Chain of Density (CoD)
A single-pass summary captures the gist but misses the details. Chain of Density fixes this through iterative refinement — each pass packs more entities and key information into the same word count, producing summaries that are progressively denser, more informative, and consistently preferred by human evaluators.
Introduced: Chain of Density was published in 2023 by Adams et al. from Salesforce Research and MIT. The technique addresses a persistent problem in AI-generated summaries: they tend to be either too verbose (lots of filler, few facts) or too compressed (missing crucial context). CoD introduces a structured iterative process where the model generates an initial summary, then refines it across multiple rounds — each time adding 1–3 new salient entities while keeping the word count constant. The result is a summary that becomes progressively more information-dense with each iteration.
Modern LLM Status: Chain of Density remains one of the most practical summarization techniques in 2026. The iterative densification approach produces consistently better summaries than single-pass methods. It is widely used in content management systems, news aggregation platforms, and automated report generation where information density matters. Many enterprise summarization tools have adopted CoD’s iterative approach as a default pattern, and the technique generalizes well across Claude, GPT-4, and Gemini models.
Density Is the Missing Dimension
Most summarization techniques focus on two dimensions: what to include and how long the summary should be. But there is a third dimension that matters just as much: information density — how much meaning is packed into each sentence. A 100-word summary with three key facts is fundamentally less useful than a 100-word summary with twelve key facts, even though they are the same length.
Chain of Density treats density as a tunable parameter. Instead of asking for one summary and hoping for the best, CoD generates a sequence of summaries at increasing density levels. The first summary is deliberately loose and readable. Each subsequent version replaces generic phrasing with specific entities, swaps vague descriptions for precise facts, and fuses redundant sentences — all while maintaining the same word count. The model effectively learns to compress more signal into the same space.
Think of it like packing a suitcase: the first attempt leaves gaps and wasted space. With each repack, you rearrange and compress items more efficiently until every cubic inch serves a purpose — without making the suitcase any bigger.
Single-pass summarization forces the model to make all inclusion decisions simultaneously — it must decide what matters before it has seen its own output. CoD breaks this into stages: first capture the broad narrative, then progressively inject the specific entities and facts that a reader would need. This iterative approach mirrors how human editors refine summaries: rough draft first, then tighten, then tighten again. Research shows human evaluators consistently prefer CoD summaries (typically around iteration 3–4) over single-pass alternatives for informativeness without sacrificing readability.
The Chain of Density Process
Five stages from source text to optimally dense summary
Generate the Initial Summary
Start with the source text and produce a first-pass summary at a fixed word count (typically 80–100 words). This initial summary is deliberately entity-sparse — it captures the broad narrative and main topic but uses general language rather than specific facts. Think of it as the “skeleton” summary that establishes the structure subsequent iterations will densify.
Source: A 2,000-word article about a new renewable energy initiative.
Summary 1: “A major energy company has launched a new renewable energy project. The initiative involves building facilities in several locations and aims to reduce carbon emissions. Industry experts say this represents a significant step forward for clean energy adoption.” (Entity count: ~3)
Identify Missing Entities
The model reviews the source text against its current summary and identifies 1–3 salient entities (names, dates, figures, locations, organizations) that are present in the source but missing from the summary. These must be genuinely informative — not trivial details, but facts that would meaningfully enhance a reader’s understanding of the topic.
Missing entities identified: (1) Company name: “Meridian Power Corp.” (2) Investment amount: “$4.2 billion” (3) Target capacity: “12 gigawatts by 2030”
Rewrite with Increased Density
Incorporate the missing entities into the summary while maintaining the same word count. This requires the model to perform fusion (combining sentences), compression (replacing verbose phrases with concise ones), and substitution (swapping generic references for specific names and figures). The key constraint is that the summary must not grow longer — density increases, length stays constant.
Summary 2: “Meridian Power Corp. has launched a $4.2 billion renewable energy initiative targeting 12 gigawatts of capacity by 2030. The project involves solar and wind installations across Texas, Nevada, and Arizona. Analysts describe it as the largest private clean energy commitment in U.S. history.” (Entity count: ~8, same word count)
Iterate 3–5 Times
Repeat the identify-and-rewrite cycle 3–5 times. Each iteration adds more entities and compresses further. Research shows that summaries typically reach optimal density around iteration 3–4 — dense enough to be highly informative, but not so compressed that readability suffers. Beyond iteration 5, summaries often become telegraphic and hard to parse, a point of diminishing returns.
