acceptodds
Under review as a conference paper at ICLR 2027

Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment

Abstract

Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained representations that support visual and semantic generalization. Co-training on web image-text data, a common remedy, applies language and action losses to separate observations, leaving VLAs with language-action misalignment that standard manipulation benchmarks do not expose. We propose Anchor-Align, which augments BC with two objectives: Vision-Language Anchoring distills layer-wise representations from a frozen VLM copy to prevent this drift, while Language-Action Alignment converts each action target into a discrete motion-direction label and jointly trains language and action prediction on the same robot observation. We conduct real-world evaluations across eight manipulation settings on single-arm xArm7 and bimanual YAM robots, using two VLA architectures with regression and flow-matching action heads. Across these settings, Anchor-Align consistently improves over BC on novel targets, layouts, and motion-sensitive bimanual tasks requiring coordinated control. At scale in simulation, we demonstrate consistent improvements on OOD perturbations, perceptual robustness, and long-horizon control across LIBERO-PRO, LIBERO-Plus, and CALVIN, respectively, suggesting that preserving pretrained representations and effective action learning are not fundamentally at odds.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language-Action Models, Out-of-Distribution Generalization, Vision Language Models, Catastrophic Forgetting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_C7pjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Anchor-Align, a VLA finetuning method that augments behavior cloning with two objectives. Vision-Language Anchoring matches intermediate representations to those of a frozen pretrained VLM, while Language-Action Alignment supervises language predictions using motion-direction labels derived from robot demonstrations. Together, these objectives aim to preserve pretrained knowledge and connect it to action learning. Experiments on LIBERO-PRO, LIBERO-Plus, CALVIN, and real-world single-arm and bimanual tasks report improvements in generalization and manipulation performance across regression and flow-matching action heads, without requiring external data or additional manual annotations.

Critical strengths
  1. Simple and practical method. Both auxiliary objectives use existing robot demonstrations, requiring no additional annotations, external training data, or language generation at inference.
  2. Real-world coverage. Real-world experiments include single-arm and bimanual robots, as well as regression and flow-matching action heads, providing evidence of applicability across different configurations.
Critical weaknesses
  1. Limited novelty. Preserving VLM representations and using language supervision for action learning have both been explored widely in prior work. The main changes here are layer-wise feature distillation and supervision with coarse motion-direction labels, which appear to be relatively small extensions of existing ideas.
  2. Comparison with stronger pretrained VLAs. The paper would benefit from comparisons with models such as π0 and π0.5 on both real-world tasks and simulation OOD benchmarks. While Appendix Table 5 includes π0.5-KI on standard LIBERO, evaluating these models under the same downstream demonstrations and OOD conditions would better establish how the proposed method compares with existing pretrained VLAs in generalization.
  3. Overstated claims in the introduction. Several claims in the introduction appear broader than the evidence supports, including that prior methods are “never supervised on the same robot observation” (L94–102) and that “losses on such auxiliary data do not transfer the VLM’s semantic understanding to the finetuned VLA” (L89–90). Approaches such as π0.5 and Knowledge Insulation already combine token prediction and continuous-action learning on robot observations, including action-related language supervision. Moreover, π0.5’s web-data ablations demonstrate benefits for language following on unseen object categories, raising the possibility that the negative results reflect issues with the experimental setup or baseline implementation, rather than a general failure of co-training to transfer semantic knowledge. The paper would benefit from narrowing these claims, distinguishing observed failures from their proposed explanations, and clarifying its specific contribution beyond existing joint supervision and knowledge-preservation methods.
Questions for authors
  1. Is the method actually leveraging pretrained language semantics, or simply learning an additional action-direction classifier?
  2. See Weakness.
5 comments
AuthorsOpenReview ↗

Author Response (1/3)

W1

Limited novelty. Preserving VLM representations and using language supervision for action learning have both been explored widely in prior work. The main changes here are layer-wise feature distillation and supervision with coarse motion-direction labels, which appear to be relatively small extensions of existing ideas

Preserving pretrained VLM representations while adapting them for continuous control during VLA finetuning is largely understudied. To our knowledge, Anchor-Align is the first VLA finetuning objective to jointly preserve pretrained VLM representations and align them with action, without additional data.

