acceptodds
Under review as a conference paper at ICLR 2027

CapVector: Learning Transferable Capability Vectors in Parametric Space for Vision-Language-Action Models

Abstract

This paper proposes a novel approach to address the challenge that pretrained vision-language-action models (VLAs) often fail to effectively improve performance and reduce adaptation costs during standard supervised finetuning (SFT). Some advanced finetuning methods with auxiliary training objectives can improve performance and reduce the number of convergence steps. However, they typically incur significant computational overhead due to the additional losses from auxiliary objectives. To simultaneously achieve the enhanced capabilities of auxiliary training with the simplicity of standard SFT, we propose a new paradigm for steering the behavior of VLAs, centered around capability vectors. Capability vectors specify the beneficial properties induced by auxiliary-objective SFT and are given by element-wise arithmetic operations between the auxiliary-objective SFT models and regular SFT models. The capability vectors are then merged with pretrained parameters to form a capability-enhanced meta model. With a regular SFT integrated with a lightweight orthogonal regularization loss, the capability-enhanced meta model attains performance comparable to auxiliary finetuned baselines with reduced computational overhead and higher training efficiency. Internal and external experiments demonstrate that our capability vectors (1) are effective and versatile across diverse models, (2) can generalize to novel environments and embodiments out of the box.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-langugae-action Models, Robot Manipulation, Model Merging

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Fyw8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CapVector extracts parameter differences between auxiliary-trained and standard-SFT VLA models, then adds these differences to the pretrained model before downstream adaptation. A regularizer is introduced to preserve the injected benefits. The goal is to reuse auxiliary-training gains across tasks with lower downstream training costs.

Critical strengths
  • The paper addresses a practical problem: reducing repeated auxiliary-training overhead when adapting VLA models.
  • Experiments span multiple backbones, auxiliary objectives, simulation benchmarks, and real-world platforms, with promising cross-environment transfer results.
Critical weaknesses
  • The practical scope of checkpoint reuse is unclear. The main experiments inject vectors into the pretrained checkpoint used as their extraction origin. Their usefulness for domain-adapted or intermediate checkpoints remains untested. If extraction must be repeated for each starting checkpoint, the practical reuse benefit could be substantially reduced.

  • The paper describes it as enforcing orthogonal updates, but the formula appears to keep certain weights close to their starting values. The authors should clarify what it actually does and whether an ordinary penalty on weight changes would work similarly.

  • The authors already have an auxiliary-trained model. Why not directly fine-tune that model on the downstream task? Without this comparison, it is unclear why subtracting the standard-SFT model and adding the difference back to the pretrained model is necessary.

  • Downstream training may be cheaper, but extracting the vector requires two source-training runs. The cost assumptions used to estimate when this investment pays off do not match the appendix measurements. The paper should show the total cost at comparable success rates.

Questions for authors
  • Can the same vector be reused on an already adapted or intermediate checkpoint without re-extraction or harming its existing capabilities?
  • Does the regularizer enforce orthogonality, or mainly limit changes to certain weights? Would ordinary weight-update regularization work similarly?
  • Does CapVector outperform directly fine-tuning the auxiliary-trained model on the downstream task?
  • Including extraction costs, how many downstream tasks are needed before CapVector saves total computation at comparable performance?
Reviewer_6XLLOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes capability vectors for vision-language-action models. Two models are fine-tuned on a small extraction dataset, one with an auxiliary objective such as spatial alignment or latent reasoning and one with standard fine-tuning, and their parameter difference is added to the pretrained model to form an enhanced initialization. Downstream tasks then use only standard fine-tuning plus a regularization term that discourages overwriting the capability vector. Experiments on LIBERO, RoboTwin, and real robots show that this matches the auxiliary-objective methods at lower training cost, transfers across tasks and embodiments, and allows different capability vectors to be combined.

Critical strengths

The paper is well written and esay to follow.

The paper proposes a novel and practical idea. Subtracting a standard fine-tuned model from an auxiliary-objective one, trained on the same data, aims to cancel task-specific fitting and isolate the gain from the auxiliary objective. Turning a training-time objective into a reusable parameter prior removes its cost from every downstream run (Tab. 8: 0.3G extra FLOPs vs. 5.0T for Spatial Forcing) while retaining most of its benefit (e.g., 91.7% vs. 87.5% for Spatial Forcing at 5k steps on LIBERO).

Critical weaknesses
  1. Key results may rest on a weak baseline and test-set tuning. The reproduced OpenVLA-OFT reaches 92.7% on LIBERO, versus 97.1% reported in [R1], which equals CapVector’s result; training on one GPU with batch size 8 may explain the gap. Extracting from the most diverse set, LIBERO-90, gives almost no gain on RoboTwin (Tab. 2: 9.4% vs. 6.7%), contradicting the task-irrelevance claim. α and λ are tuned on LIBERO-Long, which is also an evaluation task, and the α ablation varies by up to 4 points.

