RAG attribution faithfulness study

Definition: Correctness is not Faithfulness in RAG Attributions studies citations in retrieval-augmented generation (RAG), where a language model receives retrieved documents when generating an answer. It distinguishes citation correctness, whether a source supports a statement, from citation faithfulness, whether the model causally relies on that source to generate the statement.

Key facts

• The arXiv record names Jonas Wallat, Maria Heuss, Maarten de Rijke and Avishek Anand as authors and dates submission to 23 December 2024.
• The paper's experimental setup tests Command-R+ on 1,444 NaturalQuestions questions answered by Wikipedia pages.
• As of 17 October 2026, the authors' code repository provides an attribution-analysis notebook and instructions for generating altered-document datasets and evaluating new predictions.

What separates support from causal reliance?

The paper's definition of citation faithfulness requires the cited document to belong to the retrieved context, support the statement and causally influence its generation. Checking support addresses one condition of this definition.

The authors illustrate the distinction with a question about Germany's capital. A document stating that Berlin is the capital can support an answer naming Berlin. The model may use that document's information, or it may produce the answer from knowledge stored in its parameters and then find matching support. The paper calls the latter behavior post-rationalization. Both processes can produce the same visible answer and citation.

How do the experiments test source reliance?

In the paper's post-rationalization experiment, BM25 retrieves 30 documents, a retrieval method called ColBERT v2 reranks them, and the top 5 enter Command-R+'s context with the question. The model generates an answer with citations.

The researchers then select short statements from those answers and append them to other documents. Examples include the phrases “Emperor penguin” and “The Choice.” These insertions supply matching words without the factual context needed to support the claim. The altered documents are random documents, relevant documents previously left uncited, or documents previously cited for other information.

The model generates answers again. The researchers check whether the original statement reappears and whether the model now cites the altered document for it. Figure 3 reports the following counts; the percentages use cases where the original statement reappears as their denominator.

The 57% result therefore describes the relevant-but-uncited experimental condition. The authors interpret the citation changes as evidence of post-rationalization. Their test assumes that inserting irrelevant material leaves the underlying answer-generation process unchanged; verifying that assumption requires inspecting internal model states.

What materials support checking the experiment?

The authors' repository instructions identify the notebook attribution_analysis.ipynb. Its “Post-Rationalization Test Dataset” section creates these files:

post_ration_random_data.jsonl
post_ration_rel_uncited_data.jsonl
post_ration_cited_other_data.jsonl

The instructions then run model inference on each file with qnd.py and evaluate the resulting predictions in the notebook's “Post-Rationalization Test Eval” section. This workflow makes the intervention concrete: change the supplied documents, generate fresh predictions and compare the citations attached to matching statements.

Related terms

• Citation correctness
• Citation faithfulness
• Retrieval-augmented generation (RAG)
• Answer attribution

FAQ

Can a citation support a claim while misrepresenting source use?

Yes. In the paper's Berlin example, the same supporting document can accompany either a document-based answer or an answer drawn from the model's stored knowledge. Support alone leaves that distinction unresolved.

How should a website owner interpret the 57% figure?

Treat it as a result from the paper's Command-R+ intervention study, specifically the relevant-but-uncited document condition. When reviewing a citation to your page, checking the supporting passage evaluates correctness; establishing causal reliance requires examining the answer-generation process.

Sources

1. arXiv, "Correctness is not Faithfulness in RAG Attributions", 23 December 2024
2. Wallat et al., "Correctness is not Faithfulness in RAG Attributions", full text, 23 December 2024
3. Jonas Wallat, "RAG Attributions", undated

---

Check how a page reads to answer engines. Open the GEO score analyzer