acceptodds
Under review as a conference paper at ICLR 2027

Beyond Scalar Recovery: Residual Geometry in Quantized Vision-Language-Action Models

Abstract

Post-quantization recovery in vision-language-action (VLA) models is typically summarized by scalar error or task-level metrics, which measure how much error remains but not how that error is organized across a multi-step action chunk. We characterize this missing structure through teacher-relative residual geometry in a controlled OpenVLA-OFT setting. Across 13 heterogeneous ternary quantization configurations, mean pairwise Spearman similarity increases from 0.037 before adaptation to 0.557 afterward, with a demonstration-bootstrap median gain of +0.520 (95% CI [0.472, 0.568]). This geometric reorganization is not explained by uniform error shrinkage, and scalar recovery magnitude can diverge substantially from final residual alignment. The shared-alignment trend persists across additional optimization seeds, matched ground-truth-only adaptation, and a distinct 3-bit quantizer. Matched full-precision controls further demonstrate that shared alignment does not require an initial quantization perturbation. Finally, a leave-one-out template formed directly from other configurations’ adapted residuals strongly predicts the final residual geometry of a held-out configuration, reaching mean Spearman correlation 0.678 and substantially outperforming both raw geometry and a leave-one-out recovery-field predictor. Shared chunk-position and action-dimension effects explain much of this predictability, while the complete template retains additional predictive structure in reconstruction error.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, quantization, post-quantization adaptation, residual geometry, action chunking, robot learning, model compression

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

This paper studies how post-quantization adaptation changes the structure of a VLA policy's error. Using a task-adapted OpenVLA-OFT teacher on a simulated cube-stacking task, the authors build 8×6 teacher-relative residual maps for 13 ternary and three 3-bit layer-selection configurations, and compare rank-standardized maps across configurations before and after GT+KD adaptation. They report that adaptation substantially increases shared residual structure.

Critical strengths

Overall, the idea of studying how post-quantization adaptation changes the structure of a VLA policy’s error, rather than only its magnitude, is interesting.

Critical weaknesses
  1. All results come from a single task-adapted OpenVLA-OFT policy, a single simulated cube-stacking task, a selected set of layer configurations, and open-loop outputs on fixed inputs. Since the paper contributes an empirical observation rather than a method, the empirical scope is too narrow for an analysis paper.
  2. The paper shows that the remaining error becomes organized in a shared way, but does not show why this matters. There is no link between residual geometry and closed-loop task success, and no demonstration that the observation can inform quantization or adaptation choices.
  3. The paper is hard to follow because the main text is dense with numbers and switches between three residual representations.
Questions for authors
  1. Does the shared residual structure appear for another task or another VLA policy?
  2. How do the authors envision residual geometry being used in practice, and is it related to closed-loop task success?
2 comments
AuthorsOpenReview ↗

Response to Reviewer fdPC

Thank you for recognizing the motivation to study error organization beyond magnitude. Our finding is adaptation-associated reorganization of teacher-relative action residuals, not a quantization method or success predictor.

[W1, Q1] What is established, and how far does it generalize?

We study one adapted OpenVLA-OFT policy, one simulated cube-stacking task, and fixed-input predictions. No second policy, task, simulator, or closed-loop evaluation is reported; we do not claim transfer.

Nevertheless, across 13 ternary configurations (depth-based, random, and projection-based layer selections), mean pairwise Full-map Spearman alignment rises from 0.037 to 0.557 after adaptation (demonstration-bootstrap median gain +0.520, 95% CI [0.472, 0.568]). A new analysis of the submitted held-out predictions finds a positive alignment increase in all 19 held-out demonstrations (mean per-demonstration gain +0.437, 95% CI [0.387, 0.495]). This provides additional evidence of consistency within the task, not of cross-task generalization.

Controls refine the interpretation. Across six matched configurations, GT-only adaptation also raises Full-map similarity (0.186 to 0.445, versus 0.497 under GT+KD), as reported in the original submission. A distinct 3-bit quantizer and additional optimization seeds also support the within-setting trend. Crucially, matched full-precision controls achieve 0.421–0.843 alignment against counterpart-excluded adapted ternary references. Thus, quantization supplies heterogeneous initial perturbations but is not necessary for the shared adapted structure. We will make adaptation-associated residual geometry—not quantization specificity—the paper's headline.

[W2, Q2] What does this reveal that scalar recovery misses?

