acceptodds
Withdrawn from ICLR 2027

Beyond In-Distribution Preservation: Recovering Generalization in Quantized VLAs via Vulnerability-Oriented Tuning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Post-training quantization has been shown to preserve VLA performance under standard evaluation conditions, but whether it preserves the full-precision model's robustness and generalization remains underexplored. In this study, we systematically study the robustness and generalization of post-quantized VLA policies under environmental disturbances. Empirical results show that quantized policies can become fragile to subtle environmental variations despite retaining comparable in-distribution performance. We further observe that action discrepancies are concentrated in a small subset of rollout states, while teacher guidance has opposite effects depending on discrepancy: it improves generalization at high-discrepancy states but can degrade it at low-discrepancy states. These findings reveal that effective post-quantization recovery requires selectively intervening on vulnerable states rather than globally distilling the student. We therefore propose Policy-Induced Vulnerability-Oriented Tuning (PIVOT-Q), a vulnerability-aware On-Policy Distillation (OPD) framework that selectively corrects vulnerable states encountered during quantized-student rollouts using the frozen full-precision policy as a teacher. PIVOT-Q identifies vulnerable states using discounted accumulated discrepancies over a short horizon, applies phase-balanced sparse supervision, and uses a Behavioral Anchor to prevent unnecessary changes. Experiments under seven LIBERO-Plus environmental variations demonstrate consistent recovery across multiple VLA backbones and quantization methods. Notably, PIVOT-Q consistently outperforms full-state distillation across all settings while using only 7.4% of its state-level distillation budget.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Post-Training Quantization, Post-Quantization Recovery, Policy Distillation, Vulnerability-Oriented Tuning

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 5 reviews, above 27% of the others.
Reviewer_5Pg3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the robustness of the post-training quantization of vision-language-action (VLA) mode to environmental shifts, beyond in-distribution evaluation setting. The authors provide empirical study to show that quantized policies (GR00T-N1.5, π0.5, OpenVLA-OFT with QuantVLA, HoloQ-VLA, and QVLA) lose performance under environment disturbances. Besides that, they show that the performance drop can be partially recovered by selectively replacing student actions with teacher actions at high-discrepancy states, motivating targeted supervision during post-quantization recovery. Motivated by this they propose Policy-Induced Vulnerability-Oriented Tuning (PIVOT-Q), which selectively distills the quantized student from a frozen full-precision teacher on student-visited states that shows the highest vulnerability scores based on current and discounted future action discrepancies. Besides that they include a behavior anchor which regularizes the student toward the full-precision teacher’s actions on reference states to limit forgetting during recovery. The authors validate their proposed approach on LIBERO-Plus benchmark.

Critical strengths

The strengths of the paper are summarized below:

i. The paper is well-written and easy to follow. ii. The paper tackles a relevant problem, as VLA inference is computationally demanding making quantization attractive for deployment without losing the generalizability. iii. The approach requires neither reward definition nor task-success labels. iv. The inclusion of behavior anchor based regularization is sound and can prevent training instability. v. The empirical study covers multiple VLA backbones and quantization methods, providing evidence the method works beyond a single model or configuration.

Critical weaknesses

My primary concern is the novelty of the paper. It introduces a vulnerability score to guide the policy tuning or distillation. However, the score is fairly intuitive and straightforward raising question about the novelty of the paper. My decision is primarily motivated by this concern. The weaknesses aside from the novelty are summarized below:

i. Although VLAs have been designed for general purpose policy learning, their generalization to unseen environments remains limited. This poses a limitation on the proposed approach. ii. The motivating action-replacement study does not capture the broader behavioral changes caused by the intervention which necessitates introduction of behavioral anchor loss. This also makes one wonder if the behavior regularization alone accounts for a substantial fraction of the recovery. iii. It is not clear how the behavior anchor states have to be chosen. iv. Considering the simplicity of the approach the paper will benefit if the approach can be extended to broader model compression techniques such as pruning or distillation into smaller architectures. v. The necessity of the current-discrepancy term \tilde q_t is not established. Discrepancy can be established with \tilde r_t term alone.

Questions for authors

Please find my questions below:

i. How does the method perform when the full-precision teacher also fails under environmental shifts? Additional empirical analysis would help clarify whether distillation also reinforces teacher errors in the distillation/tuning process. ii. Could you provide a future-only ablation that removes the current-discrepancy term from the vulnerability score? iii. Could you provide an anchor-only ablation to determine how much recovery is attributable to behavioral regularization alone? iv. Could you comment on whether vulnerability-oriented supervision is applicable to other model compression approaches, such as pruning or distillation into smaller architectures? v. How are the reference states in the behavioral anchor bank selected? Does constructing this bank require expert demonstrations or additional annotations, and how sensitive is recovery performance to its composition?