The adaptation of the VLM backbone during finetuning creates a preservation–adaptation tension: finetuning on action causes catastrophic forgetting of VLM priors, whereas strongly preserving them (e.g., by freezing the VLM backbone) does not yield representations useful for control. Prior work mainly focuses on large-scale pretraining [1][2][3], showing how language supervision and preserving VLM representations through co-training can help. We show that these methods do not naturally transfer to finetuning. VLAs remain fragile under distribution shifts [6][7], and prominent approaches use additional pretraining data to improve generalization [2][8]. Here, we show that this is an effect of standard BC finetuning itself (L44–53), and co-training objectives can introduce task interference (further supported by [4][5]).

Anchor-Align addresses both within the finetuning objective, using supervision already available in robot demonstrations. Anchoring constrains representational drift while allowing both the continuous-action BC loss and the language–action alignment loss to adapt the VLM backbone (Figure 2 caption, L81–85). The implementation is simple; the gains are significant and consistent across 4 simulation benchmarks: LIBERO-PRO, LIBERO-PLUS, LIBERO, and CALVIN, and 8 real-world manipulation tasks spanning different embodiments (YAMs and XARM) and different backbones. The additional gains when finetuning π0.5 demonstrate that Anchor-Align complements large-scale pretraining (author response 2/3).

AuthorsOpenReview ↗

Author Response (2/3)

Comparison with stronger pretrained VLAs. The paper would benefit from comparisons with models such as π0\pi_0 and π0.5\pi_{0.5} on both real-world tasks and simulation OOD benchmarks.

A key distinction in this comparison is that Anchor-Align (based on StarVLA / VLA-Adapter backbone) requires no action pretraining. It directly adapts a pretrained VLM using only downstream demonstrations, whereas π0\pi_0 & π0.5\pi_{0.5} benefit from large-scale robot pretraining (L118–121). The π0.5\pi_{0.5} comparison allows us to assess how far this finetuning recipe can go compared to a strong robot-pretrained model under the same downstream training data budget. In real-world settings, we only compare with π0.5\pi_{0.5} because: 1) simulation results show that Anchor-Align is clearly better than π0\pi_0, and 2) real-world experimentation is very time-intensive and labor-intensive.

Real-world comparison. Anchor-Align's main strength comes from preserving and leveraging the VLM's pretrained representations, which are particularly useful for real-world observations due to large-scale vision-language pretraining. Our real-world results demonstrate this advantage: Anchor-Align outperforms π0.5\pi_{0.5} across 6 out of 8 tasks, particularly on challenging tasks involving Motion Understanding, Stability, and Target Generalization (OOD).

Extension of Figure 6(d) in the main paper

MethodReactive Grasp│Tomato Plate Transfer│Tomato Plate OOD Transfer
Success (%)│Placed (%)Retained (%)│Placed (%)Retained (%)
Standard BC50│56.3│00
π0.5\pi_{0.5} [new]40│7037.5│4018.75
Anchor-Align70│10060│7047.5

Extension of Figures 7, 9(b), and 10 in the main paper (StarVLA results)

MethodPink Mug Placement│Cable Insertion│Green Mug Placement│Broccoli│Clutter
Pickup (%)Placed (%)│Success (%)│Success (%)│Success (%)│Success (%)
Standard BC00│10│30│35│45
π0.5\pi_{0.5} [new]9025│25│50│60│75
Anchor-Align10040│50│60│50│70

Simulation comparison. We extend our simulation evaluation (Table 1) to include π0\pi_0 and π0.5\pi_{0.5} baselines. Anchor-Align substantially outperforms π0\pi_0 and achieves competitive performance with π0.5\pi_{0.5} in 8 out of 10 OOD categories. Anchor-Align underperforms π0.5\pi_{0.5} on Position Swap and Initial Robot State, two categories where strong action priors are particularly beneficial. This makes π0.5\pi_{0.5}'s advantage in these settings intuitive, given its extensive robot pretraining.

Importantly, Anchor-Align is a finetuning recipe that can also be applied directly to π0.5\pi_{0.5}. Applying Anchor-Align to π0.5\pi_{0.5} improves performance from 75.1% to 78.3% on LIBERO-PRO and from 93.8% to 94.4% on LIBERO-Plus. Notably, performance improves on Position Swap (32.8% → 41.2%) and Initial Robot State (79.4% → 82.9%), two challenging categories where large-scale action pretraining provides an advantage. On LIBERO-Plus, π0.5\pi_{0.5} already achieves at least 94% accuracy in five of seven categories, leaving limited room for further improvements. These results demonstrate that Anchor-Align further improves π0.5\pi_{0.5}, establishing the generalizability of our finetuning recipe across different VLA backbones and its complementarity to large-scale robot pretraining.

