acceptodds
Under review as a conference paper at ICLR 2027

QuoVLA: Quotient Space for Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models adapt pretrained visual-language features for robot control through action supervision. Action supervision can guide task-specific feature learning and the suppression of action-irrelevant variation. To limit this variation, we introduce QuoVLA, which combines joint feature learning with a quantized prefix adapter. The adapter processes visual features and language embeddings before a shared language backbone and action expert, using flow matching and relative temporal-complexity regularization. This regularization uses a stop-gradient raw-prefix reference to discourage excess temporal variation in the flow predictions of the learned policy. We interpret this policy through an action quotient that groups representations with the same chosen optimal action law. For this fixed action law, a conditional characterization defines an ideal compression target without assuming sufficient pretrained features or guaranteeing that training attains the target. The target motivates paired representation diagnostics alongside task success evaluation. Task success improves over the primary baseline on four simulation benchmarks and four real-robot tasks, while the diagnostics support selective compression under the evaluated perturbations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision language action

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_VGxaOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies what information a VLA representation should preserve for action prediction. It introduces an action-quotient view that groups representations with the same optimal action law, and proposes QuoVLA as a practical approximation using a quantized prefix adapter, gated STE, and relative temporal-complexity regularization. Experiments on LIBERO, LIBERO-PRO, LIBERO-Plus, RoboTwin 2.0, and four real-robot tasks show consistent gains over the main baseline, together with representation diagnostics intended to support selective compression.

Critical strengths
  1. The paper gives a useful and fairly clean perspective on VLA representation learning: action-relevant distinctions should be preserved while action-irrelevant variation should be compressed. The action-quotient formulation makes this intuition more precise than the usual “robust representation” argument.

  2. The empirical results are strong and reasonably broad. The gains are not limited to standard LIBERO, but also appear under distribution shifts, RoboTwin, and real-robot evaluation. The paired representation/action diagnostics are also useful because they try to test the claimed selective-compression behavior rather than only reporting success rates.

Critical weaknesses
  1. There is still a noticeable gap between the theoretical action quotient and the actual method. The quotient is defined through equality of the optimal action law, while the implementation uses scalar quantization of prefix features. Quantization can remove information, but it does not directly identify action-equivalence classes, so the theory currently feels more like motivation than a derivation of the algorithm.

  2. The role of RTC is somewhat heuristic. Temporal complexity of the flow prediction is not obviously equivalent to removing action-irrelevant representation variation, and the main paper does not make its independent contribution especially clear. A cleaner ablation separating quantization, RTC, and their interaction would strengthen the method claim.

Questions for authors
  1. Can the authors provide a clearer ablation of RTC, e.g. quantization only, RTC only, and quantization + RTC under the same training setting? This would help clarify how much of the gain comes from the bottleneck itself versus the temporal regularizer.

  2. Can the authors clarify what empirical evidence most directly supports the connection between the learned quantized representation and the proposed action quotient? The current diagnostics are consistent with selective compression, but do not seem to establish that the learned representation approximates the quotient in a stronger sense.

Reviewer_CRwFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes QuoVLA, a small modification to pi0.5 that inserts a single layer transformer with N-bit activation quantization and an auxiliary training signal, called 'relative temporal complexity (RTC)' that penalizes jaggedness of the predicted flow velocity. The method is motivated by an 'action quotient,' which should partition the representations. The paper reports gains over pi0.5 and other baselines in LIBERO, LIBERO-PRO, LIBERO-Plus, RoboTwin 2.0, and real-robotic tasks.

Critical strengths

Real-hardware evaluation: the four real-world tasks are evaluated substantially (100 trials per task and policy), and the confidence bounds give credence to their method. The gains over the baselines (pi0.5 in particular) are strong. This is probably the largest piece of evidence in the paper that their methodology changes the policy's behavior in a productive way since they use the same backbone as pi0.5.

Extensive simulation results: the authors evaluate on a suite of simulation benchmarks (LIBERO, LIBERO-PRO, LIBERO-Plus, and RoboTwin 2.0), and cover in-distribution performance, out-of-distribution performance, perturbations, and other sweeps. More tests corresponds to stronger power in determining if there is a real signal, which is why I think the extensive evals are important.

Critical weaknesses

Theoretical framing doesn't support method: the central aspect of QuoVLA is the 'action quotient,' which to my understanding is to partition the representations into level sets of the 'optimal action law.' Theorem 1 says this is the coarsest partition preserving that law. However, nothing connects this idea to their method. The per-token quantization they introduce is supposed to address this. However, this should uniformly reduce precision across all coordinates, regardless of action relevance or not, which was the method's motivation. There is also no mechanism for merging inputs with the optimal actions as the theorems suggest, beside just the standard implicit shaping encouraged by the learning algorithm.

