# CalVLA: Shaping Native Confidence in Vision-Language-Action Policies with Outcome Supervision

DevFeed: [CalVLA: Shaping Native Confidence in Vision-Language-Action Policies with Outcome Supervision](<https://devfeed.tech/articles/calvla-shaping-native-confidence-in-vision-language-action-policies-with-outcome-supervision-115023.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.00394>)

Author: Boniface Chukwuemeka Godwin

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Calibrated uncertainty](<https://devfeed.tech/topics/calibrated-uncertainty.md>), [lora](<https://devfeed.tech/topics/lora.md>), [Robotics](<https://devfeed.tech/topics/robotics.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [post-training](<https://devfeed.tech/topics/post-training.md>)

Tags: [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [discrimination](<https://devfeed.tech/tags/discrimination.md>), [failure](<https://devfeed.tech/tags/failure.md>), [policies](<https://devfeed.tech/tags/policies.md>), [safety](<https://devfeed.tech/tags/safety.md>), [supervision](<https://devfeed.tech/tags/supervision.md>), [vision-language-action](<https://devfeed.tech/tags/vision-language-action.md>)

## AI overview

CalVLA improves failure discrimination by adapting OpenVLA-7B's native action-token confidence from successful and failed rollouts, using low-rank adapters and an Action-Confidence Alignment squared margin-ranking loss without an auxiliary failure detector. On a held-out split of 500 LIBERO-10 configurations, adapted confidence reaches a joint AUROC of 0.7392±0.0480 across three seeds; for the same confidence summary, AUROC increases from 0.4485 before adaptation to 0.7327 after. Frozen-policy probes score 0.6146-0.6436, and their advantage over the linear probe is statistically inconclusive. Pairings use similar normalized elapsed times within tasks, not necessarily equivalent physical progress. Mean task success falls from 58% to 55%, so the result supports better failure discrimination, not better control or established safety benefits.

## Source excerpt

arXiv:2611.00394v1 Announce Type: new Abstract: Can a vision-language-action (VLA) policy signal likely task failure from its own action probabilities? We introduce CalVLA, which shapes native action-token confidence using successful and failed rollouts and an Action-Confidence Alignment (ACal) squared margin-ranking loss. Low-rank adapters update OpenVLA-7B without an auxiliary failure detector. Observations are paired within tasks at similar normalized elapsed times, not necessarily equivalent physical progress. On a held-out test split from 500 LIBERO-10 configurations, adapted confidence achieves a joint failure-discrimination AUROC of $0.7392\pm0.0480$ over three seeds, averaging scores at 130 and 260 steps. Frozen-policy probes reach 0.6146--0.6436, but superiority over the linear probe is statistically inconclusive. For the same confidence summary, AUROC rises from 0.4485 before adaptation to 0.7327 after adaptation. Mean task success decreases from 58% to 55%. These results show improved failure discrimination, not improved control or established safety benefits.