acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Replanning for Vision-Language-Action Policies via Overlap Disagreement

Abstract

Modern vision-language-action (VLA) robot policies commonly use action chunking, where the VLA predicts and executes a sequence of future actions before performing inference again. While action chunking improves computational efficiency and temporal coherence, it introduces a fundamental tradeoff: frequent replanning reduces open-loop execution drift but increases inconsistency between consecutive plans, whereas infrequent replanning improves temporal consistency at the cost of accumulating stale execution error. In this work, we model this tradeoff using a nonparametric class of surrogate costs and show that it induces an optimal replanning horizon. We then show empirically that weighted overlap disagreement, an observable quantity that measures the disagreement between overlapping segments of consecutive action chunks, has a minimizer that serves as a reliable predictor of this optimal replanning horizon. Building on this insight, we propose disagreement-guided replanning (DGR), an online adaptive replanning strategy with negligible computational cost that dynamically adjusts the replanning horizon during inference without additional training. DGR is model-agnostic, easy to deploy, requires minimal tuning, and uses a single set of hyperparameters across all models and benchmarks. Across multiple VLA architectures and benchmark suites, our method improves robustness and long-horizon task performance over baselines. Code will be made available upon acceptance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language-Action Models, Action Chunking, Replanning Horizon, Online Adaptation

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_9hUYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Disagreement-Guided Replanning (DGR), an online adaptive replanning strategy for VLA policies. It addresses the trade-off between frequent replanning, which can cause inconsistency between consecutive action plans, and infrequent replanning, which leads to accumulated execution errors. DGR uses the disagreement between overlapping segments of consecutive action chunks to dynamically adjust the execution horizon, without additional training or policy evaluations. Experiments across four VLA architectures and multiple benchmarks show improved long-horizon task performance over existing baselines, with additional validation on RoboCasa and a real-world Push-T task.

Critical strengths
  1. DGR utilizes the differences in overlapping regions between consecutive action predictions to dynamically adjust the step size of action execution. Without the need for additional training or increasing the number of model inferences, it can achieve a balance between execution efficiency and re-planning frequency.
  2. The experimental paper has verified this method on 4 VLA architectures, multiple simulation benchmarks, visual perturbation scenarios, and the real robot Push-T task. The results show that this approach is particularly effective in long-term tasks.
Critical weaknesses
  1. DGR uses the differences between overlapping action segments as the basis for adjusting the step size of the re-planning process. However, the differences in actions do not necessarily indicate execution errors or uncertainties. On the contrary, if the model consistently predicts similar but incorrect actions, this method may not be able to detect the problem in a timely manner.
  2. Prior work such as AAC (https://arxiv.org/pdf/2604.04161) adapts action execution horizons based on action entropy, while MoH (https://arxiv.org/pdf/2511.19433) uses cross-horizon consensus to dynamically select executable action prefixes. Although DGR employs overlap disagreement and an EMA-based adjustment rule, the paper could more clearly establish its unique advantages over these existing approaches and demonstrate that its design offers distinct benefits beyond alternative signals for adaptive execution.
Questions for authors
  1. The paper assumes that the degree of action inconsistency brought about by re-planning is fixed and does not change with each execution of a few steps. Moreover, the longer the consecutive execution lasts, the greater the cumulative execution error will be. However, this may not be the case in reality. For instance, when a robot grabs an object, errors may be particularly prone to occur at the moment of initial contact with the object, and the error does not necessarily increase uniformly with the number of execution steps; re-planning may also cause different degrees of action changes due to different task stages. Are these two assumptions still valid in different task stages, especially in operations involving grasping, contact, and object movement?
  2. Sometimes, the difference in predictions suddenly becomes larger. This might just be a random fluctuation of the model and does not necessarily mean that more frequent re-planning is truly necessary. If DGR subsequently shortens the execution step size, it might cause unnecessary re-planning. How can this be avoided?
3 comments

Thank you for the careful review.

W1. Consecutive predictions can agree while both are wrong. We agree and we acknowledged this limitation in the manuscript (see Limitations). DGR aims to improve the execution of an already capable policy through better replanning. If the underlying policy consistently predicts incorrect actions, the policy itself needs improvement, for example through retraining; horizon adaptation alone cannot generally correct a consistently incorrect policy.