The authors repeatedly state that the quantizer does not identify nor realize the equivalence classes of the action quotient, while still presenting it as an approximation to them, without any mechanism nor evidence that it actually is approximating them If every theoretical claim is then quickly disclaimed as not holding in general, and there is no empirical validation of the method approximating the theorems, then there is no account for why the method works and contributions reduce to an empirical one. This is fine, but then I don't understand why the maths are introduced, nor why the authors 'characterize an ideal action quotient and give sufficient conditions under which a learned representation lowers expected squared test error.'

Empirical evidence does not support stated contributions: the paper's ablations do little to help the reader understand what is important. For example, removing the RTC entirely only changes LIBERO averages by less a point (Appendix A.9). It is also unsettling that on LIBERO-Plus, the fine-tuned pi0.5 baselines perform worse than zero-shot pi0.5, which is suggestive of poor fine-tuning, making the results harder to believe.

Questions for authors
  1. I am trying to understand why QuoVLA scores higher on Hard than Easy in 45 / 50 tasks in RoboTwin 2.0. Were all RoboTwin 2.0 methods trained on the same data splits?

  2. What weight for the RTC (lambda_tc in Appendix A.9) were used for the main results reported in the paper? I couldn't find this information, and it would help address my second weakness listed above, where if lambda_tc=0, the results didn't change much relative to what you reported.

  3. The appendix states that for Table 8, 'The interventions behind these labels cannot be reconstructed from the available records. These results are not used to attribute separate effects to the reference branch and its regularizer.' I am not sure what this means. Why are these results included if the results cannot be reconstructed?

  4. I am very confused by Figure 2. It is stated in the methodology section that the quantized adapter acts on the embedding-stage tokens, before the language backbone. It is also stated again in the introduction: 'The adapter acts before the shared language backbone, and its gated surrogate gradient lets action supervision update the upstream features.' However, Figure 2 suggests that the adapter acts on the final features already processed through the VLM. Can you please clarify which this is, and update the manuscript accordingly?

  5. Have you tried to empirically validate any of the theorems with your method? If so, what were the results? For example, in Theorem 2, is a positive bound obtained by your method?

Reviewer_edaDOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

QuoVLA places a quantized Transformer adapter before the shared language backbone and action expert. It jointly learns features with flow matching, a gated straight-through estimator and relative temporal-complexity regularization against a stop-gradient raw-prefix route. An ideal action quotient motivates selective compression. Reported success reaches 99.6% on LIBERO and 88.0% on four real-robot tasks versus 64.5% for the primary baseline.

Critical strengths
  1. Introducing the concept of quotient spaces into VLA representation learning is a worthwhile direction. The paper highlights that VLA training should not only learn action-relevant information but also suppress action-irrelevant variation. This perspective is insightful and relevant to the VLA community.
  2. The evaluation covers standard manipulation benchmarks, robustness benchmarks with perturbations, and real-robot tasks. Strong performance across multiple benchmarks provides encouraging empirical evidence for the method’s effectiveness.
Critical weaknesses
  1. The manuscript repeatedly refers to an appendix, but it appears to be missing from the provided submission.
  2. The theoretical analysis does not yet explain the benefits of the specific implementation. Theorem 1 characterizes the partition induced by a chosen optimal action law, while Theorem 2 guarantees a reduction in risk only when the variation, information-loss, and prediction-deviation terms satisfy the condition in Eq. 6. These quantities are not measured, and no empirical evidence establishes that this condition holds for the trained policy.
  3. The reported ablations do not fully disentangle the contributions of RTC, adaptive gradient gating, and quantization. Although the paper includes a variant without Adaptive STE, the available comparisons do not establish the independent effects of these components or their interactions. Consequently, it remains unclear how much of the improvement stems from quantization-based selective compression versus temporal regularization or gradient scaling.
Questions for authors
  1. Please provide the exact supplement containing appendix A.1–A.12 and Eq. 25.
  2. Could the authors provide controlled ablations that isolate the contributions of RTC and adaptive gradient gating? In particular, comparisons with and without RTC under otherwise identical settings, and comparisons between learned and fixed gradient gates, would help clarify the sources of the reported gains.
  3. Could the authors provide further evidence, beyond the current paired representation diagnostics, that the learned 8-bit quantized representation approximates the proposed action quotient? In particular, does it consistently suppress variation among inputs requiring the same actions while preserving distinctions among inputs requiring different actions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.