Extension of Table 1 in the main paper

MethodLang. Reph.Object SwapPos. SwapPRO MeanLang. Instr.Bg. Text.Robot InitCam. ViewObj. LayoutLight Cond.Sensor NoisePlus Mean
MolmoAct77.882.40.053.479.584.147.410.176.577.453.459.9
OpenVLA-OFT74.495.20.056.581.595.740.394.788.695.528.274.1
π0\pi_0 [new]92.095.40.062.583.896.137.195.591.996.287.783.6
VLA-Adapter91.189.62.361.085.190.752.692.693.293.289.585.1
Anchor-Align VLA (ours)97.096.222.671.987.299.659.196.397.499.096.990.3
π0.5\pi_{0.5} [new]95.896.832.875.196.2100.079.492.097.499.394.093.8
Anchor-Align-π0.5\pi_{0.5} (ours) [new]96.497.441.278.396.499.282.994.497.4100.092.094.4
AuthorsOpenReview ↗

Author Response (3/3)

W3

Overstated claims in the introduction.

“never supervised on the same robot observation” (L94–102)

Thank you for pointing it out. We fixed it in the revision, clarifying our contribution.

“losses on such auxiliary data do not transfer the VLM’s semantic understanding to the finetuned VLA” (L89–90)” … Moreover, π0.5\pi_{0.5}’s web-data ablations demonstrate benefits for language following on unseen object categories, raising the possibility that the negative results reflect issues with the experimental setup or baseline implementation, rather than a general failure of co-training to transfer semantic knowledge.

We think there is some confusion here; please allow us to clarify. The reviewer is correct that π0.5\pi_{0.5}’s web-data ablations demonstrate improved language following on unseen object categories through its large-scale training recipe. However, our study addresses a different setting, downstream VLA finetuning, where our experiments show that auxiliary co-training interferes with learning task-relevant action representations. Similar limitations have been independently reported by ECoT [9], ChatVLA [10], and RAFT [11]. These findings do not contradict π0.5\pi_{0.5}; they demonstrate that the effectiveness of auxiliary supervision depends on the training regime and how it is incorporated.

The following is references from our submission that discuss this hindrance and empirically report performance in such cases. In Appendix A, L815–820, we discuss why this might be the case:

We hypothesize that general-purpose web supervision can act as an auxiliary objective with limited relevance to these task details, potentially competing with task-specific adaptation. This may help explain why our Co-training + KI baseline underperforms standard BC on mean LIBERO-PRO and LIBERO-Plus success.

Our representation analysis (Appendix D.4) further shows that preserving pretrained feature structure does not necessarily yield action-informative representations: the co-trained backbone retains high representational similarity but exhibits lower action decodability than the action-trained alternatives.

Q1

Is the method actually leveraging pretrained language semantics, or simply learning an additional action-direction classifier?

During the rebuttal phase, we conducted a new experiment: we remove the frozen pretrained language head and replace it with a classifier that directly predicts the six motion directions. We use the same direction labels derived from demonstrated action chunks, so the classifier’s supervision remains aligned with action generation. We train this aligned-action classifier jointly with BC and anchoring, and its loss updates both the classifier and the backbone’s trainable LoRA parameters.

MethodPROPlus
VLA-Adapter61.085.1
Aligned-Action Classifier [new]65.587.8
Anchor-Align VLA71.990.3

Anchor-Align outperforms this classifier, supporting the benefit of pretrained language semantics beyond action-aligned direction supervision.

When we examine the VLM backbone’s understanding of motion directions (Appendix, Table 8), we find that Anchor-Align strengthens the motion semantics encoded in the backbone, with gains that transfer beyond the training distribution. We train linear probes on LIBERO-Spatial and evaluate them on both LIBERO-Spatial and the out-of-distribution LIBERO-Object suite. The probes operate on fixed backbone representations and bypass both the learned projection and the frozen language head, directly assessing directional information within the backbone.

