An evaluation of OpenAI's o3 and o4-mini models on Humanity’s Last Exam reveals that their calibration errors are significantly lower than previous generations, though o3 is not statistically better than a random baseline. The analysis indicates that o3 exhibits broad underconfidence compared to predecessors like o1, which tend to predict high confidence across tasks.
- On Humanity’s Last Exam, o3's confidence distribution is nearly uniform, resulting in a lower calibration error than other models which perform worse than a random baseline.
- Calibration curves show no strong visual correlation between accuracy and confidence for o3, suggesting its lower average confidence on hard tasks drives the metric rather than precise self-awareness.
- On the saturated GSM8k dataset, o3 remains broadly underconfident while o4-mini is better calibrated; however, modifying the prompt to request reasoning for confidence scores reduces o3's calibration error from 24% to 9%.
The findings suggest that while newer OpenAI models are better calibrated than predecessors, they struggle with consistent confidence elicitation, particularly on easy tasks where they remain underconfident.