<ul><li>Post-training of large language models is essential for adapting pre-trained language models (PLMs) to align with human preferences and downstream tasks.</li><li>Post-trained language models (PoLMs) often suffer from over-confidence, assigning high confidence to both correct and incorrect outputs, which can undermine reliability in critical applications.</li><li>Disagreement-Aware Confidence Alignment (DACA) is a novel unsupervised method proposed to optimize parameters in post-hoc confidence calibration by addressing the scarcity of labeled data for individual downstream tasks.</li><li>DACA selectively uses only agreement examples for calibration to improve confidence calibration performance, as demonstrated by extensive experiments showing up to 15.08% improvement in the average ECE of LLMs like GPT-4o on common benchmarks.</li></ul>

Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator

Discover more