Chain of Density: Definition and Examples

Chain of Density (CoD) is a prompt engineering method designed to produce high-quality summaries through successive iterations. The principle is simple: ask the model to generate an initial summary, then rewrite it several times by integrating more key entities and information, without ever exceeding the initial length.

At each iteration, the model identifies missing entities from the source text, then merges them into the existing summary by compressing, reformulating, or removing the least important details. The result is a text that progressively becomes denser in useful information while remaining fluid and readable.

This technique is based on a 2023 research paper by researchers from Columbia University and Salesforce. Their work showed that summaries produced after 3 to 5 densification iterations achieve a quality level comparable to human-written summaries and are often preferred by readers.

Chain of Density illustrates a fundamental principle of prompt engineering: rather than asking for a perfect result in a single pass, guide the model through a structured process of progressive improvement. This iterative approach better leverages the reasoning and reformulation capabilities of large language models.