Nevertheless, our results in Appendix F show that when the input (and hence the policy output) is perturbed using an adversarial patch, DGR remains competitive and provides substantial gains on LIBERO-Spatial: 74% success, compared with 54% for the clean static optimal horizon and 69% for the attack-optimized static optimal horizon. This shows that DGR can compensate for occasional policy errors.

W2. Novelty. Both AAC and MoH involve generating multiple action chunks upon inference time. In addition, MoH involves retraining and architectural changes of the VLA policy prior to inference time. AAC estimates entropy from multiple candidate chunks generated based on the same observation. It would thus not work if chunks are generated using a deterministic regression policy that produces identical chunks under repeated queries. On the contrary, DGR compares chunks generated at subsequent observations, which can disagree even when action chunks are generated deterministically. Our OpenVLA-OFT results demonstrate applicability to this setting. Table 3 additionally shows approximately two to three times fewer total FLOPs than cached AAC on LIBERO-Long. This is particularly relevant for robotics, since onboard compute capacity can be limited.

MoH involves retraining a policy with multiple prediction horizons as well as weights to be used for fusing multiple action chunks. Upon inference time, it then selects an executable prefix, using consensus among horizon-specific predictions generated using the same observation, similar in spirit to AAC. While batching can reduce latency, actual inference time remains heavily dependent on the available hardware, and the overall compute, memory, and hardware requirements are higher than for DGR.

DGR is fundamentally different from these approaches in two ways. First, it relies on the comparison of two consecutively generated chunks. Thus, it can be used with any existing action-chunking policy without additional training, architectural modifications, or multiple evaluations. Thus, DGR preserves the base policy’s per-call inference cost, adding only negligible computation and memory overhead for the controller. Second, DGR's design is motivated by a surrogate analysis and the fixed-horizon problem studied offline. We empirically investigate weighted overlap disagreement as a bridge to task performance, then extend the same objective to variable horizons to derive the greedy online controller. This connection strengthens the justification for DGR’s controller: its online update is derived from an explicit objective, connected to the fixed-horizon analysis, and supported by empirical evidence. To our knowledge, this connection is also novel.

Q1. The static surrogate uses average costs across rollout states. We agree that drift and replanning inconsistency can vary across grasping, contact, and object movement. Our static surrogate uses state-averaged costs. Here, jj denotes the number of steps since replanning, rather than absolute task time tt. Assumption 1 concerns the expected drift profile ϕ(j)\phi(j): it permits nonlinear growth and sharp increases, and does not require error to increase uniformly along every trajectory. Similarly, σ\sigma represents an expected per-replan cost; Assumption 2 takes this average as independent of rr, rather than requiring every replanning event to introduce identical inconsistency.

A stage-dependent interpretation is possible under a mean-field assumption that replanning encounters the same expected mixture of task stages across fixed horizons. Stage-specific costs then aggregate into the same surrogate form, and the proof applies provided the averaged profile satisfies the stated assumptions. We will clarify this interpretation and scope.

It is important to note that DGR does not estimate σ\sigma or ϕ\phi at inference time. It uses the current disagreement and its smoothed average, allowing the execution horizon to change across stages. For example, the within-episode Push-T trace in Appendix G illustrates this behavior.

Q2. The dead zone and bounded updates limit responses to fluctuations. Directional calibration is conditioned on the score leaving the dead zone, whose width is controlled by α\alpha. Additionally, each horizon change is bounded by Δ\Delta. Therefore, α\alpha controls the controller sensitivity, while Δ\Delta controls the size of each horizon change. These two parameters can be jointly tuned to address this issue.

Thank you for the detailed clarifications and additional experiments. The comparisons with AAC and MoH help better establish the methodological distinctions and computational advantages of DGR. The adversarial perturbation results also provide useful evidence of its robustness to occasional policy errors. However, some concerns remain regarding the validity of the fixed inconsistency assumption across different task stages and the sensitivity of horizon adaptation to prediction fluctuations. The proposed dead zone and bounded updates are reasonable design choices, but their effectiveness under varying execution conditions is not yet fully established. Overall, I appreciate the authors' efforts, but I maintain my original rating.

Thank you for following up on our response. Your comments helped us identify where the aggregate modeling assumptions needed clearer explanation and motivated the additional robustness experiment below.

