What is chain of density?

Chain of Density (CoD), from Adams et al. (2023), is a prompting technique that generates several summaries of identical length, each iteration adding 1-3 previously missing entities while compressing existing wording. The result is a progressively denser, more informative summary at a fixed length.

Why keep the word count fixed?

Fixing the length is what forces densification. Because the summary cannot grow, the model must fuse, compress, and drop filler to make room for new entities, which raises information content per word rather than just adding more text.

How many iterations should I use?

Around five is the sweet spot identified in the original paper. Too few and the summary stays sparse; too many and it becomes so dense it reads awkwardly or starts losing readability.

What does the output look like?

The prompt requests a JSON array with one object per iteration, each containing the missing entities added and the denser summary. The final element is the densest version, which is usually the one you want.

When is a dense summary worse?

Extremely dense summaries can read like a list of facts and be harder to follow. If readability matters more than information density, use an earlier iteration or a standard summary prompt instead.

Chain-of-Density Summarization Prompt

Chain-of-density summarization prompt

A normal summary often skips the specific entities — names, numbers, places — that carry the real information. Chain of Density (CoD), introduced by Adams et al. in 2023, fixes that by generating several summaries of the same length, each one adding missing entities while compressing existing wording. The result is a summary that packs far more information into the same number of words. This builder generates a correct CoD prompt with your chosen length and iteration count.

How it works

You paste the source text, set a fixed word length, and choose how many density passes to run. The generated prompt instructs the model to repeat two steps: identify 1-3 informative entities missing from the previous summary, then rewrite a new summary of identical length that includes all prior entities plus the new ones. Because the length is fixed, the model must fuse and compress to make room, which is what drives densification. Output is a JSON array, one object per iteration, with the densest summary last.

Tips and notes

Take the last element. The final iteration is the densest; earlier ones are intermediate steps.
Five iterations is the default. It is the paper’s recommendation — enough to densify without becoming unreadable.
Fixed length is the trick. Do not let the model grow the summary; the constraint is what raises information density.
Watch readability. If the densest version reads like a fact dump, drop back to an earlier iteration.