acceptodds
Under review as a conference paper at ICLR 2027

Teachers Are Not Equal: Task-Flow Conditional Cross-Layer Knowledge Distillation for Efficient Flow-Based VLA Models

70 views

Abstract

Vision-Language-Action (VLA) models based on flow matching are expensive to deploy due to repeated Transformer computation at inference. One solution is to distill knowledge from a teacher VLA to a lightweight student VLA. Existing methods select a fixed teacher layer for distillation, but this may be suboptimal because layer informativeness varies with task instructions and flow stages, limiting student capacity. To address this, we introduce Task-Flow Conditional cross-layer Knowledge Distillation (TFC-KD), a framework that adaptively selects teacher layers conditioned on task instructions and flow stages. For each retained student layer, we first identify several candidate teacher layers around its initially aligned counterpart. We then use task instructions to modulate which teacher layers provide stronger supervision. Crucially, we divide the continuous flow process into three stages, allowing the preferred teacher layers to shift as action generation progresses. Task instructions and flow stages are combined to assign weights to candidate teacher layers, whose weighted attention maps over vision-language tokens serve as distillation targets. Finally, we impose an ordering constraint to keep teacher-layer selections roughly aligned with student depth while permitting local adaptation across task instructions and flow stages. On LIBERO, the trained 8-layer student achieves success, outperforming the Shallow- baseline by . When further compressed to 6 layers, TFC-KD incurs only a success drop relative to the teacher (), while reducing deployed parameters by and FLOPs by , with a speedup of inference on one A800 GPU.

Primary area:
foundation or frontier models, including LLMs
Keywords:
vision-language-action models, knowledge distillation, model compression, flow-based policy

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cLXbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper proposes a task-based hierarchical distillation method for VLA models. Instead of always learning from fixed teacher layers, the student learns from a weighted combination of teacher layers selected according to the task instruction and action-generation stage. The selection mechanism is removed after training. The eight-layer student achieves 96.72% success on LIBERO, while the six-layer version achieves 2.30× faster inference than the teacher with a small performance drop.

Critical strengths

Strengths: This paper proposes a hierarchical distillation method that transfers knowledge from a teacher model to a more lightweight student model across different tasks.

Critical weaknesses

Weaknesses

  1. The motivation for hierarchical distillation is not very convincing. The paper argues that the student model should learn from specific teacher layers depending on the task. However, this raises several questions. First, the student is initialized using only a subset of the teacher’s layers, which means its downstream capabilities could differ substantially due to the missing layers. Second, the idea that certain blocks represent particular tasks is not sufficiently justified. Establishing a clear correspondence between individual layers and specific tasks remains a significant challenge in deep learning, and relying on gradient-based evidence alone is insufficient.
  2. No real-robot demo: The evaluation relies on simulation, with the CALVIN experiment limited to five tasks, each reset independently. Real-robot experiments is missing. It is also recommending to include the the full long-horizon benchmark.
  3. Comparisons with existing adaptive distillation methods are limited. One could argue that a simple distillation method might achieve similar results under a fair experimental setup. Please specify how a fair experimental setup is ensured.
  4. The design is not well justified with clear ablation study, and this issue is related to the first weakness.
Questions for authors
  1. What evidence shows that specific teacher layers provide better supervision for particular tasks, especially after removing layers changes the student’s representations? Is there supporting evidence beyond gradient compatibility?

  2. Can you provide real-robot experiments and evaluate on the full CALVIN long-horizon benchmark?

  3. How does the method compare with simpler and existing adaptive distillation methods, and how are fair comparisons ensured?

  4. Can you provide further ablations to justify the teacher-layer neighborhood and flow-stage design choices?

11 comments
AuthorsOpenReview ↗

Response to Reviewer cLXb (1)

Thank you for assessing our motivation, comparison fairness, ablation design, and evaluation scope. We agree that the original evidence does not establish a one-to-one correspondence between individual teacher Transformer layers and specific semantic tasks. Below, additional controlled experiments clarify our claims, with responses organized by W1–W4 and Q1–Q4.

[W1 & Q1: Motivation for Condition-Dependent Teacher-Layer Supervision]

refers to: "The paper argues that the student model should learn from specific teacher layers depending on the task. However, this raises several questions. First, the student is initialized using only a subset of the teacher’s layers, which means its downstream capabilities could differ substantially due to the missing layers. Second, the idea that certain blocks represent particular tasks is not sufficiently justified." & "What evidence shows that specific teacher layers provide better supervision for particular tasks, especially after removing layers changes the student’s representations? Is there supporting evidence beyond gradient compatibility?"

