mlsanityData to 14 Oct 2026

Flow-Support-Weighted Proposals for Test-Time Verified VLA Policies

arXiv 2610.14588 ↗First posted 10 Oct 2026cs.RO cs.CVcited by 0

On the map: Vision-language-action models → Reinforcement learning and post-training also fits: Efficient VLA inference and adaptation

Abstract (arXiv)

Test-time sampling promises to turn one action prediction into a set of distinct executable candidates that a verifier can choose among. Downstream fine-tuning works against that promise. A vision-language-action model trained to solve one task distribution concentrates its flow-matching decoder around the dominant demonstrated action, so repeated samples differ only marginally and the verifier receives many copies of one decision. We call this proposal contraction and identify it as a key constraint on test-time scaling: repeated samples can remain highly redundant as the candidate budget grows. Action Dynamic Sampling (ADS) alleviates the constraint on the proposal side, during training. At each training step, ADS integrates K additional trajectories from the current policy under fresh base noise, scores each by the flow support along its path with a Hutchinson trace estimator, and corrects the policy's local action distribution toward this weighted set through a detached finite-set KL term. Training cost rises; deployment cost is unchanged. Across simulation and four physical-robot tasks, ADS raises proposal spread without the success penalty that Gaussian perturbation incurs, lifts oracle best-of-8 success on RoboCasa from 57.6% to 63.0%, improves fixed verifiers by 2.4 to 4.9 points in simulation, and raises TokenPB by up to 13.0 points on real-world manipulation.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Reinforcement learning and post-training, one of 13 directions in Vision-language-action models. Builds on 28 of its 33 references that are in the map. First posted on arXiv 10 Oct 2026.

Builds on

28 of 33 references.

Closest ancestors · 2

Same direction, newest first

  1. 2026TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action Models2 Jun
  2. 2025EVOLVE-VLA: Test-Time Training from Environment Feedback for Vision-Language-Action Models16 Dec

Other references · 26

  1. 2022Flow Matching for Generative Modelingcited by 5,702
  2. 2018Neural Ordinary Differential Equationscited by 3,779
  3. 2023Diffusion Policy: Visuomotor Policy Learning via Action Diffusioncited by 2,838
  4. 2024π0: A Vision-Language-Action Flow Model for General Robot Controlcited by 2,407
  5. 2023LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learningcited by 1,410
  6. 2023Improving and generalizing flow-based generative models with minibatch optimal transportcited by 879
  7. 2018FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Modelscited by 594
  8. 2024RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robotscited by 407
  9. 2025LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Modelscited by 303
  10. 2025HybridVLA: Collaborative Diffusion and Autoregression in a Unified Vision-Language-Action Modelcited by 181
  11. 2026StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developingcited by 112
  12. 2017Learning Sampling Distributions for Robot Motion Planningcited by 82
14 more
  1. 2025Sampling-Efficient Test-Time Scaling: Self-Estimating the Best-of-N Sampling in Early Decodingcited by 63
  2. 2025RoboMonkey: Scaling Test-Time Sampling and Verification for Vision-Language-Action Modelscited by 61
  3. 2025Verifier-free Test-Time Sampling for Vision-Language-Action Modelscited by 23
  4. 2025Sample, Don't Search: Rethinking Test-Time Alignment for Language Modelscited by 22
  5. 2026Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignmentcited by 21
  6. 2025RoVer: Robot Reward Model as Test-Time Verifier for Vision-Language-Action Modelcited by 19
  7. 2026LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Modelscited by 11
  8. 2026Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervisioncited by 8
  9. 2026Shifting Uncertainty to Critical Moments: Towards Reliable Uncertainty Quantification for VLA Modelcited by 7
  10. 2026Pre-VLA: Preemptive Runtime Verification for Reliable Vision-Language-Action and World-Model Rolloutscited by 5
  11. 2025EVE: A Generator-Verifier System for Generative Policiescited by 3
  12. 2026VeriSpace: Spatially Grounded Action Verification for Vision-Language-Action Modelscited by 3
  13. 2026Unleashing More Actions via Action Compositional Training for VLA Modelscited by 3
  14. 2026MAGMA-GEN: Validated Recovery Supervision from Ambiguous Failures via Counterfactual Re-Executioncited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies
  2. 2026On-the-Fly VLA Adaptation via Test-Time Reinforcement Learning
  3. 2026DiVeR: Decision-Critical Verifier Learning for VLA Test-Time Scaling
  4. 2026Verify, Act, Distill: Turning Test-Time Compute into Test-Time Adaptation for Vision-Language-Action Models
  5. 2026Reuse Before You Retrieve: Diagnosing Headroom and Complementarity for Test-Time Augmentation of Embodied Multimodal Policies
  6. 2026Accelerating VLA Inference with Perturbation Informed Execution
  7. 2026Looking Back to Move Forward: Temporal Verification for Generative Robot Policies
  8. 2025Steering Vision-Language-Action Models as Anti-Exploration: A Test-Time Scaling Approach