Table 8 copied from appendix

MethodSpatial (in-dist.)Object (OOD)
Frozen VLA-Adapter backbone43.822.8
VLA-Adapter86.021.4
Anchor-Align VLA94.430.3
Reviewer_7FzXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Anchor-Align, which adds two losses on top of BC finetuning for VLAs: feature matching to the frozen pretrained VLM to keep its representations, and language supervision with six direction words derived from the demo actions. The paper is evaulated on LIBERO-PRO, LIBERO-Plus, CALVIN and real world tasks, consistently better than BC with both regression and flow matching heads.

Critical strengths
  1. The method is simple and requires no extra annotation, and the comparison to BC is on the same architecture so the gains are meaningful. The evaluation is extensive: sim benchmarks, two real robots, and two types of two action heads.
  2. Good evidence that anchoring reduces forgetting without hurting action learning, it beats the lr / LoRA / weight decay baselines and GQA results also support it.
Critical weaknesses

It's unclear whether the gain comes from language semantics. The shuffled and unrelated word controls are both worse than anchoring alone, which is strange, since they still have one-to-one mapping to directions. Maybe it is just conflicting with the anchoring loss. The obvious baseline is a normal 6-way direction classifier on top of the same features, trained with BC + anchoring. Without this I don’t know if the language part matters at all. The OOD probe gain is also small (28.5 -> 30.3), no error bars. 2. Some claims are too broad. co-training + KI mixes several things together, so it failing doesn’t mean co-training doesn’t work. Real world motion tasks only compare full method vs BC, no ablation, and only 20 trials with one policy each.

Questions for authors

Same as above. And how is the bimanual case handled with one direction label?

4 comments
AuthorsOpenReview ↗

Author Response (1/2)

Q.1

How sensitive is co-training + KI to the data mixture and loss weights? Are there comparisons separating co-training from knowledge insulation?

Yes, we evaluated co-training without KI at two loss weights, 1.0 and 0.3, and compared it with Co-training + KI. For the data mixture, we fixed the web-data-to-action-data ratio at 1:3, following ChatVLA’s [3] co-training recipe.

Co-training Ablation Table

MethodLIBERO-PRO (%)LIBERO-Plus (%)
Co-training with weight 1.0 [new]31.746.6
Co-training with weight 0.3 [new]35.562.6
Co-training + KI48.572.4

Reducing the co-training loss weight from 1.0 to 0.3 improves performance on both benchmarks, while Co-training + KI outperforms both co-training-only settings. These comparisons evaluate loss-weight sensitivity and compare co-training with and without knowledge insulation; the data-mixture ratio was held fixed.

Please also clarify which external-baseline results were reproduced and which were taken from existing sources.

Implementation details:

  • MolmoAct, VLA-0, and OpenVLA-OFT were evaluated using their existing LIBERO-trained checkpoints. Their Table 1 results come from our own evaluations on LIBERO-PRO and LIBERO-Plus.
  • VLA-A [Frozen], LAP, and MAPS in Table 1 were reproduced using their officially released code and evaluated under the same setup.
  • Co-training + KI was implemented on VLA-Adapter using ChatVLA’s co-training recipe combined with Knowledge Insulation, as described above.

W1

The OOD probe improvement over anchoring alone is also small, from 28.5% to 30.3%, with no uncertainty reported. Without this comparison, the claimed role of pretrained language semantics remains unclear.

During rebuttal, we added Align VLA to Table 8, which provides a better, more complete picture. Takeaway: Anchoring and alignment independently improve OOD direction understanding and interact constructively when combined. Anchoring raises Object accuracy from 21.4% to 28.5%, while alignment alone achieves 90.4% on Spatial and 28.7% on Object. Combining them improves both further to 94.4% and 30.3%, outperforming either component on both probes. This supports jointly preserving VLM representations and adapting them for action prediction. Because the probes bypass both the learned projection and frozen language head, these improvements reflect directional information encoded in the backbone itself.

MethodSpatial (in-dist.)Object (OOD)
Frozen VLA-Adapter backbone43.822.8
VLA-Adapter86.021.4
Anchor-only VLA70.728.5
Align VLA [new]90.428.7
Anchor-Align VLA94.430.3

W2

The co-training + KI baseline combines web supervision, action-token supervision, and blocked continuous-action gradients under one data mixture. Its underperformance does not establish which part causes the problem or whether co-training itself is ineffective.

