I'm trying to get my models to write summaries that aren't just "fluff." I heard about the Chain-of-Density (CoD) prompting method. Does it actually work for highly technical reports, or does it lead to information overload for the reader? What are the specific prompt structures for 2025?
3 answers
Chain-of-Density is a game-changer for technical documentation. The core idea is to start with a sparse summary and iteratively "densify" it by identifying and adding five unique, missing entities in each pass. We’ve used it for our medical research digests, and it forced the model to stop using generic transition words and start including specific drug names, percentages, and clinical outcomes. The key is to stop at 5 iterations; any more and the text becomes a "word salad" that even experts struggle to read. It’s the best way to squeeze every bit of value out of your token budget.
Do you find that different models (like Gemini vs Claude) handle the "Density" instructions differently? I feel like some models are more "stubborn" about keeping their original summary.
CoD is great, but be careful with your audience. For a C-suite executive, a "dense" summary might be too much. Sometimes a "Chain-of-Simplicity" is more effective!
Haha, very true Patricia! Know your reader before you maximize the information density.
Christopher, you're right. In our 2024 benchmarks, Claude 3.5 was significantly better at "re-writing" rather than just "adding to the end." To fix this in Gemini, we found that you have to explicitly tell the model to "Redraft the entire summary from scratch with the new entities integrated," rather than just "adding them." If you don't force a full redraft, the model gets lazy and the flow of the summary becomes disjointed. It's a small prompt tweak that makes a massive difference in the final quality.