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

This paper studies the degradation of out-of-distribution performance in post-training-quantized vision-language-action policies. The authors observe that teacher-student action discrepancies are unevenly distributed across student rollouts and propose PIVOT-Q, a selective on-policy distillation framework. The method ranks states using temporally accumulated action discrepancies, applies sparse supervision with phase balancing, and introduces a behavioral anchoring regularizer to reduce unnecessary policy changes. Experiments on LIBERO-Plus evaluate four model-quantizer configurations and seven environmental variations. The proposed method recovers a substantial portion of quantization-induced performance degradation while using fewer supervised states than full-state distillation.

Critical strengths

1.The paper identifies an important generalization gap in quantized VLA policies that is not captured by in-distribution evaluations. 2.The proposed sparse on-policy distillation achieves substantial recovery in the GR00T-N1.5/QuantVLA setting using fewer supervised states than full-state distillation.

Critical weaknesses

1.The claimed consistent improvement over full-state distillation is not convincingly supported across all configurations, as PIVOT-Q achieves gains of only 0.1–0.4 percentage points in three of four settings, which are small relative to the reported standard deviations. 2.Although the paper analyzes discrepancy concentration and current-future correlations, it does not directly establish whether the proposed vulnerability score identifies states that causally contribute to task failure rather than merely exhibit large teacher-student action differences. 3.Although Table 3 supports the contributions of future discrepancy and behavioral anchoring, PIVOT-Q improves over uniform state selection by only 1.9 percentage points in the evaluated configuration. Whether this incremental gain remains significant across other model-quantizer settings is unclear. 4.The claimed supervision efficiency does not establish an equivalent computational advantage, since the reported 7.4% budget excludes rollout collection, full-precision teacher inference, and vulnerability estimation, without comparisons under matched total training costs.

Questions for authors

1.Are the improvements over full-state distillation statistically significant across all four model-quantizer settings? 2.Does the proposed vulnerability score predict actual task failures better than simple action discrepancies or outcome-based state selection?

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

This paper studies how post-training quantization affects the robustness of vision-language-action (VLA) policies under environmental shifts. The authors observe that quantized policies may retain strong in-distribution performance while losing considerable task success on LIBERO-Plus. They further show that teacher–student action discrepancies are unevenly distributed along student rollouts and that intervening at high-discrepancy states is more useful than intervening at randomly selected states. Motivated by these findings, PIVOT-Q selects a small number of student-visited states using current and discounted future discrepancies, balances the selections across temporal phases, and combines vulnerability-weighted distillation with a Behavioral Anchor. Across four model–quantizer settings, it improves upon the corresponding quantized policies and uses far fewer supervised states than full-state distillation.

Critical strengths
  1. The problem is important and well-motivated. Preserving accuracy on the original benchmark is not necessarily enough for deploying quantized VLA models, and the paper highlights a potentially meaningful robustness gap under changes in viewpoint, lighting, object layout, and other conditions.
  2. The main GR00T-N1.5/QuantVLA result is encouraging: PIVOT-Q improves average LIBERO-Plus success from 66.6% to 81.3%, close to the 81.4% full-precision reference. The authors also include equal-16-state sparse-selection baselines and ablations of future discrepancy and the Behavioral Anchor, which are helpful controls.
Critical weaknesses
  1. I found a concrete inconsistency in the main experimental reporting. For GR00T-N1.5 with HoloQ-VLA (W4A4), Table 2 and the README in the submitted supplementary package report different results. The original and quantized reference results agree, but the two recovery results differ and reverse the ranking. This matters because the paper repeatedly claims that PIVOT-Q outperforms FSD in all four settings. I would like the authors to identify the final evaluation records behind Table 2 and explain which figures are correct.
  2. The gains over FSD are concentrated in one setting. Outside the primary GR00T-N1.5/QuantVLA case (+4.4 points), the reported average advantages are only +0.3, +0.1, and +0.4 points. I am not sure these small margins support a broad claim of consistent superiority without paired uncertainty estimates.
  3. More generally, PIVOT-Q involves several hand-designed decisions: four temporal phases, four selected states per phase, a four-step future horizon, a 0.9 discount, equal current/future score weights, weight clipping, and an anchor coefficient. Table 3 and Appendix F provide some useful ablations, but the sensitivity of the phase allocation, total budget, and independent loss-weighting component remains unclear. In particular, the random/uniform sparse baselines change both selection and distillation weights, so they do not isolate state selection alone.
Questions for authors
  1. Table 1 demonstrates the benefit of intervening at high-discrepancy states, and Appendix E.2 relates selection statistics to final outcomes. However, neither directly establishes whether the proposed combined vulnerability score identifies states where intervention has the greatest effect on task success. Could the authors compare interventions selected by the proposed score against those selected by instantaneous discrepancy alone under a matched intervention budget? In particular, how should future discrepancies be interpreted as evidence of current-state vulnerability when they may arise from subsequent states rather than the current one?
  2. Can the Behavioral Anchor be evaluated directly through original LIBERO performance before and after recovery? Also, what are the end-to-end recovery costs, including full-precision teacher queries used for scoring candidate states, rollout collection, optimization time, and memory, relative to FSD?
Reviewer_5KnsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose PIVOT-Q, a vulnerability-aware on-policy distillation framework, which aims to help quantized VLAs recover the original VLA performance. The framework selects vulnerable states that lead to rollouts which diverge from the original VLA, and distils the quantized student using the original VLA’s rollouts. The authors evaluate the method on LIBERO-Plus, a simulation benchmark for measuring VLA generalizability, and demonstrate that they can recover model performance to a greater extent than the full-state baseline for some VLAs and quantization frameworks.

