induced Future but omits the Goal Future achieves 45.5%. Both single-future models therefore perform similarly to Direct Residual, while conditioning the corrector on both futures yields the highest success rate among the four configurations.

To avoid unnecessary corrections, our method uses an event trigger to limit when the corrector is invoked. The trigger activates only when the system predicts that the gripper is about to open or close (Eq. 3). During the approach phase before triggering, the robot directly executes the actions proposed by the base policy. Even when the trigger activates, the corrector can still output a near-zero residual. During training, the success no-op loss encourages the corrector to preserve the original actions on successful samples, while the residual magnitude regularizer discourages large corrections (Eq. 15).The system also operates in a short-horizon closed loop. At the standard 20 Hz control frequency of RoboCasa365, the corrector compares the local states predicted by the Goal Future and the Induced Future eight control steps (0.4 s) ahead. Each correction applies only to the current action chunk. After executing the chunk, the system replans from new observations, so a single Goal Future prediction does not determine the entire approach path. Together, these design choices help reduce unnecessary interventions.

To examine how residual supervision affects already successful actions, we trained a variant that uses rollout outcomes to assign residual targets. Starting from the same state as a successful reference action chunk at+a_t^+, we sampled an alternative chunk atka_t^k and evaluated whether its rollout completed the current task. If it succeeded, we retained the distinct futures of the reference and alternative branches and assigned the alternative a zero-residual target. For example, if the reference grasps an object from the left and the alternative successfully grasps it from the right, the right-side grasp receives a zero-residual target even though the two futures differ. If the alternative failed, we used at+−atka_t^+ - a_t^k as its target residual, following Eq. (19). Rollout outcomes provide these labels during training. At inference, the corrector predicts the residual from its inputs. On RoboCasa365 Composite-Unseen, the number of episodes in which the base policy succeeded but the corrected policy failed fell from 109 under the original supervision to 97 under outcome-based supervision, an 11.0% relative reduction.

AuthorsOpenReview ↗

Response to Reviewer 5djf - Part 2

[Weakness 2 & Question 4] Comparison with Correction Methods

In terms of efficiency, existing correction methods incur substantial inference overhead. FPC-VLA uses Qwen2.5-VL-7B as its supervisor and reports a correction latency of 1.766 s per keyframe. VLA-in-the-Loop uses both a 7B MLLM discriminator and a 14B video generation model, with a reported correction latency of 21 s. Such latencies are difficult to accommodate in real-time robot deployment.

We benchmarked VLA-Corrector and DualFuture on the same RTX 5090, measuring the monitoring and post-trigger correction stages separately. VLA-Corrector pvideoerforms detection at every observation step, taking 38.81 ms per detection. Once triggered, a gradient-guided recovery call takes 145.75 ms. In contrast, DualFuture evaluates its trigger once every 16 observation steps, taking 10.27 ms per evaluation, or 0.64 ms per observation step when amortized. Its average monitoring cost is therefore approximately 1/60 that of VLA-Corrector. Once triggered, DualFuture directly corrects the existing action chunk, with a combined latency of 26.69 ms for future prediction and residual correction.

We also compared the additional computational overhead of DualFuture with that of our two baselines. A single planning call to the base π0.5 policy takes 77.88 ms. Considering a cycle of 16 observation steps, all three methods incur only 10.27 ms of additional trigger-evaluation overhead when no intervention is triggered. When an intervention is triggered, DualFuture incurs a total additional overhead of 36.96 ms. Triggered Replan requires an additional VLA call, resulting in 88.15 ms of additional overhead, while Sample-and-Rerank with four candidates has an additional overhead of 345.31 ms. DualFuture’s average intervention rate is 20.96%.

For task performance, we compared the methods on RoboCasa365 Composite-Unseen using the same π0.5 backbone, training data, and evaluation protocol. The base π0.5 policy achieves 45.4% success. The open-source VLA-Corrector improves this to 46.25%, a gain of 0.85 percentage points. Triggered Replan and Sample-and-Rerank achieve 47.9% and 49.5%, respectively. DualFuture achieves 50.9%, outperforming the base policy by 5.5 percentage points and VLA-Corrector by 4.65 percentage points. These results show that DualFuture achieves higher success in this evaluation while incurring lower monitoring and correction costs.

[Question 2] Residual supervision and Harm Analysis.