Uniform error shrinkage cannot explain this change. For B72, scalar recovery is 60.7%, yet the raw-to-adapted 48-cell rank correlation is −0.043 and a multiplicative scaling fit gives R² = −0.509. The relative ranking of residual cells changes substantially despite scalar recovery. Moreover, after removing additive chunk-position and action-dimension effects, Interaction alignment increases from 0.053 to 0.321 (bootstrap gain CI [0.191, 0.374]). Thus, the result is neither uniform shrinkage nor solely a common additive profile.

The practical output is an offline error-localization diagnostic: inspect the teacher-relative 8×6 residual map alongside T6. A post-hoc comparison of submitted held-out predictions shows why:

Adapted configurationB72S2
T60.048890.04895
Δx residual0.06980.0552
rot1 residual0.04620.0655

Despite nearly identical overall T6, the two configurations have different coordinate-level residual profiles. This identifies where their remaining discrepancies differ; the scalar average cannot. Alignment quantifies shared structure; residual maps locate the discrepancies. Neither is a policy-quality score.

We have not linked 48-cell alignment to closed-loop success or demonstrated better quantization selection, adaptation, or deployment. The demonstrated diagnostic value is coordinate-level error localization, complementary to—not a substitute for—task-level evaluation.

[W3] Making the contribution easy to follow

We will reduce the density of numbers and competing representations. We will reorganize the main story around three steps: (1) Full, the rank-comparable teacher-relative 8×6 residual map, as the primary result; (2) Interaction, used specifically to test whether alignment survives removal of additive chunk/dimension effects; and (3) Nuisance, moved to supporting sensitivity analysis. We will foreground full-precision controls and the B72/S2 comparison, moving secondary numbers out of the main narrative.

We will also demote template prediction from a headline contribution: its mean-final Spearman is 0.678 versus 0.658 for the additive baseline (difference +0.019, 95% CI [−0.001, 0.036]). This focuses the contribution on observed reorganization, without claiming an established rank-correlation advantage over the additive baseline.

We will state this within-task adaptation-associated structural finding, its diagnostic use, and its limits clearly in the revision.

Thanks for the responses. I feel like the rebuttals does not fully address my concerns. Besides, although the authors updated the manuscript, the paper is still not easy to follow as most of them are wordy discussions with statistical analysis. Therefore I maintain my original score.

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

The paper studies "residual geometry" — the organization of 8×6 teacher-relative residual maps (8 chunk steps × 6 continuous motion dimensions) — in a single task-adapted OpenVLA-OFT policy on a single simulated cube-stacking task (Isaac-Stack-Cube-MyCobot-IK-Rel-Visuomotor-Mimic-v0), evaluated on 152 held-out action-chunk samples drawn from 19 held-out demonstrations. Across 13 heterogeneous ternary weight-quantization configurations (plus three 3-bit controls), the authors adapt each quantized student with GT loss + knowledge distillation to the frozen full-precision teacher and measure, before and after adaptation, (i) mean pairwise Spearman correlation of the rank-standardized 48-cell residual maps, (ii) leave-one-out alignment to a template built from the other 12 configurations, and (iii) a matched full-precision control that applies the same adaptation protocol without any quantization. The three headline results are: (a) mean pairwise Full-representation Spearman similarity rises from 0.037 (pre-adaptation) to 0.557 (post-adaptation), with demonstration-bootstrap median gain +0.520, 95% CI [0.472, 0.568]; (b) the matched full-precision controls also develop substantial post-adaptation alignment (0.421–0.843 against a counterpart-excluded reference), so the shared structure is attributed to adaptation rather than to quantization per se; (c) a leave-one-out mean-final template predicts a held-out configuration's final residual shape at Spearman ρ = 0.678, outperforming the raw shape (−0.042) and a recovery-field predictor (0.339), though the additive chunk-position + action-dimension main-effects baseline already reaches ρ = 0.658. The paper proposes no new quantization or adaptation method; it is positioned as a descriptive analysis.