Thank you for raising these questions. We apologize that our wording suggested a fixed correspondence between teacher layers and tasks. Our method does not assume that individual Transformer blocks represent or specialize in particular tasks. Instead, we investigate whether the relative suitability of teacher attention representations for supervising a compressed student varies with the task instruction and flow stage.

Our closest baseline, Shallow-π, uses attention supervision from a fixed intermediate teacher layer. Such a fixed target does not adapt its source to the task or the stage of iterative action generation. To introduce this flexibility, TFC-KD assigns soft weights to attention maps from nearby teacher layers for each retained student layer. These weights depend on the task instruction and flow stage and are learned during distillation. The resulting weighted attention target allows the contributions of candidate teacher layers to vary across conditions without assigning a particular task to a particular layer.

Regarding initialization, the parameters copied from selected teacher layers provide the initial student weights; they do not impose a permanent teacher–student correspondence. We agree that removing layers changes the student representations, and we do not assume that an initialized student layer remains equivalent to its teacher counterpart. During distillation, the student parameters and routing weights are updated, while the frozen teacher continues to provide attention targets. All compared L8 students share the same architecture and initialization from teacher layers [0, 2, 5, 7, 10, 12, 15, 17], so this starting point is held consistent across methods.

Finally, gradient compatibility is an auxiliary post-training diagnostic, not a criterion used to train the router or assign teacher layers to tasks. It measures the cosine similarity between the student gradients induced by attention distillation and the action objective, assessing their local, first-order optimization compatibility. The routing weights are learned through the attention-distillation objective and ordering regularization, without gradient-compatibility labels.

AuthorsOpenReview ↗

Response to Reviewer cLXb (2)

[W2 & Q2: Real-Robot Experiments and CALVIN Evaluation Scope]

refers to: "No real-robot demo: The evaluation relies on simulation, with the CALVIN experiment limited to five tasks, each reset independently. Real-robot experiments is missing. It is also recommending to include the the full long-horizon benchmark."

We have added closed-loop experiments on the Cobot Magic platform, an implementation of the ALOHA (A Low-cost Open-source Hardware System for Bimanual Teleoperation) architecture, to assess real-world task completion by dense π0.5\pi_{0.5}, Shallow-π\pi L8, and TFC-KD L8 under the same evaluation protocol.

Real-robot experimental setup. We select Stack Cups, Wipe Table, and Sort Objects, involving precise object interaction, continuous trajectory execution, and multi-object manipulation, respectively. We collect 50 expert demonstrations per task, totaling 150, and jointly train one policy on all three tasks. Both eight-layer students use the same demonstrations and evaluation rules and are initialized from the same dense teacher trained on these real-robot tasks. Each model undergoes 20 independent trials per task: 60 per model and 180 across all three models. Evaluation randomizes initial object positions, target positions, and relevant layouts. All models share pregenerated scene configurations, run in interleaved order, and use predefined success criteria.

Real-robot results:

ModelStack CupsWipe TableSort ObjectsAverage Success Rate
Dense π0.5\pi_{0.5}75%(15/20)75%(15/20)65%(13/20)71.67%(43/60)
Shallow-π\pi L870%(14/20)65%(13/20)55%(11/20)63.33%(38/60)
TFC-KD L875%(15/20)75%(15/20)70%(14/20)73.33%(44/60)

TFC-KD outperforms Shallow-π\pi at the same depth on all three tasks, increasing average success from 63.33% to 73.33% (+10.00 percentage points), comparable to the dense teacher (71.67%). These results support better preservation of manipulation capabilities on the tested real-robot tasks with a fixed eight-layer deployment architecture and no additional inference-time routing computation, providing further real-world validation of the simulation results.

CALVIN Evaluation Scope and Long-Horizon Evaluation Status. As suggested, we have initiated evaluation on the official CALVIN benchmark of five consecutive language instructions. The full evaluation requires substantial computation and simulation resources and remains incomplete under current resource limits; we therefore report no final metrics at this stage. The original CALVIN-5 experiments use a matched subset of five class-balanced atomic tasks with independent resets. At maximum interaction budgets of 240/160/120 steps, Shallow-π\pi L8 achieves 93.47%/86.13%/84.13% success, and TFC-KD L8 achieves 99.87%/97.87%/96.80%. These results provide atomic-task validation in another simulation environment. The ongoing official evaluation will assess consecutive multi-task execution. We will include the real-robot setup and results in the revision and clarify the scope of CALVIN-5 and the status of the official long-horizon evaluation.

