Event
Research on Confidence Calibration of VLA Models Published in TMLR
1 reports · 1 sources
Thomas P Zollo et al. published the first systematic study of confidence calibration for vision-language-action (VLA) foundation models, establishing baselines for VLA confidence estimation, examining the relationship between task success rates and calibration error and how calibration changes over time, and proposing two lightweight correction techniques—prompt ensembles and per-action Platt scaling—to address the observed miscalibration. The paper is 34 pages with 7 tables and 22 figures, has been published in TMLR, and the final version is complete.
Report timeline
- Study on confidence calibration for VLA models proposes two lightweight correction techniques
This study presents the first systematic study of confidence calibration for vision-language-action (VLA) foundation models, establishes a baseline for VLA confidence estimation, examines the relationship between task success rate and calibration error as well as how calibration changes over time, and proposes two lightweight techniques, prompt ensembles and action-wise Platt scaling, to correct the observed miscalibration. The paper is 34 pages long with 7 tables and 22 figures, and has been published in TMLR as a final version.
arXiv cs.RO ·