acceptodds
Under review as a conference paper at ICLR 2027

CAEM-BC: Condition-Adaptive Elastic Modal Flow with Bounded Continuation for Vision-Language-Action Policies

Abstract

Improving a pretrained vision-language-action policy with limited demonstrations requires learning useful action corrections and joining them to commands already executed. We present Condition-Adaptive Elastic Modal Flow with Bounded Continuation (CAEM-BC), an adapter that refines action chunks while keeping the pretrained policy frozen. A condition-dependent elastic operator defines a low-rank temporal basis for residual prediction, preserving the base proposal outside the correction subspace. Bounded continuation then joins the corrected chunk to executed history while respecting action limits. CAEM-BC uses fewer than 6M trainable parameters and trains for 3,000 steps on 40 demonstrations per task. Across three backbones and four LIBERO suites, it improves average success for every backbone. An additional Spatial evaluation improves SmolVLA success by 3.2 percentage points while reducing executed-arm jerk by 40.5%, with further gains in execution quality under new initial states and language rewrites. Controlled component interventions show that residual correction repairs task failures, while continuation preserves the repaired outcomes and improves execution continuity. These findings connect efficient adaptation to task completion through a learned correction space and history-aware execution.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action policies, robot learning, parameter-efficient adaptation, action chunking, elastic modal representations, residual learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_mDjHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CAEM-BC, a lightweight adaptation framework for frozen vision-language-action (VLA) policies. The method consists of two main components: (1) a condition-dependent low-rank temporal representation for predicting residual corrections to pretrained action chunks, and (2) a bounded continuation mechanism that modifies the corrected actions to improve continuity with previously executed commands. The temporal representation is obtained from the low-energy eigenvectors of a learned elastic operator, while the residual coefficients are predicted using a combination of flow matching and deterministic regression. The continuation mechanism is formulated as a box-constrained quadratic optimization problem involving third-order temporal differences. The authors evaluate the method on three VLA backbones across four LIBERO suites, using 40 demonstrations per task. They report small improvements in task success and more substantial reductions in the third-order differences of normalized action commands.

Critical strengths
  • Adapting pretrained VLA policies without modifying their backbone parameters is a practically relevant problem. A lightweight adapter could potentially reduce the cost of specializing large pretrained policies.
  • The experiments include SmolVLA, π0.5\pi_{0.5}, and Adapter-Pro, covering four LIBERO suites. Evaluating a common adaptation framework across different policy architectures is a positive aspect of the study.
Critical weaknesses

1. Unclear motivation and poor writing.

The paper is unnecessarily difficult to understand. The abstract fails to clearly communicate the problem, motivation, or central contribution. The introduction introduces vague expressions such as "organize a compact correction across time," "fixed temporal basis," and "condition" without explaining their meanings or practical significance. The writing is dominated by abstract terminology rather than concrete insights. Overall, the presentation reads more like an automatically assembled collection of technical descriptions than a coherent, hypothesis-driven research paper.

2. Overly complicated method without a clear central insight.

The method combines elastic operators, eigendecomposition, modal flow matching, learned trust, and constrained continuation, but provides little high-level explanation of why these components are necessary. Section 3 immediately introduces mathematical details without first establishing the key intuition. In particular, the necessity of flow matching is questionable when inference uses a single deterministic prediction. The method appears overengineered relative to the problem it addresses.

3. Marginal experimental improvements.

The main results in Table 1 show average success-rate improvements of only +1.20, +0.45, and +0.15 percentage points for SmolVLA, π0.5\pi_{0.5}, and Adapter-Pro, respectively. These correspond to merely 24, 9, and 3 additional successes across 2,000 episodes per backbone. Five of the twelve backbone-suite combinations show no improvement. Such negligible gains hardly justify the complexity of the proposed framework.

4. Ablations do not consistently support the proposed design.

The central contribution is the condition-adaptive elastic representation, yet Table 6 provides contradictory evidence. On the SmolVLA ten-task evaluation, the elastic-residual model with learned trust achieves 77% success, compared with 86% for the frozen backbone. On the disjoint task-4 evaluation, it drops to 42% versus 62% for the backbone. Fixed DCT even outperforms the elastic basis in several configurations. These results undermine the claimed necessity of the proposed representation.