Q2. The sensitivity of horizon adaptation to prediction fluctuations. To evaluate the sensitivity of horizon adaptation to prediction fluctuations, we performed an evaluation where we perturbed the predicted action chunks with zero-mean Gaussian noise. The noise had standard deviation λsd\lambda s_d along action coordinate dd, where sds_d is the empirical standard deviation along that coordinate measured from clean policy outputs. We can thus increase the variance of the action chunks by setting λ>0\lambda>0. As the included table shows, DGR as proposed has higher task success rate than its variants without the deadzone and with a high horizon update step, at all three noise levels. DGR is also competitive with the Static Optimal (r=20r=20). At noise level 22 all methods perform poorly due to the high noise. These results support the effectiveness of the safeguards under the tested fluctuations.

Methodλ=1\lambda=1λ=0.5\lambda=0.5λ=2\lambda=2
DGR62±5\mathbf{62}\pm587±3\mathbf{87}\pm317±417\pm4
No dead zone (α=0\alpha=0)50±550\pm578±478\pm413±313\pm3
Larger updates (Δ=10\Delta=10)30±530\pm574±474\pm49±39\pm3
Fixed r=20r=2060±560\pm579±479\pm418±4\mathbf{18}\pm4

Q1. Fixed inconsistency assumption across different task stages. We do not assume fixed inconsistency across different task stages. Assumption 2 concerns the expected value over the distribution of states in a task. Different task stages (e.g., grasping, contact, or transport) correspond to different subsets of the state space X\mathcal{X}, but our cost model does not assume independence of the costs from the state.

To make it explicit what we average over, let ptask,r(X)p_{task,r}(X) denote the distribution of states XX encountered at a uniformly sampled replanning event under fixed horizon rr. Let σ(X)\sigma(X) and ϕ(X,j)\phi(X,j) be the replanning inconsistency and the instantaneous drift cost incurred after jj steps if the system is in state XX upon replanning, respectively. A formulation, where state-dependence of the cost surrogate is explicit would be

\left[ \sigma(X)+\sum_{j=0}^{r-1}\phi(X,j) \right]dX.$$ We can extract the terms corresponding to the replanning inconsistency and the instantaneous drift from the above expression, showing their dependence on $r$, leading to $$\sigma_r=\int p_{\text{task},r}(X)\sigma(X)\,dX, \qquad \phi_r(j)=\int p_{\text{task},r}(X)\phi(X,j)\,dX.$$ Observe that Assumption 2 assumes independence of the **expected replanning inconsistency cost** ($\sigma_r$ above) from **the replanning horizon $r$**, rather than independence of $\sigma(X)$ from the state $X$. The motivation behind this assumption is given by the sources of inconsistency described on lines 189-194. The task and the choice of $r$ do influence $\sigma_r$ through $p_{\text{task},r}(X)$. For different tasks we find similar values of $\sigma_r$ for $r=1$ (lines 369-371), and considering that $p_{\text{task},r}(X)$ is strongly dependent on the task, but only weakly on $r$ (the overall task progression remains similar regardless of $r$), it is reasonable to consider this **average** to be an intrinsic property of the policy, justifying the modeling Assumption 2. Importantly, state (and thus stage)-dependent inconsistency costs are compatible with this aggregate model. We would like to emphasize that the cost model based on the two components is used in our paper for motivating horizon adaptation, as written on lines 112-113 and 162-164. The proposed controller DGR does itself not estimate $\sigma$ or $\phi$. Its effectiveness is assessed empirically across the reported policies and benchmarks. We have also extended the evaluation with controls matched to DGR’s replanning horizon distribution and its mean horizon, providing further evidence for the benefit of online adaptation. We hope these additions, together with the clarifications and robustness experiments above, address your remaining concerns underlying the rating.
Reviewer_cucROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies execution-horizon selection for action-chunked vision-language-action policies. It introduces a surrogate cost capturing a tradeoff between open-loop execution drift and replanning inconsistency, and proves that this cost is unimodal under explicit assumptions. It then empirically investigates weighted disagreement between overlapping portions of consecutive action predictions as a proxy for selecting a fixed execution horizon. Building on this signal, the authors propose Disagreement-Guided Replanning (DGR), which adjusts the execution horizon by comparing current disagreement with an exponential moving average. DGR requires neither additional training nor additional policy evaluations. Experiments cover four VLA architectures on LIBERO, with additional RoboCasa, adversarial-patch, and physical Push-T evaluations.

Critical strengths

A practical method with low integration cost. DGR uses consecutive action predictions already available during execution, avoiding additional policy samples, specialized training, and access to model-internal signals. This makes it potentially useful across different VLA architectures.

