A new calibration method for large language models proposed by MIT researchers. Called the 'thermometer', it requires less computing power while maintaining accuracy and can help identify model errors. Future plans include supporting more complex tasks and applying to larger models.