Questions for authors
  • Since the learned elastic operator is differentiated through eigendecomposition, how does the method handle nearly repeated eigenvalues, eigenvalue crossings, and potential discontinuities in the eigenvectors?

  • The method corrects arm trajectories while retaining the backbone's gripper commands. How does it prevent temporal misalignment between corrected arm motion and grasp/release events?

  • The adapter relies on internal action-generation features. Which representations are extracted from each backbone, and how sensitive is the method to the choice of feature layers or sampling stages?

3 comments
AuthorsOpenReview ↗

Response to Reviewer mDjH (1/2): Motivation and Experiments

Thank you for the detailed review. We have revised the paper to explain the method in more concrete terms and added simpler controls to examine its components.

Central idea and presentation. We rewrote the abstract and introduction around a frozen VLA's proposed action chunk. The adapter learns a task-specific residual in a small, condition-dependent temporal subspace; outside that subspace, the base proposal is retained before continuation. Bounded continuation then connects the corrected chunk to executed history. We define the condition and explain these steps before the operator equations (Sections 1 and 3.2–3.4).

Residual structure and FM. Five temporal modes per arm coordinate give 30 coefficients for the whole chunk and preserve the backbone proposal outside the correction subspace before continuation (Eq. 5). Under the same action-domain supervision, the time-domain MLP and elastic coefficient predictors averaged 84.40% and 85.07% success, respectively (Base 82.40%). FM adds supervision along the coefficient path during training (Lipman et al., ICLR 2023). For each action chunk proposed by the frozen VLA, the adapter calls its coefficient predictor once to obtain all 30 modal correction coefficients. Adding FM raised mean success from 85.87% to 86.07% (Table 6).

Success rates and component results. The main-table increases across the three backbones are 1.20, 0.45, and 0.15 percentage points. Table 7 retains the learned-trust elastic component results you identified: 77% on the ten-task bank and 42% on the disjoint task-4 bank, compared with Base at 86% and 62%. The complete deployment reaches 86% and 70%, respectively, with the recorded guard/selection settings. The new three-seed ablations use a separate Spatial bank. In the matched FM comparison, elastic has a 0.40-point higher mean than DCT, while DCT is higher in two of three seeds.

AuthorsOpenReview ↗

Response to Reviewer mDjH (2/2): Implementation

Eigendecomposition. The operator is symmetric tridiagonal with positive edge couplings (ki≥0.05k_i \geq 0.05), so its eigenvalues are simple in exact arithmetic. We solve the small eigensystem in float64, use a deterministic sign convention, clip gradient norms at 5, and reject non-finite gradients. The retained projector is invariant to eigenvector sign flips. Appendix B.5 discusses the numerical sensitivity associated with small spectral gaps.

Arm–gripper alignment. The arm and gripper retain their original action indices and horizon. At each step, the corrected six arm coordinates are merged with the backbone's corresponding gripper command, so grasp and release commands follow the backbone's timing. Trust scaling and action-preservation losses keep the arm correction tied to the proposal, and the next query uses the newly observed state (Appendix B.2).

Backbone features. For SmolVLA and π0.5\pi_{0.5}, we capture the input to the action-output projection at the final denoising step and form five temporal summaries, giving 3,600 and 5,120 feature coordinates, respectively. Adapter-Pro uses an 896-dimensional mean-pooled action-token representation. These capture points were fixed in the reported evaluation; we did not conduct a layer or denoising-stage sensitivity study (Appendix B.2).

AuthorsOpenReview ↗

Follow-up for Reviewer mDjH

We’ve revised the abstract and introduction and added simpler controls to better explain the design. We’ve also responded to your questions about the component results and implementation. We’d appreciate hearing which concerns you feel are still open after reading our responses.

Reviewer_e8CaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAEM-BC, a lightweight action correction framework for few-shot adaptation of frozen VLA policies. The proposed method constructs a low-rank temporal basis via a conditional adaptive elastic operator, predicts action residuals using one-step Flow, and incorporates history-aware bounded continuation optimization to enhance action correction capability and execution continuity. The authors evaluate the approach across three pretrained VLA models and four LIBERO task suites. Experimental results demonstrate that, using only 40 demonstrations per task, fewer than 6 million trainable parameters, and short training cycles, the method achieves improved average task success rates and reduced higher-order action variations.

