SentX SI Blog Meet Victoria

My RAG pipeline cites a source for every answer — how do I tell if the answer actually came from that document, or from the model's memory?

· 5 min read

SentX SI

A citation carries two claims at once: that the passage supports the sentence, and that the passage produced it. Most verification tooling tests only the first. A paper submitted to arXiv on 12 November 2026, Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation, by Tanzeel Abbas, Tehseen Zia and David Windridge of COMSATS University Islamabad and Middlesex University London, measures the second — and finds it failing in a way support checking can never see. Across five sizes of Qwen2.5-Instruct, from 0.5B to 14B, plus Llama-3.1-8B-Instruct, between 1.2% and 15.7% of citable claims on their factoid benchmark derive from the model's parametric memory rather than the retrieved context.

Source: Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (arXiv:2611.17161)

The reason standard verification is blind is structural, not a matter of a weak verifier. When a model answers from memory, it still picks a citation, and it picks one that agrees with what it has just written. The cited passage then entails the claim by construction, so an NLI checker reading the pair sees exactly the support it was asked to find. The authors run a DeBERTa entailment model in its strongest form — maximum entailment over all retrieved passages — and get results indistinguishable from chance on every model tested, repeated with three further entailment checkpoints. Restricted to the cited passage alone, one 7B model scores memory-derived claims as better entailed than context-derived ones: the contradiction sits in the text the verifier reads, and it does not register.

Source: Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (full text)

So how do you measure provenance at all? The paper's protocol is counterfactual value substitution. Find a value that appears in both the answer and the retrieved passages, replace it throughout those passages, titles included, with a false variant, and ask the question again. A model that reports the false value was using the context; one that repeats the original was using memory. The label is constructed by the intervention rather than inferred from internal states, which is what makes it measurable at scale. Rates climb with model size within the Qwen family, and family matters at least as much: Llama-3.1-8B sits at 15.7%, above any Qwen model tested, including one twice its size. At the most conservative substitution distance, the two models checked there give 4.6% and 8.5%. One artifact warning transfers directly to anyone copying the protocol: the authors' first version edited passage bodies but not titles, and a value left visible in an unedited heading inflated the measured rate roughly twofold. Every surface carrying the value has to be edited, and any case where the value survives in the question has to be discarded.

Source: Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (full text)

The most useful result for an operator is the negative one about detectors. Comparing ten attribution methods on identical claims, the best signal is the contrast between the model's token-level distribution with and without context, measured at the disputed value itself, at a mean AUROC of about 0.79. But it measures what the model knew, not what it used: restricted to claims the model could already produce without context, it falls to 0.65. And at base rates of 1.2% to 15.7%, even that ranking does not pay as a flagging policy. Precision peaks at 0.45, below the 0.5 break-even point, so every automatic flag costs more provenance honesty than it buys. Deletion-based methods sit between chance and the token contrast: leave-one-out reaches AUROC 0.54 to 0.67, and the published ContextCite system scores 0.51 on the 7B model.

Source: Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (full text)

That last line is where my own suggestion comes in, and it is mine, not the paper's. If a subset of your answers is genuinely consequential — compliance text, published content, anything downstream that acts on the claim — a black-box check is worth building. Take one of those answers, re-run the same query with the cited document removed or swapped for a neutral passage that does not contain the disputed fact, and diff the two outputs. An identical answer is a decorative citation: the document was not doing the work, and that one goes to a human reviewer. Be clear about what this does and does not buy. It is coarser than the paper's measurement and catches a subset of the problem — the cases where the answer changes not at all. Where the model partially used the context and partially used memory, the diff will be partial rather than clean, and where the evidence is redundant across passages, or the model already knew the value, removing one document changes nothing. The paper's own finding that deletion measures a passage's contribution to a claim's likelihood, rather than the claim's origin, explains that ceiling. Use it as a targeted manual-review trigger on your highest-stakes outputs, not as a pipeline-wide detector. It does have one property the paper's best signal lacks: it needs only regenerated text, no token-level log-probabilities, so it works against closed-model APIs where the distribution-contrast method is simply unavailable. And when you remove or swap the document, apply the title-leak lesson — confirm the disputed value is not still visible in another retrieved passage, a metadata field, or the question itself.

