Q-adaptive guidance improves reasoning vision-language-action driving
Original title: How Much Should a VLA Trust Its Reasoning? Step-Level Guidance as a Measured Action-Value Function
Frozen Alpamayo 1.5 with the method scores 8.135159 on the hidden Waymo set, above 8.079.
CoC-CFG applies classifier-free guidance so a reasoning trace can be weakened or strengthened at each Euler step, with w=1 recovering the original policy. Q-adaptive guidance treats the weight as a measured action value, fits small critics, and updates the controller conservatively. In the Alpasim simulator it beats fixed weights and SAC and GRPO, passing 262 of 300 closed-loop scenes versus 252 for the original policy. On the hidden Waymo end-to-end test, frozen Alpamayo 1.5 with rule-based conversions and this guidance is reported as a new published-method state of the art.
Unconfirmed: The state-of-the-art score is the authors' comparison among published methods.
- Open-loop Alpasim results beat every fixed weight plus SAC and GRPO.
- Closed loop: 262 of 300 scenes passed, versus 252 for the original policy.
- Held-out Waymo stages beat four adaptive-guidance methods, SAC, and GRPO.
- Hidden-test rater feedback score is 8.135159 versus 8.079 for the prior best.
- Organizations
- Waymo
- AI models
- Alpamayo 1.5
From the original article
1 min read at arxiv.orgReasoning vision-language-action (VLA) policies write a reasoning trace and then decode the action from it. A wrong trace yields a wrong action, and even a correct-looking trace can mislead the plan. Existing remedies ask whether the reasoning is right.
By Dabin Seo · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.