Critical strengths

None

Critical weaknesses
  1. The Introduction of the paper is overly brief, lacking an in-depth analysis of the limitations of existing VLA adaptation and action chunk continuity methods, and failing to clearly articulate the core advantages of the proposed CAEM-BC over existing approaches. In particular, the authors do not sufficiently explain why it is necessary to simultaneously introduce the conditional adaptive elastic temporal basis, Modal Flow, and Bounded Continuation, nor what irreplaceable advantages these designs offer compared to existing residual correction and action smoothing methods.

  2. As a paper targeting VLA robotic manipulation policies, the authors evaluate their method solely within the LIBERO simulation environment, lacking experimental validation on physical real-world robots. In particular, higher-order differences of normalized action commands in simulation cannot adequately reflect motion smoothness and control performance during real-world robot execution. Consequently, the absence of real-robot experiments leaves the effectiveness and robustness of the proposed method in practical robotic manipulation insufficiently validated.

Questions for authors

In Appendix B.3, the paper introduces a joint optimization objective comprising approximately ten loss terms, covering multiple aspects such as Flow Matching, deterministic regression, residual reconstruction, action preservation, smoothness, and elastic field regularization. Although the appendix provides the corresponding mathematical definitions and partial design explanations, the independent contribution of each loss term to the final performance remains unclear, as does the necessity of such a complex joint optimization objective. The authors should further clarify the design motivations and interrelationships of the primary loss terms and validate their necessity through more comprehensive ablation studies. In particular, given that the proposed Modal Flow already incorporates a regression loss that directly supervises the deterministic output, could the authors further explain the independent benefits brought by the Flow Matching loss compared to direct residual regression?

3 comments
AuthorsOpenReview ↗

Response to Reviewer e8Ca

Thank you for pointing out where the motivation and training objective were difficult to follow. We have rewritten the opening to start with the concrete action proposal from a frozen VLA. The adapter learns a correction in a low-dimensional temporal space, while bounded continuation joins the corrected chunk to executed history. Sections 3.2–3.4 now explain these roles before introducing the equations.

Why include FM alongside direct regression? Direct regression trains the predictor at the input used in deployment. FM also trains it to predict the velocity along the path to the target modal coefficients (Eq. 7), so each demonstration supervises the same predictor at intermediate states as well (Lipman et al., ICLR 2023). For each action chunk proposed by the frozen VLA, the adapter calls its coefficient predictor once to obtain all 30 modal correction coefficients. With the same readout, adding FM raised mean success from 85.87% to 86.07% across three training seeds (Table 6).

The table also includes a direct time-domain MLP residual (84.40 ± 1.39%) and an elastic direct-coefficient residual under common action-domain supervision (85.07 ± 1.10%). Removing the direct jerk-excess term from the reference gives 85.27 ± 0.46%. Appendix B.4 groups the training objectives by function and explains their roles.

Execution continuity. Table 5 isolates what bounded continuation changes during execution. On 50 paired SmolVLA task-4 trials in LIBERO, the residual-only and full models succeeded on the same 45 trials. Adding continuation reduced mean arm-command jerk by a further 38.31%. A related hardware result comes from Lozer et al. (2025): in pick-and-place trials on a Franka Panda, accelerometers on the end effector recorded less vibration with the lower-jerk trajectory. This is why we examine the commands the controller receives, including the transition from one chunk to the next.

AuthorsOpenReview ↗

Follow-up for Reviewer e8Ca

Thank you for the questions. We’ve added the direct residual baselines and the FM ablation in Table 6, and clarified how each component is used. We’d be interested to hear whether these additions address your concerns about the design. If so, we hope you might reconsider your score.

Thank you for the authors' response. However, I am still concerned about the lack of real-world robot experiments, as I could not find any newly provided experimental results addressing this concern.

Reviewer_2kXCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAEM-BC, a lightweight adaptation framework that adds an action correction module and a history-aware smoothing module after the action head of a frozen VLA model. The method uses a condition-adaptive elastic basis to predict low rank residual corrections to the original action chunks. It then applies bounded continuation, which incorporates previously executed actions into a constrained optimization problem to improve continuity both within and across action chunks. The method is evaluated on three VLA backbones across four LIBERO suites. The results show modest improvements in task success rates and substantial reductions in action command jerk.