Source: Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (full text)

The honest position this leaves you in: support checking is necessary but not sufficient, the paper's best detector is not deployable at these base rates, and separating what the model knew from what it used under black-box access is, by the authors' own account, the open problem. What you can do is bound the exposure. If you run open-weight models, the substitution protocol tells you how large the problem is in your domain — with the caveat that it only reaches claims containing a numeric value shared between answer and passages, roughly 70% of the factoid answers in their main dataset and far fewer in long-form settings. If you run closed models, the leave-one-out check on consequential answers is the audit you can actually operate, and the fraction of decorative citations it surfaces is a real number you can trend over time.

Source: Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (arXiv:2611.17161); Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (full text)

The same instinct appears in one line of SentX's own model card, which is a first-party product disclosure rather than independent certification. Under known limits, it tells users that answers can contain incorrect facts or citations, and to open the cited sources and check the supporting passage. That is the reader's habit, applied to one chat answer at a time. The value of this paper is showing what the habit finds at the pipeline level: the passage agrees with the answer, and that was never the question. The check that matters is the one you run yourself, and the citation that survives it is worth what it says.

Source: Victoria AI Model Card | Capabilities, Evidence and Limits

Sources

  1. Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (arXiv:2611.17161) — arXiv
    Supported claims and limits
    • Submitted 12 Nov 2026 by Tanzeel Abbas, Tehseen Zia and David Windridge
    • 1.2% to 15.7% of citable claims derive from parametric memory rather than the retrieved context, across five sizes of one model family plus a second comparator family
    • NLI verification gives results indistinguishable from chance across models, even restricted to the cited passage
    • The cause is structural: a model answering from memory selects a citation that agrees with its output
    • The best-separating signal does not improve provenance honesty at these base rates, and separating what the model knew from what it used remains the open problem
  2. Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation (full text) — arXiv
    Supported claims and limits
    • Protocol: a value shared by answer and passages is replaced with a false variant, titles included; reporting the false value indicates context use, repeating the original indicates memory
    • Qwen2.5-Instruct 0.5B-14B and Llama-3.1-8B-Instruct on ASQA with GTR top-5 retrieval and 800 questions per condition; ELI5 and QAMPARI assessed on Qwen2.5-7B
    • Llama-3.1-8B reaches 15.7%, above any Qwen model including one twice its size; at the most conservative substitution distance the two models give 4.6% and 8.5%
    • DeBERTa NLI in its strongest form (maximum entailment over all passages) is at chance on every model, repeated with three further checkpoints; on Qwen2.5-7B restricted to the cited passage, parametric claims score higher mean entailment than contextual claims
    • Title-leak artefact: the first version edited passage bodies but not titles, and uncorrected numbers overstated the phenomenon roughly twofold; every surface carrying the value must be edited, and cases where the value survives in the question are discarded
    • Ten attribution methods compared on identical claims; token-level with/without-context contrast at the disputed value gives mean AUROC 0.79, falling to 0.651 restricted to claims the model could produce without context
    • Precision peaks at 0.45, below the 0.5 break-even, so flagging does not pay; leave-one-out reaches AUROC 0.541 to 0.667; published ContextCite scores 0.507 on Qwen2.5-7B
    • Coverage is limited to a numeric value shared between answer and passages: roughly 70% of ASQA answers but 16.5% and 21.8% of ELI5 and QAMPARI; evaluation covers open-weight models 0.5B-14B; the method requires token-level log-probabilities that closed frontier models do not expose; 19%-24% of records were discarded as undetermined
  3. Victoria AI Model Card | Capabilities, Evidence and Limits — SentX SI
    Supported claims and limits
    • The model card is a first-party product disclosure, not an independent certification
    • Under known limits, the card states answers can contain incorrect facts, citations or calculations, and instructs users to open cited sources and check the supporting passage
Meet Victoria