Miranda Muqing Miao, Young-Min Cho, Lyle Ungar
CORAL is a method that improves the accuracy and calibration of large language models during inference without retraining.
Large language models often struggle with being well-calibrated, meaning they can be overconfident or underconfident in their predictions, especially after undergoing certain training processes. Retraining these models to improve calibration can be very costly. CORAL is a new technique that improves model accuracy and confidence without needing to change the model's original training. It works by using a special method to adjust the model's internal signals during inference, leading to better performance on various question-answering benchmarks.
