The research that made us want to build things, written down the way we'd want to be written down.

A notes repo, not a mirror. This repo is our study of one paper — Adams et al., From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting (arXiv:2309.04269) — and, because that paper gave our note-taking its spine, it is also the canonical home of the methodology every OpenCnid paper repo follows. (Yes, we summarized the summarization paper with its own method. It felt rude not to.) No papers are hosted here and none ever will be. The note is an original synthesis — our words, their findings, a citation at every level of the chain.

Important

The one-way rule. When a note and its paper disagree, the paper wins and the note gets fixed. No exceptions, no negotiation, no "but we liked our version better." That rule is the entire reason we can call these notes ground truth with a straight face.

Every note is a chain, and the chain is the artifact. A note starts sparse and gets rewritten four times at fixed length, fusing in a few more salient facts per round — methods, datasets, exact numbers, ablations, limitations — and no fact enters the chain without a locator pointing back into the paper (§ section, Table N, Figure N). We keep all five tiers, not just the densest:

tier	what it is
T1 — sparse	the problem, the approach, the headline result. For readers whose coffee hasn't kicked in yet
T2–T4	same length each, folding in 2–3 more salient entities per round, every one with a locator. The plot thickens; the word count doesn't
T5 — dense	maximally fused, still readable, every claim traceable. The one you bring to a design review
key results	exact values from the source, never rounded, each with its table or figure
our take	the only opinionated section, clearly ours, quarantined like it's contagious (opinions are)

🛠️ Cloned this and using Claude Code? There's a density-chain skill baked in: say "add a note for <paper>" and it runs the whole pipeline in house style — fetch, verify with locators, five tiers at a held budget, pinned provenance, zero numbers from memory. You have officially run out of excuses.

Why this exists

We're a new lab. The honest way to introduce yourself is to say who taught you — precisely, with page numbers. (It's also considerably cheaper than a marketing department.) OpenCnid Labs builds things because other people's research made us want to, and we write that recognition down one repo per paper, each named after the paper it studies — so the people actually searching for the research can find our study of it. This repo is the first, and it doubles as the template:

• density-chain.md — the five-tier note on the Chain of Density paper.
• METHOD.md and the synthesis prompt — the house methodology, canonical here because this is the paper it came from.
• the density-chain skill — the methodology as a runnable Claude Code skill, so every future paper repo comes out of the same oven.
• llm-research-inspirations — the map across all our paper repos: which of our work each paper shaped, and why, with receipts.

Authority runs one direction, and only one:

flowchart LR
    paper["📄 the paper<br/><i>canonical</i>"] --> note["🔗 our note<br/><i>working ground truth</i>"]
    note --> entry["🧾 inspirations entry<br/><i>interpretation</i>"]
    entry -. never .-> paper

Loading

A list entry can't cite itself into being true; it points at a note, and the note points into the paper.

How a note gets written

METHOD.md is the short version; chain-of-density-synthesis-prompt.md is the full authoring framework. The pipeline in one breath: extract evidence first — atomic claims, each with a locator and its qualifiers intact — then densify, and pick the final tier by audit rubric (entailment, attribution, qualifier integrity, readability), never "densest wins." (T5 has feelings about this. T5 will cope.) Long papers go through a map/reduce pass so a result buried mid-paper doesn't get lost to the middle of anyone's context window.

Three rules that don't bend:

1. Own words, always. At most one short attributed quote per note; never a figure, never a table, never a passage.
2. Exact numbers, located. A quantity without a locator doesn't go in.
3. Pin the version. Every note records the exact arXiv vN it studied and the date it was last checked. Papers move; a note is only ground truth relative to a pin.

🏔️ Standing on the shoulders of giants

The actual science in this story was done by Griffin Adams (Columbia), Alex Fabbri (Salesforce AI), Faisal Ladhak (Columbia), Eric Lehman (MIT), and Noémie Elhadad (Columbia) — they designed the prompt, ran the study, and personally read five hundred summaries so the rest of us could know which tier humans actually like. We wrote a note about it. Those are very different jobs, and only one of them deserves your citation.

Do not cite "some repo on GitHub." Cite them — BibTeX below.