AuthorsOpenReview ↗

Response to Reviewer cLXb (3)

[W3 & Q3: Comparisons with Existing Methods and Experimental Fairness]

refers to: "Comparisons with existing adaptive distillation methods are limited. One could argue that a simple distillation method might achieve similar results under a fair experimental setup. Please specify how a fair experimental setup is ensured."

Thank you for this suggestion. We added the comparisons below. All methods share the frozen dense π0.5\pi_{0.5} teacher, L8 student architecture, depth-aligned initialization, and LIBERO 40-task training data. The added SemCKD, C2KD, and TED implementations use 30K student optimizer updates with an effective batch size of 64 and retain the Flow Matching and teacher Velocity KD objectives. All results use training seed 42 and the same evaluation protocol: 50 episodes per task for each of three environment seeds (7, 17, and 27), totaling 6,000 episodes per model.

Distillation mechanisms. Fixed Depth-Aligned All-Layer KD supervises every student layer with the attention map of its fixed, depth-aligned teacher layer, without candidate-layer mixing. Generic Router + Monotonicity uses pooled contextual vision-language features and continuous flow time as inputs to an MLP that predicts mixing weights over candidate teacher attention maps. The mixed maps supervise student attention, using the same temperature schedule and teacher-layer ordering regularizer as TFC-KD. SemCKD uses cross-sample feature relations to learn sample-dependent soft associations between student and teacher layers and applies weighted intermediate-feature matching. C2KD uses task-related feature filters and teacher–student feature similarity to select supervision targets for the last two student layers. TED uses task-aware feature filters followed by fixed depth-aligned feature matching. During student distillation, teacher filters remain frozen, while student filters remain trainable. These implementations preserve distinct intermediate-supervision mechanisms within the shared VLA compression setting.

MethodSpatialObjectGoalLongAvg. (%)
Shallow-π\pi L897.4796.5395.5391.0795.15
Fixed Depth-Aligned All-Layer KD98.1397.0796.3391.4795.75
Generic Router + Monotonicity98.6097.4796.2790.3395.67
SemCKD92.4095.1384.5367.0084.77
C2KD88.1395.6088.8072.0086.13
TED95.0798.1393.6086.3393.28
TFC-KD L898.7398.0097.2792.9396.73

TFC-KD achieves the highest average success rate of 96.73%, exceeding Fixed Depth-Aligned All-Layer KD and Generic Router + Monotonicity by 0.98% and 1.06%, respectively. These comparisons indicate that broader supervision or generic routing with an ordering constraint alone does not explain its advantage. TFC-KD achieves an average success rate of 96.73%, compared with 84.77% for SemCKD, 86.13% for C2KD, and 93.28% for TED. Together, these comparisons support the effectiveness of TFC-KD for flow-based VLA compression. We will include the results, method-specific adaptation details, and training budgets in the revised manuscript.

Reviewer_S3nROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TFC-KD, a knowledge distillation framework designed to compress flow-based VLA models. It trains a shallow student using attention targets combined from neighboring teacher layers, with a router dynamically adjusting the mixture based on task instructions and flow stages. An ordering regularization term further encourages consistent layer alignment across the student network. Both the teacher and router are discarded after training, avoiding additional inference overhead. Experiments demonstrate higher success rates than comparable Shallow-pi baselines while maintaining similar deployment costs.

Critical strengths
  • TFC-KD achieves effective model compression without introducing additional inference overhead.
  • The improvements generalize across different backbones, compression levels, and perturbation scenarios.
Critical weaknesses
  • The existing comparisons do not clearly distinguish the benefits of adaptive target selection from those of more extensive supervision. While TFC-KD provides supervision at every retained layer, it lacks a baseline using fixed depth-aligned targets across all layers. Moreover, the generic-router baseline omits the ordering regularizer included in the complete method.

  • The router is optimized to reduce attention discrepancies rather than directly improve action prediction. Table 4 provides inconclusive evidence that its gradient compatibility is superior to Task-only, leaving the mechanism through which the selected attention targets benefit policy learning insufficiently explained.

Questions for authors
  • Could the authors include a baseline with fixed depth-aligned supervision across all student layers? Applying the same ordering regularization to the generic router would also help isolate the contribution of adaptive conditioning.

