Common Citation Format Mistakes in AI Research (And How to Fix Them)
Open ten AI papers picked at random from arXiv and you will find at least three with a citation error that would make a librarian wince - a preprint cited as if it were peer-reviewed, a model version left unspecified, or a reference list that mixes three different formatting conventions. These aren't cosmetic problems. In a field where reproducibility and provenance already struggle against fast-moving preprints and constantly updated models, a sloppy citation can send a reader - human or AI - to the wrong source entirely.
This matters more than it used to. AI systems now parse and re-cite content from the web, and they inherit the citation habits of whatever they were trained on or retrieve at inference time. If your source list is a mess, you are teaching the next generation of retrieval systems that messiness is acceptable. Below are the mistakes that show up most often, why they happen, and what to do instead.
Citing a preprint as if it were the final published version
arXiv preprints get revised - sometimes substantially - between v1 and the camera-ready conference version. A common error is citing "v1" content (a claim, a number, a method detail) but linking the DOI or venue of the final published paper, which may say something different. The fix is mechanical but rarely followed: always note the arXiv version number (e.g. arXiv:2310.06825v2) if you're citing preprint content specifically, and separately note the peer-reviewed venue if the paper was later accepted somewhere like NeurIPS or ACL. These are not interchangeable citations even though they describe "the same" paper.
If you maintain a personal research library, this is exactly the kind of detail that gets lost without a system. For a deeper walkthrough on sourcing papers correctly in the first place, see how to find and cite AI research papers correctly.
Omitting the model version or checkpoint date
"We evaluated GPT-4" or "tested with Claude" is not a citation - it's a vague gesture. Model behavior shifts across checkpoints, and a claim benchmarked against one snapshot may not replicate against a later one. The minimum acceptable citation format for a model reference includes: the exact model name and version string (not the marketing name), the API or interface used, and the date of the experiment. Without the date, nobody - including you, six months later - can tell whether a discrepancy in results comes from the model changing or from a genuine methodological difference.
This is one of the areas where AI-specific citation practice diverges hardest from traditional academic citation, because the object being cited (a live, updating system) doesn't behave like a fixed printed source. Anyone building a workflow around this should look at how AI training data optimization intersects with content freshness - the same instability that makes model versioning tricky also affects how quickly content gets absorbed and cited by these systems in the first place.
Mixing citation styles within the same document
APA, IEEE numbered references, and Chicago author-date each solve a different problem - IEEE numbered citations save space in dense technical papers, APA privileges author-year recognizability, Chicago handles footnotes and extended commentary well. The mistake isn't picking one; it's drifting between them mid-document, often because content was pasted from multiple drafts or co-authors used different reference managers. A paper that cites [12] in one section and (Vaswani et al., 2017) in another looks unfinished, and reviewers notice immediately.
Pick the format the target venue requires and enforce it with a reference manager rather than manual formatting. Tools built for this - covered in detail in a comparison of tools for managing academic citations in AI research - catch these inconsistencies automatically before submission.
Broken or missing DOIs and dead arXiv links
A DOI (Digital Object Identifier) is supposed to be permanent, but AI papers frequently get cited via a direct URL to a lab's GitHub page, a company blog post, or a Google Doc - none of which are stable. Two years later, half of those links return a 404. The safest citation hierarchy, in order of durability, is: DOI first, then a persistent identifier like an arXiv ID, then a stable institutional repository link, and only as a last resort a company blog URL with an access date attached.
Access dates matter more in AI research than in most fields precisely because so much foundational material - technical reports, model cards, system cards - lives outside traditional peer review. The DOI Foundation maintains the resolution system that makes this citation format reliable when used correctly.
Citing a benchmark result without citing the benchmark's version
This one is specific to AI research and rarely discussed. Benchmarks like MMLU, HumanEval, or GLUE get revised, re-scored, or have their test sets partially leaked over time. Different harnesses (lm-evaluation-harness vs. a custom script) can produce meaningfully different scores on the identical model and identical benchmark. Treat the benchmark itself as a citable artifact with its own version history, not a fixed, universally agreed-upon yardstick.
