acceptodds
Withdrawn from ICLR 2027

PrimingVLA: Self-Supervised Action Priming for Parameter-Efficient Vision-Language-Action Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Adapting Vision–Language–Action (VLA) models requires learning task-specific control through an action expert while preserving the multimodal capabilities of a pretrained vision–language model (VLM). Existing methods primarily align multimodal representations with action generation, but overlook the need to establish structured action priors before task-conditioned learning. To address this issue, we introduce PrimingVLA, a parameter-efficient framework built on our proposed new principle of action priming before multimodal conditioning. Its Self-Supervised Action Priming (SAP) learns temporal action structure through masked trajectory reconstruction, after which multimodal inputs are introduced for joint adaptation. A cross-prompt calibration coordinates their adaptation by bounding relative action-prompt updates using VLM-prompt updates as a reference. Experiments on LIBERO, LIBERO-PLUS, and LIBERO-PRO demonstrate effective task adaptation and characterize robustness under distribution shifts, with ablations supporting the benefits of staged learning and coordinated updates. Real-world experiments on a Piper robotic arm further demonstrate its effectiveness in robotic manipulation. Neural Tangent Kernel monitoring throughout training provides direct evidence that the sensitivity gap between the two pathways progressively narrows. Remarkably, PrimingVLA requires updating fewer than 0.05M parameters ( of the full model) while keeping both pretrained backbones frozen, enabling efficient adaptation and rapid task switching.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language-Action Models, Robotic Manipulation, Parameter-Efficient Fine-Tuning, Dual-Prompt Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ReXuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PrimingVLA adapts a frozen VLA model by training only a few parameters across VLM and action prompts. It first primes action prompts through masked action reconstruction, then jointly adapts both prompt banks using observation-conditioned flow matching. Cross-Prompt Calibration constrains action-prompt updates relative to VLM-prompt updates. Various experiments show the effectiveness of PrimingVLA.

Critical strengths
  • The topic, achieving effective task performance with very few trainable parameters, is quite important for VLA policy fine-tuning.
  • The training design is intuitive and easy to understand.
  • Various experiments have been performed to assess the effectiveness of PrimingVLA.
Critical weaknesses
  • The connection between staged learning and the challenges of adapting a frozen VLA with very few prompt parameters is insufficiently explained.
  • CPC lacks sufficient justification and comparison with simpler alternatives, such as using a lower learning rate for action prompts.
  • Bridge and Adaptation use the same inputs, objective, trainable parameters, and CPC mechanism, but differ mainly in optimizer settings. The necessity of separating them, particularly resetting the optimizer, is neither clearly motivated nor directly ablated.
  • The real-world evaluation lacks quantitative success rates and comparisons with baseline methods.
  • Beyond the concerns regarding the motivation and evaluation, the paper lacks polish in both its writing and figures. For example, Figure 4 is poorly cropped and contains truncated content, reducing readability and overall presentation quality.

Overall, the paper appears insufficiently mature in its current form. The proposed training recipe is not adequately motivated, and the paper does not clearly explain why its components address the challenges of parameter-efficient VLA adaptation. Together with the insufficient real-world evaluation and noticeable issues in writing and figure presentation, I believe that the paper does not currently meet the standard expected for acceptance at ICLR.

Questions for authors

See weakness above.

Reviewer_3ZndOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PrimingVLA, a parameter-efficient framework for VLA adaptation that learns action structure before multimodal conditioning. It combines Self-Supervised Action Priming (SAP) through masked trajectory reconstruction with Cross-Prompt Calibration (CPC) to coordinate prompt updates. Experiments on LIBERO benchmarks and real-world robotic tasks demonstrate promising adaptation performance with fewer than 0.05M trainable parameters.

Critical strengths

1.The method achieves a favorable balance between parameter efficiency and task adaptation performance. 2.The staged training strategy and cross-prompt calibration show promising empirical effectiveness.

Critical weaknesses
  1. The technical novelty and underlying mechanism are not sufficiently established.
  2. Experimental comparisons and real-world validation are insufficient to support the stronger performance and generalization claims.
Questions for authors
  1. How can the authors distinguish the benefits of SAP from those of additional training or better prompt initialization? In particular, does masked action reconstruction outperform an action-only flow-matching warm-up under exactly matched training budgets?
  2. Can the authors provide comparisons with π0.5-based PEFT baselines under identical pretrained checkpoints, demonstration data, training budgets, and evaluation protocols? How consistent are the reported gains across different random seeds?
  3. What are the quantitative success rates of PrimingVLA on the four real-robot tasks over repeated trials? How does it compare with direct dual-prompt adaptation without SAP and CPC under the same demonstrations and evaluation conditions?
Reviewer_FYvVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PrimingVLA, a method for adapting a pretrained vision-language-action model using very few trainable parameters. The main idea is to learn action patterns before introducing visual and language inputs. First, action prompts are trained through masked action reconstruction. Then, both vision-language prompts and action prompts are jointly trained for observation-conditioned control. Cross-Prompt Calibration (CPC) limits action-prompt updates relative to vision-language-prompt updates to coordinate learning. PrimingVLA achieves 89% success on LIBERO. The paper also evaluates robustness on LIBERO-PLUS and LIBERO-PRO and presents qualitative experiments on a Piper robotic arm.

Critical strengths
  • Very few trainable parameters. The task-specific prompts are small and easy to store or switch.
  • Simple design. It reuses the existing model and demonstrations without an extra teacher network.
Critical weaknesses
  • The full method uses 4,000 training updates, while the direct-training baseline uses 3,000. The improvement could partly come from extra training rather than the proposed training order.

  • It is unclear whether CPC works better than simply using a lower learning rate for the action prompts.

  • The baseline comparisons are not fully controlled with different models, training budgets, and evaluation protocols.

  • Real-robot success rates and actual training time and memory costs are not reported.

Questions for authors

Check weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.