We hypothesize that adapting the VLM backbone for action prediction is important during finetuning. In Co-training + KI, the backbone is updated through web and action-token supervision, while gradients from the continuous-action expert are blocked. This setup can introduce task interference [1][2], whereby auxiliary objectives hinder effective learning of the primary action task. Such interference is well established in multi-task learning [1][2]. Similar effects have been observed in VLAs. ChatVLA [3] reports that adding visual-text data to robot training reduces real-world task success, attributing this degradation to partial task interference between control and understanding. Our co-training ablations (Q.1) further show that co-training without KI, evaluated at loss weights of 1.0 and 0.3, performs worse than Co-training + KI. Thus, the observed underperformance is not specific to knowledge insulation; adding KI improves performance in the tested setup. Many other works report similar observations [4][5]. These results support the explanation that task interference from auxiliary web data accounts for the underperformance of the tested co-training configurations during finetuning.

AuthorsOpenReview ↗

Author Response (2/2)

Q2

In Appendix E, is the diagnostic question evaluated through the original language head or through the learned projection as well?

We are evaluating Anchor-Align in Appendix E, which consists of both the learned projection and the original frozen language head. The model is trained on LIBERO expert demonstrations, whereas the diagnostic is evaluated on LIBERO-PRO rollouts. As stated in Table 9, “Language-action alignment on LIBERO-PRO rollouts,” this evaluation covers object swap, position swap, and language rephrase, matching the evaluation settings in Table 1. The diagnostic therefore evaluates alignment on OOD rollout observations, not on the demonstrations used for training.

Is the question adapted to each task, and what fraction of rollout frames is retained for each method?

Yes, we use the same question template across tasks and replace the task clause with the instruction associated with the evaluated rollout. We retain a rollout frame only when the action displacement satisfies ∥dt∥2≥0.001\lVert d_t\rVert_2 \geq 0.001, and its dominant direction agrees with those of the next five frames (App. E.1). These criteria remove near-stationary and transient motions. This evaluation is conducted on 1,500 rollouts per method, totaling about 15K frames (average of 10 frames per episode).

Q3

Are anchoring-only and alignment-only results available for the bimanual tasks?

Real-world evaluation is time-consuming and labor-intensive, and for that reason we have kept our ablation analysis to simulation results.

Also, how are separate arm predictions implemented when the two arms have different direction labels, given the single pre-action representation and projection described in Appendix B.1?

App. B.1 describes the single-arm case. For the bimanual YAM, we use two arm-specific projections WprojL,WprojR∈Rd×d\mathbf{W}^{L}_{\text{proj}}, \mathbf{W}^{R}_{\text{proj}} \in \mathbb{R}^{d \times d}. Both read the same pre-action hidden state hpre\mathbf{h}^{\text{pre}}, and each produces vocabulary logits through the same frozen language head: a^alang=WlmWprojahpre\hat{\mathbf{a}}^{\text{lang}}_{a} = \mathbf{W}_{\text{lm}}\mathbf{W}^{a}_{\text{proj}}\mathbf{h}^{\text{pre}} for a∈L,Ra \in \\{L, R\\}. Each arm has its own direction label. We average that arm's displacement over the 8-step action chunk. An arm's label is masked when its mean displacement is below 0.002 m, so a chunk where only one arm moves supervises only that arm. The alignment loss is the cross-entropy over all valid (arm, sample) labels. Since hpre\mathbf{h}^{\text{pre}} is 896-dimensional, the two linear readouts can extract different directions from it. We will add this formulation to the App. B.1 and B.4.

Q4

Have you compared against a conventional six-way direction classifier trained jointly with BC and anchoring, using the same backbone representation and comparable tuning?

During the rebuttal phase, we conducted a new experiment: we remove the frozen pretrained language head and replace it with a classifier that directly predicts the six motion directions. We use the same direction labels derived from demonstrated action chunks, so the classifier’s supervision remains aligned with action generation. We train this aligned-action classifier jointly with BC and anchoring, and its loss updates both the classifier and the backbone’s trainable LoRA parameters.

MethodKey TechniquePROPlus
VLA-AdapterStandard BC61.085.1
ShuffleControl61.484.9
ScatterControl63.385.7
Aligned-Action Classifier [new]Control65.587.8
Anchor-Align VLAFull method71.990.3

The Aligned-Action Classifier outperforms Shuffle and Scatter, but Anchor-Align exceeds it by 6.4% on LIBERO-PRO and 2.5% on LIBERO-Plus. The classifier already uses the same action-derived direction labels and actively adapts the backbone alongside BC and anchoring. These gains therefore demonstrate the benefit of semantic alignment beyond auxiliary classification.

