A driving system learns how strongly reasoning should steer its plans
How Much Should a VLA Trust Its Reasoning? Step-Level Guidance as a Measured Action-Value Function · arXiv:2610.40002
In plain words
A vision-language-action system looks at a scene, writes a reasoning trace, then uses it to make a driving plan. This paper introduces a way to turn the trace’s influence up or down at each step, and a controller that learns those settings by scoring completed plans during training.
In two simulator examples, the controller reduced the trace’s influence when the original plan followed an unavailable lane change or went too far.
Across 358 held-out simulator scenes, its third stage also lowered a measure of plan error against recorded futures by 16.5% compared with the system’s default setting.
The research
A driving system can produce a trace that misreads the scene, but checking whether the trace is correct still leaves a harder question: how much should that trace steer the plan? The paper also shows that a correct-looking trace can produce a plan that runs far beyond the recorded drive.
CoC-CFG makes trace influence adjustable at each planning step, and Q-adaptive guidance learns weights from measured plan scores.
During training, the controller scores plans made after trying different weights at individual steps, then learns from score differences against a reference weight.
On held-out Alpasim scenes, its third stage reduced plan displacement cost against recorded futures; in separate simulated rollouts, it passed more scenes than the two reported fixed-weight settings.
The Waymo comparisons also favored an adaptive stage on held-out frames. Together, these results support the approach in the reported driving evaluations.
The evidence covers Alpamayo 1.5 driving evaluations in Alpasim and Waymo, with rule-based conversions for Waymo. It does not establish how the approach performs with other policies or in other task domains.
What the evidence shows
- Correctness is not enough
A vision-language-action system first writes a reasoning trace and then uses it to generate a driving plan.
The key question is not only whether the trace matches the scene, but how strongly it should steer the plan: a wrong trace can point to an unavailable lane change, and a correct-looking one can still lead the plan too far from the recorded drive.
- Control trace influence
CoC-CFG, the paper’s method for adjusting trace influence, compares the action expert’s output with the trace present and with only the trace removed. It combines those outputs at every plan-generation step.
Weight 1 leaves the original behavior unchanged; lower weights reduce the trace’s influence, and higher weights amplify it. This creates a direct way to control the trace’s effect on the plan.
- Learn from scored branches
Q-adaptive guidance learns a weight for each step from training scenes. The researchers try candidate weights at a saved generation state, finish each plan using the current controller, and score it against the known recorded future.
They compare each score with the reference weight’s score, then train critics to predict those measured differences. This gives the controller evidence about individual step choices.
- Two Alpasim evaluations
In Alpasim, the third controller stage reduced open-loop displacement cost by 16.5% against the default weight on 358 held-out scenes. That measure scores plans against logged futures without replanning.
In a separate closed-loop simulator test, where the system drives through reconstructed scenes, it passed 262 of 300 scenes, versus 252 for the default and 254 for the training-scene anchor weight.
- Evidence on Waymo
The researchers also tested Waymo driving, where a score averages eleven categories comparing plans with trajectories rated by people. On 358 held-out validation frames, the second controller stage scored above the tested fixed weights and other listed methods.
Its hidden-test score also exceeded the best published comparison cited by the paper, though that comparison does not establish identical systems or evaluation procedures.
Deep
- Original paper authors
- Dabin Seo, Eunhyek Joa
- Submitted to arXiv
- License
- arXiv non-exclusive