Publications

Drive the Thoughts: Runtime Monitoring of VLA Reasoning-Trajectory Consistency

Tian Yu, Lu Feng, and Sebastian Elbaum

Manuscript under review, 2026

Can a driving model’s reasoning tell us what it should do next? We study whether vision-language-action (VLA) models’ reasoning can serve as a specification for monitoring their planned trajectories.

A benchmark of 150 manually annotated reasoning–trajectory pairs; lane-relative monitors achieve an F1 score of 0.75 for inconsistency detection, 0.13 above the strongest baseline.

@unpublished{yu2026drivethethoughts,
  title = {Drive the Thoughts: Runtime Monitoring of {VLA} Reasoning-Trajectory Consistency},
  author = {Yu, Tian and Feng, Lu and Elbaum, Sebastian},
  year = {2026},
  note = {Manuscript under review},
  url = {https://arxiv.org/abs/2608.29583}
}