<ul><li>In-Context Learning (ICL) in Large Language Models (LLMs) allows adaptation to new tasks with few examples but suffers from systematic biases affecting classification performance.</li><li>Existing calibration techniques shift decision boundaries in the logit space without altering orientation, leading to limited effectiveness.</li><li>Supervised Calibration (SC) is proposed as a loss-minimization framework that optimizes LLM's predictive probabilities through per-class affine transformations in the logit space, overcoming limitations of existing methods.</li><li>SC integrates purpose-built regularization techniques for stability and control, achieving superior performance in different shot settings across multiple datasets.</li></ul>

Boosting In-Context Learning in LLMs Through the Lens of Classical Supervised Learning

Discover more