Critical strengths
  1. Methodologically careful on a narrow axis. The matched GT-only and full-precision controls are the right controls to run if one wants to attribute the shared-alignment phenomenon to adaptation rather than quantization or teacher distillation. The demonstration-level bootstrap (rather than action-chunk-level) correctly accounts for within-demonstration correlation. The provenance of calibration/test splits, trainable parameter subsets, and seeds is specified. This is an honest execution of a narrow experiment.
  2. One honest finding worth reporting: the matched full-precision controls demonstrate that the shared post-adaptation residual structure is a property of the adaptation protocol itself, not of quantization. This is the paper's clearest contribution and should probably have been its headline framing (and title).
Critical weaknesses
  1. Title and framing misrepresent what the paper demonstrates. "Beyond Scalar Recovery: Residual Geometry in Quantized Vision-Language-Action Models" promises a result about quantized VLAs. The paper's own controls (§5.1) establish that the central phenomenon — shared post-adaptation residual geometry — does not require quantization. A reader who stops at the abstract will acquire a false impression of what the paper found. The accurate title would be something like "Adaptation Produces Shared Residual Structure Across Fine-Tuning Configurations in a Simulated VLA Task," and under that title the significance is much weaker.
  2. The scope-to-claim ratio is extreme and no attempt is made to generalize. One policy, one simulated task, one robot morphology, 152 test action-chunk samples, open-loop action-space evaluation. For this evidence base, the paper nevertheless claims that "residual geometry" is a complementary view of post-quantization VLA adaptation in general (abstract, discussion, conclusion). There is no second task, second policy, second dataset, closed-loop evaluation, or cross-quantizer sanity check beyond a single 3-bit variant. The one quantitative transfer story offered — the ρ=0.678 leave-one-out predictor — reduces to the additive main-effects baseline (ρ=0.658) under the paper's own statistical test, which the paper acknowledges and then ignores.
  3. No actionable artifact. The paper proposes no method, no quantization scheme, no initialization strategy, no diagnostic for practitioners to use. If I accept every claim in the paper at face value, I do not know what to do differently when quantizing or fine-tuning a VLA. The leave-one-out template predictor is not evaluated as an initialization or warm-start for a new configuration's adaptation; it is evaluated only for its ability to predict a geometry that itself has no demonstrated downstream relevance.
Questions for authors
  1. Given that §5.1 shows the shared adapted structure emerges under matched full-precision controls (no quantization), why does the title and introduction frame the paper around quantization? Can you reconcile the title with the control result, or restate the framing to match what the controls demonstrate?
  2. The complete mean-final template beats the additive-only main-effects baseline by Spearman +0.019 with 95% CI [−0.001, 0.036], i.e., the bootstrap interval includes zero. The Conclusion still asserts that "the complete template retains additional predictive structure." Can you justify treating a result whose confidence interval includes zero as a finding worth stating in the Conclusion?
  3. Is there any evidence — in this paper or elsewhere — that two VLA policies with similar "residual geometry" (in your 48-cell sense) exhibit similar closed-loop task success, or that reducing residual geometry misalignment improves anything a robotics practitioner cares about? Without such evidence, why should a reader believe residual-geometry alignment is a useful axis of comparison for VLAs?
  4. Have you evaluated the shared-residual-geometry phenomenon on any second task, second policy, second simulator, or closed-loop success? If not, on what basis should the reader expect the trend to transfer at all?
  5. In 4.3 you note that R48 reaches 79% scalar recovery but only 0.483 alignment, while 3b-P48 reaches 7.6% recovery but 0.840 alignment. Doesn't this show that residual-geometry alignment can be high precisely for configurations that have failed to recover? If so, why is residual-geometry alignment a useful diagnostic?
  6. The paper closes with a leave-one-out template predictor at ρ=0.678. Have you tested whether using this predicted template as a warm-start initialization, regularization target, or auxiliary loss for adapting a new (held-out) quantized configuration actually improves either scalar recovery or closed-loop success? Without such a downstream use, the predictor is a curiosity, not a contribution.
1 comment
AuthorsOpenReview ↗

Response to Reviewer WuqS

Thank you for the detailed review. Our central finding is adaptation-associated shared residual geometry: it persists after removing additive effects (Interaction alignment) and appears without quantization. We address all three weaknesses and six questions.

[W1, Q1] Framing: adaptation, not quantization specificity

The 13 ternary configurations provide heterogeneous perturbations; quantization is not necessary. Across them, mean pairwise Full-map Spearman alignment rises 0.037 to 0.557 after adaptation (demonstration-bootstrap median gain +0.520, 95% CI [0.472, 0.568]). Matched full-precision controls also reach 0.421–0.843 alignment to counterpart-excluded adapted ternary references. We will revise the title, abstract, and contributions to foreground adaptation-associated shared residual geometry, without claiming a unique mechanism.