A substantial improvement in the strongest experimental setting. On π₀ LIBERO-Long, DGR achieves 89±1% success, compared with 79±2% for the suite-level static optimum and 82±2% for AAC. This is meaningful evidence that the proposed approach can improve long-horizon execution. Evaluation across four architectures provides additional evidence of applicability beyond a single policy family.

Useful reporting of limitations. The paper reports that standard DGR underperforms the static optimum on RoboCasa and explicitly acknowledges failure cases involving consistently incorrect predictions. The physical Push-T results are also appropriately qualified: the improvement over the best tested fixed horizon amounts to only one successful episode.

Critical weaknesses
  1. The value of online adaptation is not sufficiently isolated. Fixed-horizon correlations do not establish that disagreement relative to its moving average identifies beneficial horizon updates. Controls matching DGR’s average horizon, inference budget, or horizon distribution are needed.

  2. Directional calibration is insufficiently validated. The reported 90% sign agreement uses realized trajectories rather than counterfactual continuations under both candidate updates. Sign-agreement frequency also does not establish the magnitude-weighted expectation required by Assumption 3.

  3. The theory supports the surrogate rather than task performance. Proposition 1 proves unimodality under assumed cost structure. Reconstructing drift from disagreement using that structure demonstrates compatibility, rather than independently validating the mechanism.

  4. Some experimental comparisons remain inconclusive. AutoHorizon parameters are transferred from π₀.₅ to π₀, several gains are small relative to reported errors, and the RoboCasa experiments do not separate the effects of groupwise adaptation and inpainting.

Questions for authors

Does DGR retain its advantage over a fixed horizon matched to its average execution length or inference budget, and over a shuffled schedule preserving its empirical horizon distribution? These comparisons would materially affect my assessment of the contribution of online adaptation.

4 comments

Thank you for the careful review and for suggesting additional experiments. We have incorporated the new results and clarifications into the manuscript.

W1 and Q1. Motivation behind DGR and the value of online adaptation. The correlations between weighted overlap disagreement Loverlap(r)\mathcal{L}_{overlap}(r) and task failure across fixed horizons motivate the objective in §4. Eq. (6) extends that objective to variable horizons and recovers Loverlap(r)\mathcal{L}_{overlap}(r) for constant rr. We derive DGR’s greedy update from this extension under the local approximations and Assumption 3 (Appendix A.3). Nonetheless, we agree that the benefit of online adaptation may have been unclear.

We have added new experiments to validate the benefit of online adaptation. We evaluate a control sampling from DGR's task-specific horizon distribution on π0\pi_0 Spatial, Object, Goal, and Long, obtaining 94±1%94\pm1\%, 97.8±0.7%97.8\pm0.7\%, 94±1%94\pm1\%, and 83±2%83\pm2\% success, respectively. A fixed horizon matched to DGR's task-specific mean achieves 96.0±1%96.0\pm1\%, 95±1%95\pm1\%, 85±2%85\pm2\%, and 75±2%75\pm2\%, respectively. Comparing these results to those reported in Table 1 for DGR, we observe that DGR performs 6% points better than using the marginal horizon distribution on π0\pi_0-LIBERO-Long, and 10 and 14 % points better than using DGR's task-specific mean horizon on π0\pi_0-LIBERO-Goal and π0\pi_0-LIBERO-Long, respectively. The differences on the other suites are modest relative to the reported uncertainty. Note that each suite contains 10 tasks, and the gains are spread over multiple tasks. These results support the value of DGR's online adaptation: reproducing its mean horizon, or even sampling from its marginal horizon distribution, does not reproduce its LIBERO-Long performance.

At the same time, we would like to emphasize that even when DGR only matches Static Optimal, it avoids the expensive, exhaustive suite-specific horizon sweep used to tune that baseline. DGR uses the same default hyperparameter configuration across the LIBERO comparisons. Matching an optimized baseline without suite-specific horizon tuning already has merit as well as practical value, particularly when evaluation rollouts must be collected on a real robot.

