Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations

The article introduces the \texttt{Call Playbook} dataset, which consists of five classification tasks based on real-world B2B conversations, aimed at enhancing in-context learning (ICL) for complex, multi-party interactions. The proposed method includes novel knowledge extraction techniques that reduce token usage by 99% and improve macro-averaged AUC by up to 7% compared to traditional ICL, maintaining performance even with increased context length. This advancement addresses critical challenges in NLP applications, providing more transparent and efficient classification logic for practitioners working with specialized domain tasks.