[W2, Q4] Scope and robustness

We evaluate one OpenVLA-OFT policy, one simulated cube-stacking task, and fixed-input predictions for 152 chunks from 19 held-out demonstrations. We report no cross-task, cross-policy, or closed-loop evidence; transfer remains untested.

A new analysis of submitted predictions finds higher mean pairwise Full-map alignment in all 19 demonstrations (mean per-demo gain +0.437, 95% bootstrap CI [0.387, 0.495]). This provides additional evidence of consistency within the task, not of cross-task generalization.

[W2, Q2] Additive baseline and prediction

Mean-final template Spearman is 0.678 versus 0.658 for additive chunk-position/dimension effects: +0.019 (95% CI [−0.001, 0.036]). The CI includes zero: we claim no established incremental rank-correlation advantage and will revise the Conclusion.

Both Spearman and centered-log RMSE appear in the original Table 1 and Appendix F. Centered-log RMSE is 0.218 versus additive 0.225 (paired improvement CI [0.005, 0.011]). This is a modest incremental scale-normalized profile-reconstruction signal, not reduced absolute action error. Independently, Sec. 4.1 reports mean pairwise Interaction alignment increasing from 0.053 to 0.321 (gain CI [0.191, 0.374]) after removing additive chunk/dimension effects. Additive effects explain most endpoint prediction, not all shared reorganization. Template prediction remains supporting, not the headline.

[W3, Q3, Q5] Why high alignment need not imply successful recovery

Alignment measures error organization, not error magnitude or success. The apparent paradox partly reflects percentage recovery. R48 starts with T6 0.392 and ends at 0.082 (~79% reduction); 3b-P48 already starts at 0.0546 and ends at 0.0504 (7.6% reduction). Thus, low percentage recovery need not mean worse final error: 3b-P48's final T6 is lower than R48's.

The alignment references also differ: R48's 0.483 is leave-one-out alignment against a template from the other 12 ternary configurations; 3b-P48's adapted 0.840 (up from 0.480 before adaptation) is alignment against a fixed adapted reference from all 13 ternary configurations, with no 3-bit configurations in the reference. They are not identical reference comparisons; neither measures success.

We have not linked 48-cell alignment to closed-loop success. It distinguishes shared from configuration-specific error patterns; the maps localize error, rather than predict policy quality.

A new post-hoc comparison of submitted held-out predictions illustrates this:

Adapted configurationB72S2
Overall error (T6)0.048890.04895
Δx residual0.06980.0552
rot1 residual0.04620.0655

Despite nearly identical T6, B72 concentrates more error in Δx and S2 in rot1. The 8×6 residual map alongside T6 localizes errors concealed by scalar averaging. This is a reproducible offline error-localization diagnostic, not a validated improvement in policy selection or deployment.

[W3, Q6] Template-guided interventions

We have not trained with template-based warm-starts, regularizers, or auxiliary losses, and claim no improvement in recovery or closed-loop success. A leakage-free test needs calibration-only adapted residuals excluding the target, plus matched training against uniform KD and an additive-template control; test-set templates cannot be training signals. Matching an error template need not reduce error magnitude. Our contribution is the controlled finding and offline diagnostic, not an untested training method.

Presentation and revision. We will foreground Full-map geometry and full-precision controls, use Interaction for the nonadditive check, make Nuisance secondary, and simplify figures. We will detail changes in the revised PDF.

Reviewer_J3LxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how post-quantization adaptation changes the residual geometry of VLA action predictions. Using OpenVLA-OFT with different quantization configurations, the authors observe increased alignment of residual patterns after adaptation and show that shared residual templates can predict the final residual structure.

Critical strengths

The paper provides an interesting empirical observation that scalar recovery metrics do not fully capture changes in action prediction errors.

The analysis includes multiple quantization configurations, optimization seeds, and matched full-precision controls, which strengthen the empirical findings.

The statistical analysis is reasonably thorough, including demonstration-level bootstrap confidence intervals and several controlled comparisons.

Critical weaknesses
  • All experiments use a single VLA model and one simulated manipulation task. The analysis is restricted to offline action predictions, without closed-loop evaluations or real-robot experiments. This substantially limits the generality and practical significance of the findings.

  • The work primarily characterizes an empirical phenomenon without proposing a new quantization or adaptation method. Moreover, the full-precision controls exhibit similar residual alignment, suggesting that the observed behavior is not specific to quantization.