Critical strengths
  1. This paper focus on correcting imperfect action predictions and maintaining continuity across consecutive action chunks. The history-aware bounded continuation mechanism incorporates previously executed actions into the optimization of predicted action sequences. This provides a practical way to improve chunk continuity while preserving the corrected actions.

  2. The proposed method provides a structured low-dimensional space for residual action prediction. The framework keeps the original VLA backbone and requires relatively few trainable parameters making it useful for adapting existing VLA models.

  3. The experiments cover three different VLA backbones and four LIBERO suites. This demonstrates that the method is applicable across different models.

Critical weaknesses
  1. The paper emphasizes execution reliability and continuity, but all experiments are conducted in simulation. The reported jerk metric measures normalized action commands, which does not necessarily correspond to the physical jerk of a robot end-effector. Given the practical motivation of the work, real-robot experiments would significantly strengthen the evidence for improved execution quality and practical applicability.

  2. Although the method improves average success rates across all three backbones, the absolute improvements are relatively small. Some task suites show no improvement. Table.3 reports 37 repaired failures but 32 regressions for SmolVLA on LIBERO-Spatial, indicating that the correction can both improve and degrade individual trajectories. Additional statistical analysis or multi-seed evaluation for main results should provide.

  3. The condition-adaptive elastic basis is an key design, the evidence that it consistently outperforms simpler fixed temporal bases is limited. The supplementary multi-seed results do not show a consistent advantage over fixed DCT. It is therefore unclear how much the flow matching objective contributes beyond direct residual coefficient regression.

Questions for authors
  1. Could the authors further clarify the advantage of the condition-adaptive elastic basis over a fixed DCT basis? In particular, how should the mixed results in the supplementary multi-seed comparison be interpreted?

  2. Given that execution continuity is one of the main motivations, have the authors evaluated whether the reduction in action command jerk translates into smoother physical motion?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2kXC

Thank you for these questions. We agree that the size and variability of the gains, the DCT comparison, and the interpretation of jerk all deserve a clearer account.

Success and seed variation. In the main Spatial evaluation, SmolVLA has 37 repairs and 32 regressions, for a net gain of five successes (Table 3). The new three-seed study uses a separate SmolVLA Spatial evaluation bank: the full model succeeds on 423/500, 439/500, and 429/500 episodes, giving 86.07 ± 1.62%; the fixed Base succeeds on 412/500 (82.40%). The Base has one evaluation, while the trained variants have three training seeds. Table 6 and Appendix D, Table 13 report the results and individual counts.

Elastic basis versus DCT. A fixed DCT basis gives every query the same temporal modes. Our operator uses the current condition to set the cost of changing the correction between adjacent steps (Eq. 4). A larger coupling penalizes a change more strongly at that pair of steps, so the retained modes can place variation at different points in the chunk. We compare the bases at the same rank. With FM, elastic averaged 86.07% and DCT 85.67%, with DCT ahead in two of three seed pairs. Without FM the means were 85.87% and 84.73% (Table 6). Thus the adaptive basis offers local control over where a correction varies, and the success advantage on this bank varies by seed.

Physical motion. We used LIBERO to measure how abruptly the arm commands change when the policy starts a new chunk. On the 392 SmolVLA Spatial episodes that both policies solved, CAEM-BC reduced the mean third difference of the normalized arm commands by 35.54% overall and by 28.48% across chunk boundaries (Table 3). Robot experiments give this comparison a physical motivation. On a Franka Panda pick-and-place task, Lozer et al. (2025) reduced planned trajectory jerk and measured lower vibration at the end effector. On a three-axis Cartesian robot, Béarée and Olabi (2013) reshaped the jerk profile and brought residual vibration below 23% of the level measured with a conventional jerk-limited profile.

AuthorsOpenReview ↗

Follow-up for Reviewer 2kXC

Thank you again for your review. We’ve added the per-seed comparison with DCT and clarified what the command-smoothness metric measures. Could you let us know what you think of the new results, and whether they address your main concerns? We’d be happy to clarify further.

