“Chain of Density” for Effective Text Summarization with Large Language Models

If you ever played with LLMs to perform text summarization, you’d know that condensing large volumes of information into digestible summaries is a challenging task.

Join our next cohort: Full-stack GenAI SaaS Product in 4 weeks!

However, there is a light at the of the tunnel!

Researchers from Salesforce AI, MIT, and Columbia recently showed the Chain of Density (CoD) prompt to summarize news articles (and probably others), that are not only succinct but also rich in details.

During their study, the researchers found out that:

Humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries.

In this article, I want to walk you through CoD’s entity-centric approach, encompassing key actors, locations, and concepts, without overwhelming you with excessive density that may hinder comprehension.

Let’s have a brief look at CoD before trying it out.

Understanding CoD Prompt

When you employ CoD prompt, LLM begins with the generation of an initial entity-sparse summary. This summary serves as a foundation, which is then iteratively…