  • How does performance change when task or flow-stage conditioning is shuffled under identical training conditions? Such experiments would help determine whether the conditioning signals are responsible for the observed improvements.

  • Could the authors clarify the precise flow-bin boundaries, temperature scheduling, and loss coefficients? Additional details on the generic router's input features and architecture would also improve reproducibility.

7 comments
AuthorsOpenReview ↗

Response to Reviewer S3nR (1)

Thank you for recognizing the deployment efficiency of TFC-KD and its generalization across backbones and perturbations. We have added controlled experiments to distinguish the contributions of Task-Flow conditioning, supervision coverage, and layer-ordering regularization. We use W1, W2, and Q1–Q3 to link our responses to your comments and questions.

[W1 & Q1: Contribution of All-Layer Supervision]

refers to: "While TFC-KD provides supervision at every retained layer, it lacks a baseline using fixed depth-aligned targets across all layers." & "Could the authors include a baseline with fixed depth-aligned supervision across all student layers?"

We have added Fixed Depth-Aligned All-Layer KD, which supervises each student layer with the attention map of its initialization-source teacher layer. For 18→8-layer compression, the fixed teacher-layer indices are [0, 2, 5, 7, 10, 12, 15, 17]. This baseline supervises all 8 student layers without adaptive candidate mixing or Task-Flow routing.

MethodSpatialObjectGoalLongLIBERO Average Success Rate (%)
Shallow-π L897.4796.5395.5391.0795.15
Fixed Depth-Aligned All-Layer KD98.1397.0796.3391.4795.75
TFC-KD L898.7398.0097.2792.9396.73

Fixed all-layer supervision improves over Shallow-π by 0.60 percentage points. TFC-KD further improves over this stronger baseline by 0.98 percentage points, indicating that broader supervision alone does not fully explain its advantage. With the same all-layer supervision coverage, this result further supports the effectiveness of adaptive teacher-target construction in TFC-KD.

[W2 & Q2: Controlling for Monotonicity Regularization]

refers to: "Moreover, the generic-router baseline omits the ordering regularizer included in the complete method." & "Applying the same ordering regularization to the generic router would also help isolate the contribution of adaptive conditioning."

We have added Generic Router + Monotonicity, using the same penalty on expected teacher-layer ordering as TFC-KD. All four variants share a temperature schedule that decreases linearly from 3.0 to 0.5 over the first 10K optimizer updates and remains at 0.5 thereafter. The monotonicity coefficient is 0 without regularization and 0.05 with regularization, yielding the following 2×2 comparison.

Routing MechanismWithout Ordering RegularizationWith Ordering Regularization
Generic Router95.4895.67
Task-Flow routing96.2796.73

Under the same temperature schedule, Task-Flow routing outperforms Generic Router by 0.79% without monotonicity regularization and 1.06% with it. These results support the effectiveness of Task-Flow routing and show that monotonicity regularization alone does not explain the performance gains.

[Q3: Condition-Shuffling Experiments]

refers to: "How does performance change when task or flow-stage conditioning is shuffled under identical training conditions? Such experiments would help determine whether the conditioning signals are responsible for the observed improvements."

We added condition-shuffling experiments that independently permute router task representations or flow-stage labels within each training batch, while preserving the correct instructions, flow times, and training targets for the policy and teacher. All four variants omit monotonicity regularization; the reference is therefore Task-Flow routing without this regularizer (96.27%), rather than full TFC-KD (96.73%).

Router ConditioningSpatialObjectGoalLongLIBERO Average Success Rate (%)
Correct task + correct flow stage98.6797.5396.8792.0096.27
Shuffled task + correct flow stage98.3397.9395.7388.9395.23
Correct task + shuffled flow stage98.4097.5395.0789.1395.03
Shuffled task and flow stage98.0097.4793.5383.8093.20

Relative to the unshuffled reference, task shuffling, flow-stage shuffling, and joint shuffling reduce success by 1.04%, 1.24%, and 3.07%, respectively. Because policy inputs and action targets remain correct, these results support the importance of matching routing conditions to samples for closed-loop policy performance. We will include these experiments in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer S3nR (2)

[W3: Attention Supervision and Action-Gradient Compatibility]

refers to: "The router is optimized to reduce attention discrepancies rather than directly improve action prediction. Table 4 provides inconclusive evidence that its gradient compatibility is superior to Task-only, leaving the mechanism through which the selected attention targets benefit policy learning insufficiently explained."

