Highlight-Then-Summarize: Learning to Compress Evidence for Long-Context Understanding

Researchers have introduced Highlight-Then-Summarize, a method designed to improve how large language models process long documents by identifying and compressing task-relevant evidence. This technique addresses the challenge of sparse information by filtering out redundant content before the model performs its reasoning. By reducing the volume of data models must navigate, this approach aims to increase efficiency and accuracy in tasks involving extensive texts, code, or complex conversations.