  • The mean-final predictor achieves a Spearman correlation of 0.678, compared with 0.658 for the simple additive baseline. This small improvement raises questions about how much additional structure is captured beyond shared action-dimension and chunk-position effects.

Questions for authors
  • Does the observed residual alignment generalize to other VLA architectures and more diverse manipulation tasks, such as LIBERO?

  • Does stronger residual alignment correlate with better closed-loop task success? Can this be demonstrated in simulation or on real robots?

  • Can the learned residual geometry be used to improve quantization or adaptation, rather than serving only as a descriptive diagnostic?

2 comments
AuthorsOpenReview ↗

Response to Reviewer J3Lx

Thank you for recognizing the importance of residual structure beyond scalar recovery and the rigor of our controls and demonstration-level statistical analysis. We address the concerns about scope, practical significance, and template predictability below.

[W1, Q1–Q2] Scope, robustness, and closed-loop relevance

Across 13 ternary configurations, adaptation increases mean pairwise Full-map Spearman alignment from 0.037 to 0.557, with a demonstration-bootstrap median gain of +0.520 (95% CI [0.472, 0.568]). This is a substantial reorganization of teacher-relative action error, not just scalar error reduction.

To test whether a few demonstrations drive the result, we additionally recomputed mean pairwise Full-map Spearman alignment separately for each of the 19 held-out demonstrations (eight sampled chunks each). Full-map alignment increased in all 19 demonstrations, with mean per-demonstration gain +0.437 (95% bootstrap CI [0.387, 0.495]). This provides additional evidence of consistency within the task.

We distinguish within-task robustness from generalization: our study evaluates one task-adapted OpenVLA-OFT policy, one simulated manipulation task, and fixed-input predictions. We have not tested LIBERO, other VLA architectures, or a connection to closed-loop success. The contribution is a controlled characterization of residual organization, not a claim of improved robot execution.

[W2, Q3] Why the finding matters beyond scalar recovery

Residual geometry reveals where remaining teacher-relative error concentrates across action-chunk steps and motion dimensions, which aggregate error obscures.

First, adaptation is not uniform shrinkage. B72 achieves 60.7% scalar recovery, but its raw-to-adapted cell-rank correlation is −0.043, and a multiplicative scaling fit gives R² = −0.509. More importantly, after removing additive chunk-position and action-dimension main effects, mean pairwise Interaction alignment increases from 0.053 to 0.321 (gain CI [0.191, 0.374]; originally reported in Sec. 4.1). Thus, the shared reorganization is not merely an additive profile.

Second, a new post-hoc analysis of submitted held-out predictions provides a concrete scalar-matched diagnostic:

Adapted configurationB72S2
Overall error (T6)0.048890.04895
Δx residual0.06980.0552
rot1 residual0.04620.0655

Despite nearly identical T6, B72 has higher Δx error, whereas S2 has higher rot1 error. Examining the 8×6 residual map alongside T6 therefore identifies error concentrations that a scalar average cannot. This is a reproducible offline error-localization diagnostic, not an experimentally validated improvement to downstream performance.

Matched full-precision controls yield 0.421–0.843 Full-map alignment against counterpart-excluded adapted ternary references. This establishes that quantization is not necessary for shared residual structure; heterogeneous quantization configurations serve as controlled starting perturbations for studying an adaptation-associated effect. The GT-only, seed, and distinct 3-bit controls further support its within-setting robustness.

[W3] What does the full template add over additive effects?

The mean-final template reaches Spearman 0.678, versus 0.658 for the additive baseline; the paired gain is +0.019 (95% CI [−0.001, 0.036]). We do not claim a statistically established incremental rank-correlation advantage.

Both Spearman and centered-log reconstruction RMSE were already included in the original submission (Table 1; Appendix F). Spearman measures residual-cell rank ordering; centered-log RMSE measures reconstruction accuracy of the scale-normalized log-residual profile. The complete template modestly reduces RMSE from 0.225 to 0.218 (paired improvement CI [0.005, 0.011]). Thus, additive effects explain most endpoint predictability, while a small additional reconstruction signal remains. The independently observed Interaction-alignment increase supports nonadditive shared reorganization, a distinct claim from incremental endpoint prediction.

Manuscript revision

We are clarifying the adaptation-centered framing, elevating the full-precision controls, simplifying the figure hierarchy, incorporating the new within-task analyses, and positioning template prediction as supporting characterization. We will identify the exact changes when the revised PDF is uploaded.

I maintain my original score after reading the whole responses, partially resolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.