We explored how to refine residual supervision so that the Corrector preserves alternative actions that already succeed. The current approach uses a success no-op loss to encourage the Corrector to output zero residuals on successful samples, together with a residual-magnitude regularizer that constrains the size of corrections (Eq. 15). We tested a variant that sets the target residual for each alternative action according to its rollout outcome. If the alternative completed the current task, we assigned it a zero-residual target, even if its approach direction or grasp position differed from the reference. If it failed, we used the difference between the successful reference action and the alternative action as the target residual, at+−atka_t^+ - a_t^k, following Eq. (19). On RoboCasa365 Composite-Unseen, this change reduced the number of episodes in which DualFuture turned a base-policy success into a failure from 109 to 97, an 11.0% relative reduction.

Following your suggestion, we systematically analyzed all 109 harmful episodes under the original supervision on RoboCasa365 Composite-Unseen. Because ground-truth labels for the causes of harm were unavailable, we manually reviewed the video of each episode. The tight coupling between the predicted Goal Future and Action Future representations and the Corrector’s residual prediction makes it difficult to attribute errors to a specific component from video alone. We therefore examined whether each intervention was necessary and whether the resulting correction was appropriate. In 23 episodes (21.1%), an unnecessary intervention disrupted a viable action. For example, an additional end-effector adjustment moved the gripper away from its intended grasp position while it was already approaching the object along a feasible path. In the remaining 86 episodes (78.9%), the post-trigger correction was inappropriate. In one case, adjusting displacement along one Cartesian axis without corresponding adjustments along the other axes moved the gripper away from the target contact point. In another, the forward correction was too small and the gripper closed at its originally scheduled time, before reaching the object. These findings suggest that DualFuture could achieve higher task success by improving the trigger’s ability to recognize when intervention is needed and the Corrector’s ability to produce residuals of appropriate magnitude that are coordinated across action dimensions.

AuthorsOpenReview ↗

Response to Reviewer 5djf - Part 3

[Question 3] Future Prediction Quality and Robustness.

To assess how the corrector is affected by the shift from realized futures during training to predicted futures at inference, we conducted two experiments on RoboCasa365 Composite Unseen. We first measured how closely the predicted Goal Future and Action Future matched their realized counterparts. We then tested how task success changed when we perturbed the predicted Goal Future supplied to the corrector.

Averaged across the three camera views, the mean absolute error (MAE) between predicted and realized representations was 0.37043 for the Goal Future and 0.37707 for the Action Future. As a reference, we replaced each prediction with a random Gaussian representation processed with the same LayerNorm. The corresponding MAEs were 1.02993 and 1.03072. Relative to this baseline, the predictor reduced MAE by approximately 64% for the Goal Future and 63% for the Action Future.

To isolate sensitivity to Goal Future errors, we added Gaussian noise to the predicted Goal Future at inference while keeping the predicted Action Future unchanged. For each patch feature vector GG, we supplied the corrector with

\qquad \epsilon \sim \mathcal{N}(0,I), \qquad \alpha \in \{0.1, 0.2, 0.5\}.$$ Across 1,454 sampled noise-injection calls, the mean cosine similarities between the original and perturbed features after LayerNorm were 0.9950, 0.9806, and 0.8945 at the three noise levels, respectively. The corresponding task success rates were 51.00%, 50.63%, and 50.13%, compared with 50.90% without added noise. Even at the highest tested noise level, the observed success rate dropped by only 0.77 percentage points. This suggests that correction performance is relatively insensitive to Gaussian perturbations of the Goal Future over the tested range. **[Question 5] Generalization across VLA Backbones** To assess whether DualFuture generalizes beyond π0.5 and whether its two modules require retraining for a new policy, we evaluated DualFuture with a frozen GR00T N1.5 backbone on RoboCasa 365 Composite Unseen. We first applied the future predictor and corrector trained on π0.5 directly to GR00T N1.5 at inference time. Without retraining either module, task success increased from 44.00% to 48.88%, a gain of 4.88 percentage points. We then trained DualFuture specifically for the frozen GR00T N1.5 policy using the same training setup as for π0.5. Under the same evaluation protocol, task success reached 49.38%, a gain of 5.38 percentage points over the frozen baseline, comparable to the 5.50-point gain on π0.5. Direct transfer thus recovered approximately 91% of the gain from policy-specific training. These results show that both modules can be reused together on GR00T N1.5, with policy-specific training yielding a further improvement of about 0.5 percentage points.

I sincerely thank the authors for their thoughtful and detailed responses, which have helped me better understand the paper and its contributions. I appreciate the authors' efforts in addressing my concerns and conducting additional experiments. I now find the experimental evaluation to be comprehensive, the motivation to be well-justified, and the proposed method to be well-aligned with its motivation. While I acknowledge the merits of this work, I feel that the methodological novelty may not be sufficiently substantial to warrant a score of 4. Nevertheless, considering the authors' clarifications and additional results, I am happy to raise my score from 2 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.