Ignoring how AI systems themselves parse your citations
Here's the angle most style guides miss entirely: citation format doesn't just serve human readers and journal reviewers anymore. Large language models retrieving and summarizing your content will parse your reference list, and poorly structured citations degrade how reliably they attribute claims back to you. A citation buried in unstructured prose ("as shown by recent work from Google") gives a retrieval system nothing to anchor to. A citation with clear structure - author, year, title, stable identifier, ideally marked up with proper schema - gives it something concrete to cite in return.
This is where the overlap with schema markup for AI citations becomes directly relevant: structured, machine-readable citation data isn't just good academic hygiene, it's a visibility lever. If you want your research summaries or blog posts about AI research to actually get cited by tools like ChatGPT or Claude, the underlying reference formatting has to be parseable, not just readable. For teams publishing research-adjacent content at volume, a platform like ForgR can help keep that structural consistency intact across dozens of posts without each one drifting into a different citation convention.
Failing to distinguish primary and secondary sources
A surprising number of AI blog posts cite a news article about a paper instead of the paper itself, or cite a Twitter/X thread summarizing findings rather than the underlying technical report. This creates citation chains where errors compound - a journalist's simplification gets cited as if it were the original claim, and the nuance from the actual paper (confidence intervals, limitations sections, dataset caveats) disappears entirely. Always trace back to the primary source before citing, even if it takes an extra five minutes.
A quick reference: fixing the format
| Mistake | Fix |
|---|---|
| Preprint cited as final version | Note arXiv version + venue separately |
| No model version/date | Include exact checkpoint name and test date |
| Mixed citation styles | Pick one format, enforce with a reference manager |
| Dead links | DOI > arXiv ID > stable repo > blog URL with access date |
| Benchmark score without version | Specify benchmark version and evaluation harness |
| Secondary source cited as primary | Trace back to the original paper or report |
None of this requires exotic tooling - it requires discipline and a system. For a step-by-step process on finding and citing sources correctly from the start, the guide on how top AI conference papers actually cite sources walks through real examples worth studying before you draft your own reference list.
Key takeaways
- Always separate the arXiv preprint version from the peer-reviewed venue when both exist — they can contain different claims
- Include exact model checkpoint names and test dates whenever citing an LLM's behavior, since results drift across versions
- Pick one citation style (APA, IEEE, or Chicago) per document and enforce it with a reference manager, not manual editing
- Prioritize DOIs and arXiv IDs over blog URLs, which frequently break within a year or two
- Never cite a benchmark score without noting the benchmark version and evaluation harness used
- Structure citations for machine parseability, not just human readability, if you want AI models to cite you back accurately
Frequently asked questions
What is the most common citation mistake in AI research papers?
Citing a model or preprint without specifying its exact version and date. Since AI models and preprints change over time, an unversioned citation makes results impossible to reproduce or verify later.
Should I cite the arXiv preprint or the conference-published version of a paper?
Cite whichever version contains the claim you're referencing. If you're quoting content specific to a preprint revision, note the arXiv version number explicitly rather than assuming it matches the final published paper.
Why do dead links happen so often in AI research citations?
Many AI papers cite company blogs, GitHub repos, or informal reports rather than stable identifiers like DOIs. These sources get moved, renamed, or taken down far more often than peer-reviewed journal archives.
Does citation format actually affect whether AI models cite my content?
Yes. Clear, structured citations with identifiable authors, dates, and stable links are easier for retrieval systems to parse and re-attribute than vague or unstructured references buried in prose.
How should I cite a benchmark result like an MMLU score?
Specify the benchmark version, the evaluation harness or script used, and the prompt format if relevant. Raw percentage scores without this context are difficult to compare across papers.
Is it acceptable to cite a news article instead of the original research paper?
Only as a secondary reference alongside the primary source, never as a replacement. News summaries often omit caveats, limitations, and confidence intervals present in the original paper.