IC-1490Zero-shot CLIP is overconfident on STEM questions, with softmax confidence loosely related to actual accuracy

Jianhao Shen, Ye Yuan, Srbuhi Mirzoyan, Ming Zhang, Chenguang Wang

SourceMeasuring Vision-Language STEM Skills of Neural Models

The paper plots the relationship between CLIP's softmax confidence and its actual accuracy on the STEM test set. The zero-shot model is overconfident: its predicted confidence is systematically higher than its realized accuracy, and the two are only loosely correlated. After fine-tuning on the STEM training split, the model becomes more calibrated, suggesting the miscalibration is partly due to the zero-shot transfer setting rather than an inherent architectural property.

Evidence
observational
Caveat
No numerical calibration metric (e.g., ECE) is reported; the finding is based on visual inspection of the confidence-accuracy curve in Figure 7.
Model
CLIP / CLIP-ViT (LC)
Related findings
IC-1489, IC-1491
Extraction
automatic-extraction