Thank you for raising this concern. We apologize for the lack of clarity regarding the purpose of Table 4 in the original manuscript. We would like to clarify that Table 4 is not intended to demonstrate that reducing attention discrepancies directly improves action prediction performance. Instead, it analyzes the contributions of different components from an optimization perspective by comparing gradient compatibility across ablation variants. The results show that TFC-KD achieves higher gradient compatibility than Task-only, Time-only and Static Soft Mix. The contributions of these components to actual policy performance are evaluated through the closed-loop ablation experiments in Table 2. Therefore, Table 4 should be interpreted as complementary mechanistic evidence supporting the ablation analysis, rather than direct proof of a causal relationship between attention supervision and action prediction performance. We will clarify the purpose and scope of this analysis in the revised manuscript to avoid potential misunderstandings.

[Q4: Routing Configuration and Reproducibility]

refers to: "Could the authors clarify the precise flow-bin boundaries, temperature scheduling, and loss coefficients? Additional details on the generic router's input features and architecture would also improve reproducibility."

Thank you for this suggestion. The configurations are as follows.

  • Flow-bin boundaries

    • Generic Router: No discretization; continuous flow time t is supplied directly.
    • TFC-KD: Three bins: [0, 1/3), [1/3, 2/3), and [2/3, 1].
  • Temperature scheduling

    • Generic Router: The temperature decreases linearly from 3.0 to 0.5 over the first 10K optimizer updates and remains at 0.5 thereafter.
    • TFC-KD: The same schedule.
  • Loss coefficients

    • Generic Router: Velocity KD λ₁=1.0, attention KD λ₂=1.0, and monotonicity λ₃=0; the Generic Router + Monotonicity variant uses λ₃=0.05.
    • TFC-KD: λ₁=1.0, λ₂=1.0, and λ₃=0.05.
  • Generic Router input features and architecture

    • Input: A stop-gradient pooled representation of the contextual vision-language prefix, concatenated with continuous flow time (2049 dimensions).
    • Architecture: A 2049 → 64 → 40 MLP with SiLU activation. The outputs provide routing logits for 8 student layers with 5 candidate slots each; out-of-range teacher candidates are masked.

Sensitivity to the Number of Flow Stages. We fix the teacher-layer candidate radius at r=2, allowing each student layer to receive supervision from its depth-aligned teacher layer and up to two neighboring layers on either side (at most five candidates), and vary only the number of uniformly partitioned flow stages B.

Flow Stages BStage BoundariesSpatialObjectGoalLongLIBERO Average Success Rate (%)
1None (all t share one stage)98.4097.6795.8790.7395.67
21/298.5397.9396.8792.2096.38
31/3, 2/398.7398.0097.2792.9396.73
41/4, 1/2, 3/498.7397.9397.2792.7396.67

Average success increases from 95.67% with one stage to 96.73% with three stages, while four stages achieve a similar result of 96.67%. These results support three stages as a simple and effective discretization in the evaluated setting. Here, B=1 retains task conditioning and monotonicity regularization, whereas the original Task-only ablation omits the latter; the two settings are therefore not equivalent.

We will include the complete configuration and sensitivity results in the revised manuscript.

Thank you again for your constructive feedback. We hope these responses address your concerns and warrant reconsideration of your score. We welcome any remaining questions.

AuthorsOpenReview ↗

A Friendly Reminder to Review Our Responses

Dear Reviewer S3nR,

Thank you again for your careful evaluation and constructive feedback. We have organized our responses under [W1 & Q1], [W2 & Q2], [Q3], [W3], and [Q4] to address your concerns individually.

In [W1 & Q1], we add Fixed Depth-Aligned All-Layer KD to control for supervision coverage. In [W2 & Q2], we compare Generic Router and Task-Flow routing with and without monotonicity regularization under the same temperature schedule. In [Q3], we report condition-shuffling experiments that alter only the router inputs while preserving correct policy inputs and training targets. In [W3], we clarify the intended scope of the gradient-compatibility analysis. In [Q4], we provide the routing configurations and a sensitivity study on the number of flow stages. The supporting results and explanations are included directly in our responses.

When convenient, we would be grateful if you could review our responses and let us know whether any concerns remain unclear or insufficiently addressed. If you find these clarifications and additional evidence satisfactory, we would sincerely appreciate your consideration of updating your assessment.

