Large language model adaptation through few-shots learning and post-hoc calibration

Abstract

In the context of adapting language models to a specific task, using prompt engineering often yields performance gains without requiring access to the internal parameters of the model. Another form of adaptation, much less studied in the literature, is achieved through posthoc calibration techniques, where only the model’s output scores are accessed and modified via a function to enhance task performance. This “gray-box” approach, which only utilizes the values from the model’s output layer, offers a computationally inexpensive alternative to supervised learning techniques and has not been thoroughly investigated in the literature. This work presents some preliminary results in which the combination of prompt engineering and post-hoc calibration shows improvements in multiple-choice social behavior question tasks on two large-scale language models (Phi-1.5 and Phi-2). The results obtained so far suggest that these two techniques are complementary, paving the way for the development of techniques that systematically combine prompt engineering with post-hoc calibration to improve model performance.