Test-Time Perturbation Learning with Delayed Feedback for Vision-Language-Action Models

2026-04-20

AI Summary

Existing Vision-Language-Action (VLA) models often exhibit brittleness under minor environmental shifts due to trajectory overfitting. This work proposes the PDF framework, which achieves validator-free test-time adaptation for the first time by integrating uncertainty-driven data augmentation, action voting, and a lightweight perturbation module. A novel delayed feedback mechanism is introduced to recalibrate action confidence, complemented by an adaptive augmentation scheduler to mitigate spurious correlations in multimodal agents. Evaluated on the LIBERO benchmark, the approach improves task success rates by 7.4% and achieves a 10.3% gain in human-normalized scores on Atari environments, significantly outperforming both the original VLA model and current test-time adaptation methods.

Technology Category

Computer Vision: Large Vision ModelsMultiagent Systems: Adversarial AgentsHumans and AI: Human-Aware Planning and Behavior Prediction

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingResponsible Web: Machine-in-the-loop, human agency and autonomyUser Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Vision-Language-Action models (VLAs) achieve remarkable performance in sequential decision-making but remain fragile to subtle environmental shifts, such as small changes in object pose. We attribute this brittleness to trajectory overfitting, where VLAs over-attend to the spurious correlation between actions and entities, then reproduce memorized action patterns. We propose Perturbation learning with Delayed Feedback (PDF), a verifier-free test-time adaptation framework that improves decision performance without fine-tuning the base model. PDF mitigates the spurious correlation through uncertainty-based data augmentation and action voting, while an adaptive scheduler allocates augmentation budgets to balance performance and efficiency. To further improve stability, PDF learns a lightweight perturbation module that retrospectively adjusts action logits guided by delayed feedback, correcting overconfidence issue. Experiments on LIBERO (+7.4\% success rate) and Atari (+10.3 human normalized score) demonstrate consistent gains of PDF in task success over vanilla VLA and VLA with test-time adaptation, establishing a practical path toward reliable test-time adaptation in multimodal decision-making agents. The code is available at \href{https://github.com/zhoujiahuan1991/CVPR2026-PDF}{https://github.com/zhoujiahuan1991/CVPR2026-PDF}.

Problem

• Vision-Language-Action models
• trajectory overfitting
• spurious correlation
• environmental shifts
• test-time adaptation

Innovation

• Test-Time Adaptation
• Vision-Language-Action Models
• Delayed Feedback
• Uncertainty-Based Augmentation
• Perturbation Learning