W2. Evidence for Assumption 3 We thank the reviewer for the suggestion. Under the two local approximations in Appendix~A.3 (lines 657-674), we estimated the weighted directional condition using counterfactual continuations. Note that the definition of μk\mu_k depends on ρˉ\bar \rho. For fixed-horizon continuation policies r=5,10,20r=5,10,20, we sample triggered histories from π0\pi_0 LIBERO-Long. At each history, we branch the simulator with r−Δr-\Delta and r+Δr+\Delta, followed by the same continuation policy, with five repeated rollouts per branch. The resulting 95% bootstrap confidence intervals [0.44,1.76][0.44,1.76], [0.84,3.73][0.84,3.73], and [0.41,8.6][0.41,8.6], respectively, all in units of 10−410^{-4}. These results support Assumption 3 under the stated local approximations in all three tested configurations. We replaced the paragraph with the new counterfactual evaluation.

W3. The empirical bridge between surrogate and success rate Proposition 1 establishes unimodality of the surrogate C(r)C(r) and characterizes its minimizer under Assumptions 1 and 2. Proving that this minimizer necessarily maximizes task success rate would require additional assumptions linking policy errors, environment dynamics, and task outcomes. Instead of this, our paper provides an empirical bridge between the two: we show that the profile of weighted overlap disagreement is compatible with the class of cost functions, in the sense that it can be decomposed in two terms with corresponding behavior, and that its minimizer predicts the failure-rate minimizer. Our assumptions together with the empirical bridge between success rate and disagreement form an internally consistent framework that allows us to derive DGR, which performs well in practice. We do agree that the empirical evidence could be considered a compatibility test of the success rate with the surrogate, rather than an independent validation of the mechanism and the underlying assumptions.

W4. Experimental comparisons. We agree that several success-rate gains are modest compared to adaptive baselines. Observe, however, that DGR has much lower computational cost than those baselines, is independent of the VLA architecture, and is easy to deploy. On LIBERO-Long, standard DGR uses approximately one-third to one-half of cached AAC's total FLOPs. Furthermore, the OpenVLA-OFT results demonstrate DGR's applicability to a non-diffusion policy. Additionally, all results using DGR were obtained with a single default hyperparameter value. This is practically important, as there is no need for suite-specific tuning for DGR. The baselines use suite-specific parameter choices provided by the authors, where available. For AutoHorizon, we have clarified that the π0\pi_0 result is a transfer comparison using π0.5\pi_{0.5} settings.

We ran the groupwise-only ablation with inpainting guidance disabled. Across the same 20 tasks, standard DGR, groupwise DGR without inpainting, and guided groupwise DGR achieve 41%41\%, 42%42\%, and 48%48\%, respectively. Groupwise replanning alone remains close to standard DGR, while adding inpainting within the groupwise setting improves the mean by approximately 6 percentage points. This supports the importance of conditioning a group's new plan on the other groups' already committed actions. We include this ablation in the revised manuscript to make the component-wise contributions clearer.

Thanks for you responsible and detailed reply. I maintain my positive score.

Thank you for taking the time to read our response. We are glad that it addressed your concerns, and we appreciate you maintaining your positive score.

Reviewer_aEGyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Disagreement-Guided Replanning (DGR), a training-free method for adapting the execution horizon of vision-language-action (VLA) policies. DGR compares overlapping action predictions from consecutive plans and uses their disagreement to adjust when to replan. The authors motivate the approach through a cost model of execution drift and replanning inconsistency and evaluate it across multiple VLA models and manipulation tasks.

Critical strengths
  1. The paper addresses a clearly defined problem and provides an intuitive motivation for adaptive replanning.
  2. The proposed method is simple and lightweight, making it straightforward to integrate into existing VLA policies.
  3. The paper includes comparisons with fixed-horizon and adaptive baselines, as well as a discussion of limitations and failure cases.
Critical weaknesses
  1. The method is poorly presented. The paper lacks a clear overview explaining how DGR operates and how it differs from existing approaches, making it especially difficult for readers to understand the method and assess its novelty and contribution. Furthermore, the paper introduces several objectives and assumptions but without clearly explaining their relationships.
  2. The theoretical analysis leaves a key gap between horizon selection and task performance. The authors prove that their cost function has an optimal replanning horizon, but do not establish that this horizon also maximizes task success. Moreover, DGR adjusts the horizon based on action disagreement, whose connection to task success is supported empirically rather than theoretically. Thus, the proof does not establish that DGR’s adjustments improve task performance.
  3. Performance improvements are not consistent across settings. Table 1 only shows a notable gain on π0\pi0 in LIBERO-Long, but improvements elsewhere are generally modest, with some results below other methods.
Questions for authors

See Weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.