Thank you again for your time and thoughtful suggestions.

Best regards,

The Authors

Reviewer_76KrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

We introduce Task-Flow Conditional Cross-layer Knowledge Distillation (TFC-KD), a framework that improves the efficiency of Vision-Language-Action (VLA) models by adaptively selecting teacher layers based on task instructions and stages of action generation. Instead of relying on fixed layer correspondences, TFC-KD dynamically weights candidate teacher layers and distills their attention maps while preserving alignment with student network depth. On the LIBERO benchmark, an 8-layer student achieves a 96.72% success rate, outperforming the Shallow-π baseline by 1.57 percentage points. A compressed 6-layer student retains 95.52% success while reducing parameters by 44.37%, FLOPs by 58.07%, and inference time by achieving a 2.30× speedup on an A800 GPU.

Critical strengths
  • The method seems to be novel for the application area of VLA as far as I can tell.
  • The direct comparison shows improved performance of the resulting model compared to the reference model.
  • Good ablations, although some things need to be looked at in the appendix.
Critical weaknesses
  • Comparison to other methods: The SOTA evaluation is very limited; especially, baseline comparisons to other layer-specific KD methods like TED, SemCKD, and C2KD are missing. While the paper states that "These methods do not condition teacher-layer selection on the flow stage of iterative action generation in flow-based VLAs.", I don't see why they can not be adapted to this scenario (at least for benchmarking, not claiming top performance). Maybe there is something special about the VLA method, but I don't see it right now. In this case, it would be good if the authors could elaborate.
Questions for authors

Overall, the architecture could be interesting. Student-teacher architectures are well studied, but I'm not sure how far this topic has been recognized in the context of VLA. The comparison with the state-of-the-art (?) seems a bit incomplete in this respect, but the overall reported numbers show that the method is able to reach better performance compared to the reference method. So, just based on the evaluation, I'm a bit on the fence if the method provides a sufficient contribution to justify publication and would usually give it a clear borderline rating. I'll give the paper a weak reject for now, but if the authors can clarify how this is superior to other layer KD works and/or if reviewers from the VLA community consider this sufficiently interesting, I'm happy to upgrade my rating.

Post-rebuttal comment: I increased my score from 2 to 3 as the points have been resolved.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 76Kr

Thank you for your positive assessment of the novelty of TFC-KD. To address the limited cross-layer KD comparisons, we have added closed-loop evaluations of SemCKD, C2KD, and TED in the flow-based VLA framework.

[W1 & Q1: Comparisons with Existing Cross-Layer Distillation Methods]

refers to: "Comparison to other methods: The SOTA evaluation is very limited; especially, baseline comparisons to other layer-specific KD methods like TED, SemCKD, and C2KD are missing."

Thank you for this suggestion. In principle, SemCKD, C2KD, and TED can all be applied to flow-based VLAs. We evaluated all three methods on the four LIBERO task suites using a common training and evaluation protocol. Results follow.

MethodSpatialObjectGoalLongLIBERO Avg. (%)
Shallow-π\pi L897.4796.5395.5391.0795.15
SemCKD92.4095.1384.5367.0084.77
C2KD88.1395.6088.8072.0086.13
TED95.0798.1393.6086.3393.28
TFC-KD L898.7398.0097.2792.9396.73

Under the shared π0.5\pi_{0.5} L8 distillation and LIBERO closed-loop evaluation framework, SemCKD, C2KD, and TED achieve average success rates of 84.77%, 86.13%, and 93.28%, respectively. TFC-KD achieves 96.73%, exceeding them by 11.96%, 10.60%, and 3.45%, respectively. TFC-KD performs better than SemCKD and C2KD on all four suites and better than TED on Spatial, Goal, and Long. These results support the effectiveness of Task-Flow conditional supervision.

Thank you again for your constructive feedback. We hope this response addresses your concerns and that you will consider raising your score. We will promptly address any remaining issues you identify.

AuthorsOpenReview ↗

A Friendly Reminder to Review Our Responses

Dear Reviewer 76Kr,

Thank you again for your careful evaluation and constructive feedback. In [W1 & Q1: Comparisons with Existing Cross-Layer Distillation Methods], we address your concern about the missing comparisons with SemCKD, C2KD, and TED.