Why do the shuffled and unrelated-word controls perform worse than anchoring alone?

We interpret this as task interference from semantically incompatible auxiliary supervision. Anchoring encourages preservation of the pretrained representations, while these auxiliary objectives require learning language associations that are inconsistent with the demonstrated motion. Preserving the representations does not prevent an additional loss from interfering with their adaptation for control.

This is precisely what makes our alignment novel and relevant: the VLM backbone learns to predict task-relevant information, i.e., the direction of the demonstrated motion, through language supervision that directly supports action generation. It is this constructive interaction between anchoring and alignment that makes Anchor-Align so effective.

AuthorsOpenReview ↗

References

  1. Tianhe Yu et al. Gradient Surgery for Multi-Task Learning. NeurIPS, 2020.

  2. Lucio M. Dery et al. Auxiliary Task Update Decomposition: The Good, The Bad and The Neutral. ICLR, 2021.

  3. Zhongyi Zhou et al. ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model. EMNLP, 2025.

  4. Michał Zawalski et al., Robotic Control via Embodied Chain-of-Thought Reasoning, Conference on Robot Learning (CoRL), 2024.

  5. Shresth Grover et al., RAFT: Reasoning-Aligned Fine-Tuning Improves Robustness of VLA Post-Training, arXiv preprint arXiv:2611.04368, 2026.

AuthorsOpenReview ↗

Summary of New Results and Request for Your Feedback

We thank the reviewer for highlighting our method’s simplicity, the effectiveness of anchoring, and strong empirical results.

We note that the review was updated today, which consolidates several of the original points. So, below we summarize our new results and indicate where the corresponding concerns are addressed in our earlier responses.

  • We added an aligned-action classifier as a new control. Anchor-Align outperforms it on both benchmarks, supporting the benefit of semantic alignment beyond auxiliary classification (Author Response 2/2, Q4).

  • Our additional linear-probe study shows that anchoring and alignment each substantially improve OOD direction understanding in the backbone. Their combination performs best, demonstrating their constructive interaction in Anchor-Align (Author Response 1/2, W1).

  • We evaluated co-training without KI at different loss weights using a fixed data mixture. Co-training + KI outperforms both co-training-only settings, showing that the observed underperformance is due to co-training hurting action learning (Author Response 1/2, Q.1 and W2).

  • Bimanual supervision uses arm-specific projections and direction labels (Author Response 2/2, Q3).

Looking forward to hearing your feedback on our responses. If any concerns remain, we are happy to run targeted additional experiments during the remainder of the day.

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

Vision-Language-Action (VLA) models lack semantic generalization despite leveraging a Vision-Language Model (VLM) backbone trained on an internet-scale dataset. The authors propose two auxiliary losses, denoted the Anchor-Align loss, to 1) anchor the model to the VLM prior and 2) align actions with the language prior. Two types of pretrained VLAs, flow-matching-based and regression-based, are finetuned with the behavior cloning (BC) loss along with the Anchor-Align loss.

The evaluation includes three simulated environments (LIBERO, LIBERO-Plus, and CALVIN) and real-world experiments on 8 tasks, demonstrating that policies trained with the Anchor-Align loss generalize better to out-of-distribution settings, suggesting that the proposed method leverages the generalizability of the VLM.

Critical strengths

This is a well-executed paper. The experiments include a sufficient number of benchmarks and baselines. The results speak for themselves, supporting the effectiveness of the proposed method. Moreover, the ablation study on the language-action alignment method justifies the particular design choice of the loss.

The language-action alignment loss targets a subtle failure mode of other baseline methods for preventing catastrophic forgetting: the misalignment of language and action supervision. The realization of the method appears novel.

Critical weaknesses

In Table 1, 5 out of 7 baselines perform worse than the base model VLA-Adapter by a significant margin, despite being modifications intended to prevent catastrophic forgetting. This raises a question about baseline choice or execution. A brief discussion of the shortcomings of the other baselines would be helpful for understanding the results.

Questions for authors
  • Could you explain why some baselines underperforms compared to VLA-Adapter?
  • Could you clarify which dataset is used for co-training VLA?
2 comments
AuthorsOpenReview ↗

Author Response

In Table 1, 5 out of 7 baselines perform worse than the base model VLA-Adapter by a significant margin, despite being modifications intended to prevent catastrophic forgetting. This raises a question about baseline choice or execution. A brief discussion of the shortcomings of the other baselines would be helpful for understanding the results.

