Depth-Wise Probing and Pruning of the Planning Token in a Driving Vision-Language-Action Model

2026-08-07

AI Summary

This work investigates the information evolution of planning tokens across decoding layers in vision-language-action (VLA) models for driving and assesses their necessity for trajectory generation. By freezing the native planner as a trajectory-space probe and combining linear probing with trajectory compatibility evaluation, the study reveals that semantic intent can be decoded with high accuracy from shallow layers—achieving 97.7% accuracy in the first layer—while trajectory compatibility progressively improves with network depth. Leveraging angular deviation of planning tokens, the authors propose a layer importance ranking method enabling efficient pruning: removing 8 out of 32 layers incurs only ~5% relative error, yields a 1.33× speedup in inference, and preserves performance without significant degradation.

Abstract

Vision-language-action (VLA) models route driving decisions through a deep language model, but it is unclear how much of that depth the action itself requires. We study a representative driving VLA whose entire plan is carried by a single planning token that a generative planner decodes into a trajectory. Borrowing the planner as a trajectory-space logit lens, we decode the planning token from every one of the 32 decoder layers and measure two signals: the linear decodability of the navigation command and trajectory compatibility with the frozen native planner. Our diagnostic shows that semantic intent is linearly decodable early: command-probe accuracy reaches 97.7\% after the first decoder layer, compared with 16.7\% chance. In contrast, compatibility with the frozen native planner improves gradually across depth, with open-loop Avg-L2 reaching its minimum of 2.11\,m only at the final layer. Learned readouts from the first layer recover much of this gap, indicating that planning information is already present early but is not yet represented in the format expected by the deployed planner. Ranking decoder layers by the angular deviation they induce in the planning token permits removal of 8 of 32 layers within an approximately 5\% relative open-loop error increase and yields a measured 1.33$\times$ decoder speedup. At the evaluated sample size, no family-specific degradation is statistically resolved. These findings are limited to the evaluated ORION checkpoint and Bench2Drive setup.

Problem

• vision-language-action model
• planning token
• depth-wise probing
• model pruning
• trajectory compatibility

Innovation

• depth-wise probing
• planning token
• trajectory compatibility
• decoder pruning
• vision-language-action model