Tag
This paper investigates whether probabilistic calibration in language models can be improved through fine-tuning, comparing soft-target and hard-target methods across 12 models. The results show that calibration is a trainable capability, though gains sometimes reduce downstream arithmetic reasoning capabilities.