Verifier-free Test-Time Sampling for Vision Language Action Models

2025-10-07

AI Summary

Vision-Language-Action (VLA) models for high-precision robotic control are constrained by single-shot inference, limiting simultaneous accuracy and generalization. This paper proposes MG-Select, a test-time sampling framework that requires no external verifier and involves no additional training. Its core innovation lies in generating a reference action distribution under maximal uncertainty—induced via stochastic state masking and language conditioning—and using KL divergence to quantify candidate actions’ consistency with this distribution as a confidence score. A joint training strategy—incorporating both conditional and unconditional action distribution modeling alongside dropout regularization—enhances intrinsic distribution calibration. Experiments demonstrate that MG-Select improves performance by 28% on in-distribution tasks and 35% on out-of-distribution tasks. Moreover, on the RoboCasa pick-and-place benchmark with only 30 demonstrations, it achieves a 168% relative performance gain.

Technology Category

• Intelligent Robots: Learning & Optimization for ROB
• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Language and Vision

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Vision-Language-Action models (VLAs) have demonstrated remarkable performance in robot control. However, they remain fundamentally limited in tasks that require high precision due to their single-inference paradigm. While test-time scaling approaches using external verifiers have shown promise, they require additional training and fail to generalize to unseen conditions. We propose Masking Distribution Guided Selection (MG-Select), a novel test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules. Our approach utilizes KL divergence from a reference action token distribution as a confidence metric for selecting the optimal action from multiple candidates. We introduce a reference distribution generated by the same VLA but with randomly masked states and language conditions as inputs, ensuring maximum uncertainty while remaining aligned with the target task distribution. Additionally, we propose a joint training strategy that enables the model to learn both conditional and unconditional distributions by applying dropout to state and language conditions, thereby further improving the quality of the reference distribution. Our experiments demonstrate that MG-Select achieves significant performance improvements, including a 28%/35% improvement in real-world in-distribution/out-of-distribution tasks, along with a 168% relative gain on RoboCasa pick-and-place tasks trained with 30 demonstrations.

Problem

• VLAs lack precision in single-inference robot control tasks
• Test-time scaling requires external verifiers with limited generalization
• Proposes training-free framework using internal model properties for action selection

Innovation

• MG-Select uses KL divergence for action selection
• Reference distribution generated via masked VLA inputs
• Joint training with dropout improves distribution quality