Summary 4 (final): “Meridian Power Corp.’s $4.2B initiative — 12GW solar/wind capacity across TX, NV, AZ by 2030 — marks the largest private U.S. clean energy commitment. CEO Maria Chen projects 15,000 construction jobs, 3,200 permanent positions, and a 40% reduction in Meridian’s carbon footprint. The DOE has approved $800M in tax credits for the project.” (Entity count: ~15, same word count)
Select the Optimal Density Level
Choose the summary iteration that best balances density and readability for your use case. For general audiences, iteration 2–3 is typically ideal. For expert readers who want maximum information density, iteration 4–5 may be preferred. The beauty of CoD is that it produces a spectrum of summaries rather than a single output, giving you the power to choose the right density level for each context.
For a newsletter: use Summary 2 (readable, well-paced). For an executive brief: use Summary 4 (maximum information density). For a tweet-length teaser: use Summary 1 as a starting point. Each serves a different communication need from the same source material.
See the Difference
Why iterative densification outperforms single-pass summarization
Single-Pass Summary
Summarize this research paper on climate change adaptation strategies in 80 words.
This paper examines how communities can adapt to climate change. The researchers studied several approaches to dealing with rising temperatures and changing weather patterns. They found that some strategies are more effective than others. The paper recommends that policymakers take action to help communities prepare for future climate impacts. Overall, adaptation is becoming increasingly important as climate change accelerates.
Chain of Density (Iteration 4)
After 4 iterations of entity injection and compression at a fixed 80-word count:
A meta-analysis of 147 climate adaptation programs across 23 countries (2018–2024) found that nature-based solutions (wetland restoration, urban canopy expansion) reduced heat-related mortality by 34% versus engineered alternatives. Coastal communities using managed retreat showed 60% lower flood damage than those relying on seawalls alone. The study, led by Dr. Sarah Chen at Stanford, recommends integrating indigenous land management practices, citing 89% success rates in Australian bushfire-adapted communities.
Practice Responsible AI
Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.
48 US states now require AI transparency in key areas. Critical thinking remains your strongest tool against misinformation.
Chain of Density in Action
See how iterative densification transforms summaries across domains
A 15-page quarterly earnings report from a technology company, covering revenue, product launches, headcount changes, and forward guidance.
Iteration 1 (entity-sparse): “The company reported strong quarterly results, with revenue exceeding expectations. Several new products were launched, and the company provided positive guidance for the next quarter. Hiring continued across multiple departments.”
Iteration 3 (mid-density): “Q3 revenue of $48.2B (+14% YoY) beat consensus by $1.8B, driven by cloud services (+31%) and AI infrastructure (+67%). Three new products launched: an enterprise AI assistant, a custom chip for inference workloads, and an upgraded developer platform. Headcount grew 8% to 142,000.”
Iteration 5 (maximum density): “Q3: $48.2B revenue (+14% YoY, $1.8B beat), cloud $21.4B (+31%), AI infra $8.7B (+67%). Launched: Enterprise AI Copilot (200K waitlist), Tensor-4 inference chip (3x perf/watt vs. prior gen), DevPlatform 3.0. Headcount 142K (+8%), 60% AI-focused roles. Q4 guide: $50–52B, 23% margin expansion. Buyback: $15B authorized.”
Note: Always verify financial figures in AI-generated summaries against the original source documents. CoD produces denser summaries, but density does not guarantee accuracy — cross-reference all numbers before using in decisions or reports.
A 12-page neuroscience paper on the effects of sleep deprivation on cognitive performance, including methodology, results, and clinical implications.
Iteration 1 (entity-sparse): “Researchers studied how lack of sleep affects thinking ability. They tested participants under different sleep conditions and found significant impacts on several measures of cognitive function. The findings suggest that sleep is important for maintaining mental performance.”
Iteration 3 (mid-density): “A controlled study of 240 adults (ages 22–45) found that 36 hours of sleep deprivation reduced working memory accuracy by 38% and reaction time by 27%. Executive function tasks showed 42% more errors. fMRI scans revealed decreased prefrontal cortex activation correlating with performance decline.”
