Why Can't Large Language Models Cite Data Like They Cite Documents?

Large language models now sit between people and the information they seek. A growing body of research asks whether these models can point back to the sources behind their answers.

That research treats citation as a verification tool and applies it to textual documents. Scholarly citation does two more things: it assigns credit and it records provenance. Both functions apply to data as much as to text. The paper submitted on 26 Aug 2026 argues that data citation for large language models is an open challenge, distinct from document-level citation grounding and harder to solve. Current approaches ask a model to name the document that supports a claim. A document is a static container. Data lives in tables and knowledge graphs. Query results change over time. A model trained on a corpus absorbs statistical patterns into its parameters. The original datasets disappear inside the weights. No page number exists for a fact that emerged from millions of rows.

The author writes that training data attribution must turn influence estimates into references for corpora absorbed into model parameters. Influence estimates are not citations. They measure how much a training example shifts the model's behavior. A citation must identify a specific dataset, a specific subset, or a specific query result at the right granularity and fixity. Fixity means the cited object does not change after the citation is made. Most datasets do not guarantee that.

The Gap Between Documents and Data

The first direction targets training data attribution. Researchers need methods that map a model's output back to the training examples that shaped it. Those methods must produce references that scholars recognize. A reference to a corpus is not a reference to a row. The paper does not say how many training examples a typical model absorbs. It does not say whether current influence functions scale to trillion-token corpora. The second direction targets inference-time citation. When a model retrieves data to answer a query, it must cite the dataset and the subset. The query result itself must be citable at the moment it is produced. The paper does not specify which identifier schemes exist today. It does not name a single standard that satisfies fixity.

Three Research Directions

The third direction targets knowledge graph facts. A knowledge graph stores triples — subject, predicate, object. A model that outputs a fact from a graph must cite the triple. The paper asks what a reference to a single triple denotes. It asks how credit propagates along provenance chains. A triple may come from an extraction pipeline that drew from multiple sources. The paper does not describe an existing system that traces that chain end to end.

Why Cross-Community Work Is Required

Progress on all three directions depends on joint work across the database, information retrieval, knowledge representation, and artificial intelligence communities. Database researchers understand query provenance and fixity. Information retrieval researchers understand grounding and ranking. Knowledge representation researchers understand graph semantics and triple-level identity. Artificial intelligence researchers understand model internals and influence estimation. No single community holds all the pieces. The paper is seven pages long. It is classified as a journal paper. Its ACM classes are H.3.7 and H.2.8. The DOI is 10.48550/arXiv.2608.25663. The author is Gianmaria Silvello. The submission history shows one version. The abstract ends with a call for joint work. The paper does not propose a benchmark. It does not propose a metric. It does not name a conference where this work will be discussed