Critical strengths
  1. The authors present a well-motivated definition of vulnerability, and ablate critical components of the definition to demonstrate the superior performance of PIVOT-Q when it uses their definition. Consequently, their main claim that PIVOT-Q recovers quantized VLA performance in simulation is well-supported by results presented in the paper.
  2. The paper also presents a comprehensive analysis of the action discrepancy distributions across three VLAs quantized using two quantization frameworks, which shows that the problem the paper tries to address is likely pervasive across multiple VLAs and is not easily mitigated by switching the quantization method.
Critical weaknesses
  1. I think the most critical weakness of the paper is its absence of real-world experiments. While LIBERO-Plus is a good simulation benchmark, given that the practical utility of VLAs is derived from their real-world deployment, the complete absence of real-world experiments limits readers’ interpretations of the utility of the proposed method. While it is possible that the method does in fact improve quantized VLA performance in simulation, it is possible that the method also simultaneously degrades real-world generalizability.
  2. Additionally, the benefits of the proposed approach over Full State Distillation is unclear: the LIBERO-Plus experiments only show that PIVOT-Q improves over FSD by 0.1 average points for π0.5\pi_{0.5} quantized using QuantVLA and by 0.3 for GR00T-N1.5 quantized via HoloQ-VLA. Thus, the reported improvements in the quantized VLA are modest for most VLAs and quantization methods.
Questions for authors
  1. Can the authors provide any evidence to show that quantized VLA performance and generalization capabilities do not degrade in the real-world after PIVOT-Q simulation training?
  2. Given that PIVOT-Q improves over FSD by achieving higher or comparable performance while using fewer states, do the authors observe a sizeable decrease in training time?
  3. Can authors explain why performance does not increase as much for π0.5\pi_{0.5} when quantized using QuantVLA (under both FSD and PIVOT-Q) for LIBERO-Plus tasks, whereas it recovers a sizeable amount of performance for GR00T-N1.5 using the same quantization algorithm for both methods?
Reviewer_7oqFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work identifies that quantized VLA policies are more fragile than full-precision policies under subtle environmental disturbances, such as changes in lighting, initial object pose, etc. To address this, the authors propose PIVOT-Q, which identifies and intervenes on vulnerable states, states that have high discrepancy between the quantized and full precision policy. In these vulnerable states (+ anchor states), they perform on-policy distillation with the full-precision policy as the teacher and the quantized policy as the student. Their empirical experiments show PIVOT-Q consistently recovers the performance lost by quantization, and can outperform full-state distillation using only 7.4% of its state-level distillation budget.

Critical strengths

The paper is clearly written and the on-policy distillation framework they present, PIVOT-Q, shows strong performance with low distillation budget. The proposed framework consistently improves the quantized policies across several model-quantizer settings and outperforms full-state distillation while supervising only ~7% as many states.

Critical weaknesses

The results presented in the main text (Tables 1-3) are between the full precision vs the quantized policy in environments with subtle changes. This provides no comparison of the performance between the quantized and the full precision policy when they are both evaluated on in-distribution environments. Therefore, the claim that there exists a generalization gap, that quantized policies perform worse than full precision policies under environmental shift than in-distribution environments, is only mentioned but not shown. Providing the comparison under the in-distribution environment would significantly strengthen the author’s work as this would more clearly show the generalization gap.

Questions for authors
  1. For the results, can you provide any intuition for why PIVOT-Q can sometimes outperform the full-precision model? Is this expected?
  2. In the vulnerability aware state selection, why did you choose to use normalized ranks instead of a continuous normalized value? What is the motivation behind this design choice?

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

Related papers

Discussion (0)

Sign in to comment.