We evaluate all three methods on the four LIBERO task suites under a common training and evaluation protocol. Their average success rates are 84.77%, 86.13%, and 93.28%, respectively, compared with 96.73% for TFC-KD. The suite-level results and analysis are included directly in our response to help assess the contribution of Task-Flow conditional supervision in the evaluated VLA setting.

When convenient, we would be grateful if you could review our response and let us know whether any concerns remain unclear or insufficiently addressed. If the additional evidence addresses your concerns, we would sincerely appreciate your consideration of updating your assessment.

Thank you again for your time and thoughtful suggestions.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up on the Additional Distillation Comparisons

Dear Reviewer 76Kr,

Thank you again for your careful evaluation and constructive feedback. As the discussion deadline approaches, we would like to follow up on our response to your concern about the missing comparisons with existing cross-layer distillation methods.

In [W1 & Q1], we provide comparisons with SemCKD, C2KD, and TED under a common VLA training and evaluation protocol. Their average LIBERO success rates are 84.77%, 86.13%, and 93.28%, respectively, compared with 96.73% for TFC-KD. Our response includes the suite-level results and analysis, which we also incorporate into the revised manuscript.

You indicated in your review that you would be open to updating your assessment if the additional evidence clarified our contribution relative to existing methods. When convenient, we would appreciate your feedback on whether these comparisons address your concern and whether any questions remain. If you find the evidence satisfactory, we would be grateful for your consideration of updating your assessment.

Thank you again for your time and thoughtful consideration.

Best regards,

The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Friendly Reminder for Feedback

Dear Reviewer 76Kr and cLXb,

Thanks for your attention to this message. This is a friendly reminder that the author-reviewer discussion period is ending. We understand you are busy, but we would appreciate any additional feedback or suggestions you might have on our work. Your insights are valuable to us. Thank you.

Sincerely,

Authors of Submission 39407

AuthorsOpenReview ↗

Friendly Reminder for Feedback

Dear Reviewers 76Kr, S3nR, and cLXb,

We sincerely thank Reviewer S3nR for their constructive feedback, recognition of our work, and decision to raise their score after considering our responses.

As the author–reviewer discussion period is ending, we would like to kindly follow up with Reviewers 76Kr and cLXb, from whom we have not yet received responses. We understand you are busy, but we would appreciate any additional feedback or suggestions you might have on our responses and revised manuscript. Your insights are valuable to us. Thank you.

Sincerely,

Authors of Submission 39407

AuthorsOpenReview ↗

Manuscript Revision Update

Dear Reviewers,

Thank you again for your constructive feedback and careful reassessment. We have incorporated the additional experiments and clarifications into the revised manuscript. For ease of review, key revisions are highlighted by reviewer: blue for 76Kr, green for S3nR, and red for cLXb. Revisions related to comments shared by multiple reviewers are highlighted in purple.

The revisions are summarized below:

  • Motivation and method clarification

    • Introduction and Section 4: Clarified the motivation for conditional teacher-layer supervision and its relationship to student initialization and local teacher-layer candidates. (cLXb)
  • Distillation baseline comparisons

    • Section 5.3: Added comparisons with SemCKD, C2KD, and TED, alongside fixed all-layer supervision and regularized generic routing. (76Kr, S3nR, cLXb)
    • Appendix J: Documented the baseline settings, loss coefficients, and auxiliary pretraining. (76Kr, S3nR, cLXb)
  • Controlled ablations and design sensitivity

    • Section 5.4: Added controlled ablations, condition-shuffling experiments, and sensitivity analysis of the candidate radius and number of flow stages. (S3nR, cLXb)
    • Appendix E: Expanded the suite-wise ablation results and comparisons across multiple training seeds. (S3nR)
    • Appendix K: Detailed the condition-shuffling and design-sensitivity protocols. (S3nR, cLXb)
  • Optimization and routing configurations

    • Appendix B: Specified the optimization settings, router architectures, temperature schedule, and regularization coefficients. (S3nR)
  • Cross-benchmark and real-robot evaluation

    • Section 5.6: Added real-robot results and clarified the scope of the CALVIN-5 atomic-task evaluation. (cLXb)
  • Mechanism analysis and attention visualizations

    • Section 6: Clarified the interpretation of the gradient diagnostics and added teacher-attention similarity analysis. (S3nR, cLXb)
    • Appendix L: Presented the attention-analysis protocol, quantitative results, confidence intervals, similarity heatmaps, and attention-map visualizations. (cLXb)

Thank you for helping us strengthen the experimental evidence and improve the clarity of the manuscript.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.