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

TL;DR - An arXiv analysis of a driving vision-language-action model (ORION on Bench2Drive) that probes the single "planning token" at each of 32 decoder layers, showing semantic intent emerges almost immediately while planner-compatible formatting accrues slowly with depth. It matters because it exposes redundant decoder depth in VLA stacks and enables ~1.33× inference speedup with modest accuracy loss.

• Using the generative planner as a trajectory-space "logit lens," navigation-command probe accuracy hits 97.7% after just the first decoder layer (vs. 16.7% chance), indicating intent is linearly decodable early.
• Compatibility with the frozen native planner improves only gradually, with open-loop Avg-L2 bottoming out at 2.11 m at the final layer; learned readouts from layer 1 recover much of that gap, suggesting a representation-format mismatch rather than missing information.
• Ranking layers by the angular deviation they induce in the planning token allows pruning 8 of 32 layers for ~5% relative open-loop error increase and a measured 1.33× decoder speedup.
• Scope is explicitly limited: results come from one ORION checkpoint and the Bench2Drive setup, and no family-specific degradation was statistically resolved at the evaluated sample size.