The pursuit of effective summarization has led to InfoSumm, a distillation approach to extract powerful summarizers from data without relying on LLMs’ capacity or human-written references. Using information-theoretic objectives, InfoSumm formulates saliency, faithfulness, and brevity as mutual information measures between document and summary.

• Presents a novel framework for automatic summarization, focusing on information-centric measures.
• Trains a powerful, compact summarizer with only 568M parameters, rivaling ChatGPT’s performance.
• Offers a controllable and efficient alternative to proprietary LLMs in summarization tasks.

This innovative approach to automatic summarization could shift the reliance away from vast LLMs towards more scalable and controllable models, allowing for a wider range of applications while maintaining high-quality summary generation.