Thank you for the added multi-seed results and ablations, they address some of my concerns. However, the evaluation remains simulation-based. While the physics-based model is useful, it does not establish that the command-smoothness gains transfer to real hardware. The multi-seed results are also limited to one backbone and one suite, and the DCT/FM gains are modest and variable. I therefore maintain my original rating of 2.

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

The paper introduces CAEM-BC, a novel and efficient residual learning method that combines a condition-adaptive elastic modal flow method with bounded continuation, to improve both success rates and execution continuity.

Critical strengths
  1. A novel inductive bias for residual learning. Limiting the correction to the low-energy modes is new in this literature, to my knowledge, and it has a real structural content.
  2. Bounded continuation is simple, well-characterized, and likely transferable. It is a strictly convex quadratic program with a box constraint, deployed via 64 projected-gradient steps. Empirically it reduces jerk at essentially zero success cost. I think it's probably transferable for other chunked policies. And the overhead is small.
  3. Two two components are complementary, and the paper isolates them. The intervention study cleanly attributes task repairs to the residual branch and the jerk improvements to continuation (plus the CAEM components)
  4. Clean presentation. The main line of argument is easy to follow.
Critical weaknesses
  1. Success-rate evidence is broad but thin. The main table averages +1.20 / +0.45 / +0.15 points across the three backbones, but the benchmark is already near saturation for two of the three backbones, and the cleanest decomposition uses only a single task. It would help to include more benchmarks, or extend the mechanism decomposition to more tasks, or state more explicitly that the execution quality is the primary claim rather than success rate.
  2. The method is too complex, and the necessity is not demonstrated. The residual branch structure itself is already not simple, and the objective has 10 grouped terms! In the current design, structure and auxiliary losses are stacked on top of each other, and no experiments separate them. A simple unstructured MLP residual baseline trained with the same losses (and/or the elastic structure without the jerk loss) would directly answer how much the elastic construction itself contributes.
  3. Some terminology is ambiguous. The paper used a lot of physics words, which is not bad by itself, but it would be better to build the physics picture in the beginning of / before Section 3.2. And some words like "carrier" is hard to follow, its meaning is only inferrable until appendix.
Questions for authors
  1. Can you show that the learned basis varies with the condition, for example, visualizing the stiffness/tethering fields across task phases?
  2. The history h enters the condition pack but the backbones have no access to history. How much of the success and smoothness gains come from h alone? An ablation removing h from c (separate from the \mu=0 continuation ablation) would answer this.
1 comment
AuthorsOpenReview ↗

Response to Reviewer fjQR

Thank you for the encouraging review and for recognizing the complementary roles of residual correction and bounded continuation. Your suggestions helped us improve the paper.

We added a direct MLP residual baseline, a variant without the jerk-excess loss, and separate controls for history in the predictor and continuation (Table 6). Mean success was 86.07% with both history inputs, 84.47% after masking predictor history and retraining, and 84.93% after removing the continuation history term at inference. We also explain the temporal picture before the equations and define the continued chunk, previously called the “carrier,” when it is introduced.

For condition dependence, Appendix F reports how the retained subspace changes within 100 held-out demonstrations per family. The mean projector distance from the initial subspace is 0.159 for SmolVLA and 0.174 for π0.5\pi_{0.5}.

Thank you again for your positive assessment and helpful suggestions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

Thank you all for the careful reviews and constructive suggestions. We have revised the paper to make the motivation and the roles of its components clearer. We also revised the abstract to describe the lightweight residual adapter and summarize the main results more directly.

The main changes are:

  1. Motivation and method. The introduction now starts from the two problems we address: correcting a frozen VLA's action proposal and connecting the corrected chunk to previously executed commands. Sections 3.2–3.4 explain the temporal basis and continuation before developing their mathematical formulations.
  2. Simpler controls and history. Section 4.4 and Table 6 add comparisons with a direct MLP residual, fixed DCT, removal of flow matching and the jerk-excess loss, and separate interventions on the two uses of history. Appendix D, Table 13 gives the per-seed counts and definitions.
  3. Scope and implementation. We clarify that jerk is measured from normalized arm commands in simulation, not physical end-effector motion (Section 4.1). Appendix B describes the backbone features, arm–gripper indexing, training objectives, and eigensolver treatment in more detail.

We respond to each review below and would be glad to clarify any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.