Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models

The paper introduces SVA (Search, Value, and Act), which uses Monte Carlo tree search in simulation to explore the output distribution of a frozen vision-language-action policy and trains a lightweight Q-value model from empirical returns. At deployment, the frozen VLA proposes multiple actions and the evaluator selects the candidate with the highest uncertainty-regularized value, without simulator access. A diagnostic pass@k study reports success increasing from 33% at pass@1 to 92% at pass@32. The abstract also reports that a 9B VLA outperformed a 27B model by 7 points with 27% lower inference latency in the stated evaluation setting.