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

TL;DR

This work addresses the precision gap in Vision-Language-Action models by introducing MG-Select, a verifier-free test-time scaling framework that leverages a condition-masking reference distribution and KL-divergence-based confidence to select among multiple sampled actions. A joint imitation-learning strategy with condition masking further improves the quality of the reference distribution and the robustness of action selection. Empirically, MG-Select yields substantial gains across simulation and real-world robotic benchmarks (e.g., RoboCasa, SIMPLER-WidowX, LIBERO) and reduces inference latency via efficient deployment, demonstrating a practical path to high-precision, robust VLA policies without external verifiers.

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.

Figures (4)

• Figure 1: Overview of MG-Select. (1) Autoregressive VLA $\pi_\theta$ samples action tokens in parallel from the predicted distribution, while simultaneously computing token-wise KL divergence from the condition-masking distribution to the predicted distribution. (2) Best-of-N selection is then performed using an action confidence score $C_{\tilde{a}}$ obtained by aggregating these token-wise scores.
• Figure 2: Qualitative results of MG-Select in real-world pick-and-place tasks. We visualize one of our real-world experiments in the "Box to Bowl" task: (a) grasping an object from the box and (b) releasing it into the bowl. The rollout shows that MG-Select can generate high-precision actions at critical moments for task success, whereas the base policy ($\pi_0$-FAST-DROID) often struggles at these steps.
• Figure 3: Real-world performance on out-of-distribution tasks with Franka Research 3. We report the average success rate (%) over 16 trials for each task.
• Figure 3: Inference latency comparison on LIBERO-Object liu2023libero. We compare vanilla MG-Select with its efficient deployment variant using single prefill, based on $\pi_{0}$-FAST pertsch2025fast.