📥 Want the PDF? One command, straight from the source

We don't keep a copy here (see: entire ethos, above). arXiv hosts it beautifully, for free, forever — no middlemen, no photocopier smell:

curl -L -o chain-of-density.pdf https://arxiv.org/pdf/2309.04269v1

Prefer the version of record? The ACL Anthology page has it, DOI and all.

📚 Cite the humans, not us

This repo is not a citable source; it's a signpost pointing at one. If you take one thing away, take this:

@inproceedings{adams-etal-2023-sparse,
    title = "From Sparse to Dense: {GPT}-4 Summarization with Chain of Density Prompting",
    author = "Adams, Griffin  and
      Fabbri, Alex  and
      Ladhak, Faisal  and
      Lehman, Eric  and
      Elhadad, No{\'e}mie",
    editor = "Dong, Yue  and
      Xiao, Wen  and
      Wang, Lu  and
      Liu, Fei  and
      Carenini, Giuseppe",
    booktitle = "Proceedings of the 4th New Frontiers in Summarization Workshop",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.newsum-1.7/",
    doi = "10.18653/v1/2023.newsum-1.7",
    pages = "68--74"
}

Kept honest by machine

index.json is the machine-readable face of every paper repo: the source pin, the verification date, the tags. Trellis, our current project, consumes those indexes and owns freshness — when a source revs (v2, errata, retraction), the note gets flagged before we get embarrassed.

Honest notes

• CoD is a tool, not a truth serum. The original study covered news articles and GPT-4; later work found denser-but-less-readable summaries in clinical settings and information loss on noisy multi-document inputs. That is exactly why we extract evidence before densifying, keep every tier, and select by audit instead of by density.
• Readers have a ceiling. Adams et al. found preference peaks near human-written density and falls past it. T5 is dense, not maximal.
• Summaries are lossy by construction. The locators are the refund policy: any claim can be walked back into its source in one hop.
• We will get things wrong. When we do, the fix lands source-first and the correction is public history. If we've mangled your paper, open an issue — correcting the record is the project.
• A human and an AI wrote this repo together. The human kept asking for more jokes; the AI kept adding citations. We disclose this because disclosure is sort of our whole thing.

Layout

density-chain.md                        the five-tier note on this repo's paper
METHOD.md                               the house methodology (canonical copy)
chain-of-density-synthesis-prompt.md    the full authoring framework
index.json                              machine-readable pin + verification metadata
AGENTS.md                               the agents' front door — consuming + maintaining notes
prompts/                                reusable session prompts (batch runs, handoffs)
.claude/skills/density-chain/           the baked-in skill — clone, then "add a note for <paper>"
assets/                                 banner art (the dots are load-bearing)

Every future paper repo follows the same shape, minus the methodology files — those live here and get linked, not copied.

What's next

• More paper repos. The density-chain skill shipped (see the toolbox note above), which drops the marginal cost of recognizing a paper to roughly one sentence.
• The staleness bot: Trellis opening an issue the day an arXiv v(N+1) lands on anything we've pinned.
• Coverage: the opening slate is Anthropic-heavy on purpose — their papers are a large part of why we're here — then outward to the rest of the field.

References — the methodology stands on published work (nine papers deep; click to expand)

• The method: Adams et al., From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting (2309.04269)
• Domain adaptation: Shrestha & Mahmoud (2506.14192)
• Mixed later evidence: Nagar et al. (2025.acl-long.134); Lee et al. (2025.neusymbridge-1.1)
• Lost in the middle: Liu et al. (2024.tacl-1.9)
• Citation + coverage is hard: Laban et al., SummHay (2024.emnlp-main.552); Buchmann et al. (2024.emnlp-main.463)
• Abstraction vs. factuality: Dreyer et al. (2023.findings-eacl.156); Xiao & Carenini (2023.codi-1.9)
• Redundancy control: Xiao & Carenini (2012.00052)

License

Notes and prose: CC BY 4.0 © OpenCnid Labs. The papers we summarize belong to their authors — that's the point.

---

_{No papers were harmed, stored, or even lightly cached in the making of this repository.}