Could you explain why some baselines underperform compared to VLA-Adapter?

Table 1 includes both independently developed VLAs (such as MolmoAct, VLA-0, VLA-Adapter, and OpenVLA-OFT) and methods applied to VLA-Adapter (such as VLA-A [Frozen], LAP, Co-training + KI, MAPS). MolmoAct, VLA-0, and OpenVLA-OFT are evaluated using their official LIBERO-trained checkpoints. Their performance on LIBERO-PRO and LIBERO-Plus highlights the difficulty of generalization in these models under perturbations, as evaluated in Table 1. We train VLA-A [Frozen], LAP, and MAPS using officially available code. Co-training + KI is our reimplementation of co-training with Knowledge Insulation on VLA-Adapter (Sec. 4.1 and Appendix A). LAP and MAPS perform competitively with VLA-Adapter on LIBERO-PLUS and consistently outperform it on LIBERO-PRO.

We attribute the underperformance of VLA-A [Frozen] and Co-training + KI to insufficient adaptation of the VLM backbone for continuous-action prediction during finetuning. VLA-A [Frozen] prevents backbone adaptation by keeping the VLM backbone of VLA frozen, whereas Co-training + KI blocks the continuous-action loss from directly updating the backbone. These results highlight the importance of adapting pretrained representations for downstream control while preserving their generalization capabilities. Anchor-Align jointly preserves pretrained VLM representations and aligns them with demonstrated motion, while allowing the continuous-action loss to adapt the backbone. Prior studies report similar lower downstream performance when the backbone is frozen [1, 2].

Could you clarify which dataset is used for co-training VLA?

We follow ChatVLA’s [3] co-training strategy for combining general-purpose vision-language data with robot-action data, motivated by its extensive analysis and strong empirical results. Specifically, we use 25K VQA samples from VSR, GQA, and COCO and adopt ChatVLA’s 1:3 sampling ratio between web and robot-action data (Appendix A & B.3).

[1] Moo Jin Kim et al., OpenVLA: An Open-Source Vision-Language-Action Model. CoRL, 2024.

[2] Yihao Wang et al., VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model. arXiv preprint arXiv:2509.09372, 2025.

[3] Zhongyi Zhou et al. ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model. EMNLP, 2025.

AuthorsOpenReview ↗

Summary and Request for Your Feedback

We thank the reviewer for highlighting our method’s simplicity, extensive experimental evaluation, strong empirical results, and the novelty of language-action alignment. Below, we summarize our clarifications addressing both questions.

  • We clarified the baseline implementations and explained that VLA-A [Frozen] and Co-training + KI underperform due to insufficient backbone adaptation for continuous-action learning.

  • We clarified that co-training follows ChatVLA’s recipe, using VQA samples from VSR, GQA, and COCO alongside robot-action data at a 1:3 sampling ratio.

We believe these clarifications address both questions raised in your review. We look forward to hearing your feedback and are happy to provide additional evidence or run targeted experiments during the remaining day.

Comments on the paper

2 comments
Area_Chair_M2K6OpenReview ↗

Please Review the Author Rebuttal and Reassess Your Rating

Dear Reviewers,

Thank you for your time and effort in reviewing the submissions.

The authors have now submitted their rebuttals. I kindly ask you to carefully review the authors' responses and assess whether they have adequately addressed your questions and concerns.

If the rebuttal resolves your concerns, please consider updating your review and adjusting your rating accordingly. If any concerns remain, please feel free to raise them in the discussion.

Your timely feedback and active participation are greatly appreciated, as they are essential to reaching a fair, well-informed decision.

Thank you again for your valuable contributions!

Best regards,
Area Chair

AuthorsOpenReview ↗

Updated manuscript with additional rebuttal experiments and request for feedback

Dear AC,

Thank you for the reminder and for encouraging the discussion.

Dear Reviewers,

During the rebuttal period, we have added experiments and clarifications that we believe address the concerns raised in your reviews and further strengthen our work. We have also updated the manuscript, with changes highlighted in blue.

We would appreciate your feedback on whether our responses and additional experiments resolve your concerns. If any questions remain, we are happy to provide further clarification or additional evidence during the remaining discussion period.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.