Iteration 5 (maximum density): “240 adults (22–45, M=31.4), 36hr sleep deprivation: working memory −38%, reaction time −27%, executive function errors +42%. fMRI: prefrontal cortex activation −31%, compensatory parietal increase +18%. Recovery: 2 nights sufficient for reaction time, 4 nights for executive function. Clinical: shift workers (n=89 subset) showed 2.3x baseline impairment. Authors recommend max 16hr wake periods for safety-critical roles.”
Note: When using AI to summarize scientific research, always verify statistical claims, sample sizes, and effect sizes against the original paper. AI-generated summaries may conflate findings from different sections or misrepresent confidence intervals.
A 1,500-word news article about a new international trade agreement involving multiple countries, with economic projections and political context.
Iteration 1 (entity-sparse): “Several countries have signed a new trade agreement that is expected to boost economic growth. The deal covers various sectors and includes provisions for reducing tariffs. Leaders expressed optimism about the agreement’s potential impact.”
Iteration 3 (mid-density): “The Pacific Rim Digital Trade Accord, signed by 15 nations including Japan, Australia, and South Korea, eliminates tariffs on digital services and sets cross-border data flow standards. Economists project $340B in new trade volume by 2028. The U.S. and China are notably absent from the agreement.”
Iteration 5 (maximum density): “Pacific Rim Digital Trade Accord: 15 nations (Japan, Australia, South Korea, Singapore + 11 ASEAN partners), zero tariffs on digital services, unified data-flow/privacy standards. Projected $340B new trade by 2028, 2.1M tech jobs. U.S./China absent; EU observer status. Ratification: 12/15 needed, expected Q2 2026. Key dispute: India’s data localization carve-out for financial services.”
Note: AI-generated news summaries should always be verified against the original source. Entity-dense summaries pack many claims into compact space — if even one entity is wrong, the entire summary may mislead. Always fact-check before sharing or publishing.
When to Use Chain of Density
Best for summarization tasks where information density matters
Perfect For
When decision-makers need maximum information in minimum reading time — CoD packs every sentence with actionable facts, figures, and names.
News digests, research roundups, and newsletter curation where readers expect high-density summaries that capture the essential facts from lengthy source materials.
Automated systems that process hundreds of documents and need consistent, entity-rich summaries — CoD produces reliably dense output across diverse input types.
When the same source needs summaries at different density levels — CoD naturally produces a spectrum from casual overview to expert-dense briefing in a single pipeline.
Skip It When
Texts under 500 words rarely benefit from iterative densification — there are not enough entities to justify multiple refinement passes.
Stories, opinion pieces, or persuasive writing where tone and flow matter more than entity density — CoD optimizes for information, not emotional impact.
When summaries are needed instantly — CoD requires multiple sequential LLM calls (one per iteration), making it slower than single-pass approaches.
Use Cases
Where Chain of Density delivers the most value
Research Literature Review
Summarize dozens of academic papers into entity-dense abstracts that capture key findings, methodologies, sample sizes, and statistical results in consistent format.
Legal Document Summarization
Compress lengthy contracts, rulings, and regulatory filings into dense summaries that preserve all parties, dates, obligations, and key clauses without loss of critical detail.
Medical Record Synthesis
Create dense patient summaries from extensive medical records, ensuring diagnoses, medications, dosages, lab values, and treatment timelines are all captured in compact format.
Financial Report Digests
Transform quarterly earnings reports, SEC filings, and market analyses into information-packed briefs with all key metrics, projections, and risk factors preserved.
Meeting Minutes Compression
Convert lengthy meeting transcripts into dense action-oriented summaries that capture all decisions, owners, deadlines, and key discussion points in minimal space.
News Aggregation Platforms
Generate consistent, fact-packed summaries for hundreds of daily articles, giving readers the essential who/what/when/where/why in a standardized, scannable format.
Where Chain of Density Fits
CoD bridges basic summarization and structured iterative refinement
Research on Chain of Density consistently finds that human evaluators prefer summaries from iterations 3–4 over both the sparse initial summary and the maximally dense final iteration. This suggests a “Goldilocks zone” of density where summaries are packed with useful information but still flow naturally as readable prose. When implementing CoD in production, consider defaulting to iteration 3 for general audiences and offering iteration 4–5 as an “expert mode” for domain specialists who prioritize information extraction over readability.
Related Techniques
Explore complementary summarization and refinement techniques
Densify Your Summaries
Try Chain of Density on your own documents or explore summarization techniques with our interactive tools.