A new preprint says the next hard problem in making AI answers verifiable may be more basic than adding a source link: working out what, exactly, a language model should cite. Its argument is that data citation for large language models is distinct from—and harder than—document-level citation grounding, where an answer is tied to a passage in a text. The paper asks how models can cite data in a way that lets readers check an answer’s basis, trace where the underlying material came from, and give fair credit to the people who created or curated it.
The paper presents its agenda around three interconnected challenges: material absorbed during training, structured data consulted while a model generates an answer, and facts drawn from knowledge graphs. Each raises a different question about how fine-grained a citation should be and how long its trail should remain usable. The author presents these as research directions, not as a tested solution.
Training data leaves an attribution gap
The training-data challenge is to turn influence estimates into references for corpora absorbed into model parameters. Influence estimates are meant to indicate which training material may have shaped a particular output, but the paper treats the conversion from that signal to a citation as an open research direction. The practical question is how to connect a generated passage to identifiable corpus material without assuming that the attribution is already exact.
The technical gap is substantial. The preprint says current training-data attribution methods are computationally expensive and approximate, and that they do not produce scholarly-style citations. It also leaves open how influence estimates can become rigorous, efficient references for individual training items and contribution weights. In the paper’s framing, this is a research direction rather than a validated attribution system.
The citation problem becomes more concrete at answer time
The second challenge appears at inference time—the stage when a model produces an answer. If a system uses structured data at that point, a useful reference would need to identify the dataset’s creator, title, version, persistent identifier and access information. It would also need to point to the relevant subset or query result at an appropriate level of detail, so that the citation names the material actually used rather than merely the larger source.
The paper says current large language model systems do not support this kind of subset citation. It treats inference-time and retrieval-augmented citation as the most practical starting point for building a broader approach, but provides no implemented citation system or performance evaluation. That means the proposal identifies what a reference should contain; it does not measure whether the design improves verification, provenance tracking or credit.
A single fact may need its own source trail
Knowledge graphs pose a problem at the level of the individual fact. A citation to the whole graph is too coarse: it can conflate independent assertions, while an individual triple—a single knowledge-graph fact—may not carry the bibliographic metadata needed to identify its source. The unresolved question is what an LLM should cite when it draws on one fact rather than the entire graph.
When a fact has been assembled from several sources, citation is also an accounting problem. The paper argues that credit should be divided in proportion to each source’s contribution and carried through the provenance graph. That makes the task a network-propagation problem: the system would have to follow links among sources and pass credit through them. Existing approaches, the paper says, do not address this problem at LLM-pipeline scale.
The proposed fix is a coordinated redesign
The proposed answer is therefore broader than a new reference format. The paper calls for five elements to be designed together: a unified citation model; architectures that preserve provenance; scalable credit distribution; benchmarks for data citation; and standards or infrastructure. The aim is to connect the citation shown to a reader with the chain of training, retrieval and fact construction behind an answer.
The author’s wider point is that no single technical community can settle the issue in isolation. Citation rules, model architecture, data management, provenance and credit all meet in the same output, so the proposed agenda depends on co-design across those areas. But the preprint does not claim to have supplied that common system. It offers a map of the problem and the work needed to test, standardize and operate data citations in model outputs.
That distinction matters for anyone trying to judge AI answers. The analysis does not show a measured rate of citation failure or data misattribution, and it does not establish that data citations improve verification, provenance, credit or trust. Its contribution is a definition of the problem: a credible citation trail must be designed for the different ways a model obtains and combines information, from absorbed corpora to retrieved subsets and individual graph facts.
Paper data and sources
Original title: Data Citation for Large Language Models: A Challenge
Authors: Gianmaria Silvello
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text