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

2026-04-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

Vision-Language-Action (VLA) models, while effective in sequential decision-making, exhibit fragility to minor environmental changes due to trajectory overfitting. Researchers propose Perturbation learning with Delayed Feedback (PDF), a verifier-free test-time adaptation framework that enhances decision performance without requiring base model fine-tuning. PDF addresses spurious correlations through uncertainty-based data augmentation and action voting, utilizing an adaptive scheduler to balance performance and efficiency. To bolster stability, PDF incorporates a lightweight perturbation module that retrospectively adjusts action logits using delayed feedback, thereby correcting overconfidence. Experiments demonstrate PDF's consistent gains, achieving a +7.4% success rate on LIBERO and a +10.3 human normalized score on Atari, establishing a practical approach for reliable test-time adaptation in multimodal decision-making agents. The code is available on GitHub.

Key takeaway

For research scientists developing Vision-Language-Action models, you should consider integrating Perturbation learning with Delayed Feedback (PDF) to enhance model robustness and performance at test time. This framework offers a verifier-free approach to mitigate trajectory overfitting and overconfidence, leading to significant gains in task success rates without fine-tuning the base model. Implementing PDF can lead to more reliable multimodal decision-making agents.

Key insights

PDF improves VLA model robustness at test-time by mitigating overfitting and overconfidence through adaptive perturbation and delayed feedback.

Principles

• Mitigate spurious correlations via uncertainty-based augmentation.
• Retrospectively adjust action logits with delayed feedback.
• Balance performance and efficiency with adaptive scheduling.

Method

PDF uses uncertainty-based data augmentation and action voting, managed by an adaptive scheduler. It learns a lightweight perturbation module to adjust action logits retrospectively, guided by delayed feedback to correct overconfidence.

In practice

• Apply PDF to enhance VLA robustness in dynamic environments.
• Utilize PDF for improved success rates in robotic manipulation.
• Implement PDF for better performance in Atari game agents.

Topics

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

Code references

• zhoujiahuan1991/CVPR2026-PDF

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Learning, Fast and Slow: Towards LLMs That Adapt Continually — Combining slow parameter updates with fast in-context learning improves LLM adaptation and reduces forgetting.
• A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt — TTT's effectiveness hinges on spectrally matching update steps and aligning update directions with query relevance.
• Learning Perturbations to Extrapolate Your LLM — Learnable continuous perturbations in embedding space significantly improve LLM out-of-domain extrapolation.
• From Abstraction to Instantiation: Learning Behavioral Representation for Vision-Language-Action Model — BehaviorVLA improves VLA model robustness by learning temporally coherent behavioral representations for generalized manipulation across environments.
• Learning, Fast and Slow: Towards LLMs That Adapt Continually [R] — Combining slow parameter updates with fast context optimization enhances LLM adaptation and mitigates forgetting.
• A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt — Test-time training's instability is explained and guided by a decision-theoretic, Bayesian inference framework for adaptive updates.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.