  2. Method and efficiency claims are overstated. The “orthogonal” loss (Eq. 6) is a weighted L1 penalty on coordinates where γ is large, not an orthogonality constraint, and its computation for LoRA, where A has a different shape from γ, is unclear. The cost analysis assumes auxiliary fine-tuning costs 2× standard SFT, but Tab. 8 measures about 1.28×, which moves break-even from 3 to about 9 downstream tasks.

[R1] M. J. Kim, C. Finn, and P. Liang. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. arXiv:2502.19645, 2025.

Questions for authors
  1. Is the gap between your OpenVLA-OFT baseline (92.7%) and the 97.1% reported in [R1] due to the reduced batch size or GPU count? If possible, please compare CapVector against OpenVLA-OFT trained with its official recipe.

  2. Why does extracting from LIBERO-90 give almost no gain on RoboTwin, while LIBERO-Spatial gives a large one?

  3. Were α and λ selected on the same LIBERO-Long evaluation used to report results?

  4. In the LoRA setting, how exactly is Eq. 6 computed between the full-size γ and the low-rank A?

1 comment

Dear Reviewer 6XLL,

I hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we are eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Best regards, Authors

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

The paper proposes transferring the benefits of auxiliary-objective VLA fine-tuning through parameter differences. It constructs a capability vector by subtracting a standard-SFT checkpoint from an auxiliary-objective checkpoint trained on the same extraction data, adds this vector to the pretrained model, and adapts the resulting initialization to downstream tasks with an additional regularizer. The intended benefit is to reuse auxiliary supervision across tasks without repeatedly incurring its training overhead. Evaluation covers LIBERO, RoboTwin 2.0, three VLA backbones, two auxiliary objectives, and several real-world embodiments. The paper also investigates vector composition, preservation during adaptation, and amortized extraction cost.

Critical strengths
  1. Reusing auxiliary supervision through an initialization is a useful practical contribution. The method addresses a concrete cost in VLA adaptation: repeatedly running auxiliary models or objectives for each new task. Its strongest evidence is transfer between different simulation environments. With LIBERO-Spatial as the extraction dataset, Table 2 reports 31.8% average RoboTwin success for CapVector versus 6.7% for the reproduced OpenVLA-OFT baseline, approaching Spatial Forcing at 33.1%. This is a substantial improvement in a setting where extraction and downstream environments differ.

  2. The validation extends beyond a single backbone or benchmark. Experiments include LoRA and full fine-tuning, spatial and latent-reasoning supervision, and real robots operated by external collaborators. Table 4 also reports complementary benefits from combining the two vectors. The regularizer ablation, merging-weight sweep, and TIES-style comparison in Appendix C strengthen the empirical case that this is a reusable adaptation strategy.

Critical weaknesses
  1. The comparisons do not adequately control for additional source-data training. In Table 2, CapVector uses LIBERO data for capability extraction before RoboTwin fine-tuning, while equivalent source-data training is not reported for the standard-SFT or auxiliary-objective baselines. Baselines trained on the same DextD_{ext} before downstream adaptation are needed to distinguish the benefit of capability extraction from that of additional source-data exposure.

  2. Guidance on selecting capability-extraction data is insufficient. Table 2 shows substantially different RoboTwin success rates using LIBERO-Spatial, Long, and LIBERO-90 (31.8%, 23.7%, and 9.4%). Although Figure 3 examines visual richness, the paper does not explain these source-dependent differences or provide actionable selection criteria. This leaves practitioners uncertain about which data will yield a reusable capability vector and whether source selection requires costly downstream trial and error.

Questions for authors
  1. Data-matched comparison: Did the standard-SFT and auxiliary-objective baselines in Table 2 receive any training on the same LIBERO D_ext used by CapVector? If not, please compare against training those baselines on D_ext before RoboTwin adaptation, including direct source-checkpoint reuse. Report source-data exposure and training budgets so the effect of capability extraction can be separated from additional data and computation.

  2. End-to-end cost validation: The experiments demonstrate reuse across multiple downstream settings, but the three-task break-even point in Section 3.7 remains an estimate. Can you report cumulative training costs, including both extraction runs, alongside success rates for the downstream adaptations sharing a single capability vector? This would establish whether the reported multi-task benefits are achieved at a lower total cost than auxiliary-objective fine-tuning.

  3. Extraction-data selection: What explains the large transfer gap between LIBERO-Spatial, Long, and LIBERO-90 in Table 2? Were source-data exposure and convergence comparable? Beyond the visual-richness experiment in Figure 3, what criteria can guide DextD_{ext} selection for a new downstream task, and what validation data and selection budget are required?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.