Chain-of-Density

CoD operates through a multi-turn refinement loop where a language model repeatedly revises its own summary. Starting from an initial sparse summary, each iteration identifies missing salient entities and fuses them into the existing text without adding words. The process typically runs for 5 iterations, with each step producing a progressively denser output that packs more unique entities per token while maintaining the original word budget.

Unlike traditional summarization that prioritizes sentence importance, CoD explicitly tracks entity coverage as its optimization target. The technique identifies key named entities, technical terms, and factual anchors that are absent from the current summary and instructs the model to incorporate them. This entity-first approach is particularly valuable for legal documents where missing a specific party name, statute reference, or case citation constitutes a critical failure.

A defining characteristic of CoD is its strict length invariance. The summary must remain at the original target word count throughout all iterations, forcing the model to abstract and fuse concepts rather than simply appending new information. This constraint mirrors real-world legal use cases where summaries must fit within fixed display constraints—such as a single screen in a case management system or a standardized brief format.

Research on CoD reveals a measurable inverse relationship between information density and human readability. As entity density increases across iterations, human evaluators consistently rate summaries as less coherent and more difficult to parse. This creates a deployment design choice for legal applications: high-density summaries for expert rapid scanning versus lower-density summaries for client-facing or non-specialist audiences. The optimal stopping point depends on the end-user's domain expertise.

CoD summaries serve as highly efficient retrieval targets in Retrieval-Augmented Generation (RAG) architectures. Because each CoD summary packs maximum entity density into minimal tokens, it produces information-rich embeddings that improve semantic search recall for legal document collections. When a user queries for a specific case detail, the dense vector representation of a CoD summary is more likely to match than a sparse, narrative summary that omitted key entities.

A critical risk in iterative summarization is hallucination amplification—where a factual error introduced in one iteration compounds across subsequent refinements. CoD implementations for legal domains must incorporate factual consistency verification at each iteration step, typically using Natural Language Inference (NLI) models to check that newly fused entities are entailed by the source document. Without this guardrail, the density optimization can inadvertently produce highly confident but legally inaccurate summaries.