Many deployed question-answering systems search in two places. They search a large, fresh source such as the web or a recent crawl, because that is where answers are easy to find. They also have to cite a smaller, authoritative collection, such as a frozen archive, an enterprise document store, or a regulator's corpus, because that is what counts as proof. When the world changes, these two sources disagree about when a fact held. The fresh source says the current officeholder is B. The authoritative archive only ever mentions A. We propose to study a specific error that appears in exactly this gap. A retrieval-augmented language model (an LLM that answers from passages retrieved at query time) states B and cites an archive document that actually says A. We call this a temporally laundered citation: the citation looks authoritative, but it supports a different, older value than the one asserted. The reverse also happens. The model states its stale memorized value A and cites a current document that says B.
The opportunity is that the older setting of "answer projection" already had the right frame. Early TREC systems discovered an answer on the web and then had to find a supporting document in the restricted AQUAINT collection. They treated any failure to find one as "unsupported." That binary verdict discards information once time enters the picture. If the archive attests A as of 2018 and the web attests B as of 2021, the honest output is neither "B, citing the archive" nor "unsupported." It is "B per the fresh source; the authoritative corpus last attests A, dated 2018." We will build a benchmark that isolates this three-way situation, measure how often current systems launder citations, and test a simple fix. The fix retrieves archive evidence by the question's relation slot rather than by the answer string, and it labels each archive citation with the value and date it actually supports.
Core Claim
In two-corpus retrieval under knowledge change, a substantial share of citations to the authoritative corpus point to documents that attest a different temporal state of the same fact. Neither answer-string projection nor standard single-pass citation generation detects this. We claim two things. First, retrieving archive passages by relation slot (the entity plus the property asked about) and then classifying each passage as supporting the candidate value, attesting a different value at a stated date, or being silent will sharply cut laundering. Second, it will raise the rate at which systems correctly surface the archive's last attested state, without lowering accuracy on the current answer. A secondary, mechanistic claim tests a tension in the literature. The white-box citation signal that marks valid citations by high internal recall should misfire on the stale-memory variant of laundering, because in that case the model's internal memory agrees with the wrong answer.
Where This Sits in Prior Work
The projection framing comes from the TREC-era systems. Integrating Web-based and Corpus-based Techniques for Question Answering mined candidates from the web and projected them into AQUAINT, placing 6th of 25 groups. Lin et al. and Qiu et al. followed the same discovery-then-verify pattern. Huang et al. and Jijkoun et al. documented the bottleneck that arises when the target corpus phrases the fact differently or lacks it. These papers supply our task definition. They all treat both corpora as static and search the target for the answer string. That search is guaranteed to fail when the target holds only an older value, which is the case we study. The verification-first variants of Ittycheriah et al. and Adafre et al. rerank local candidates with web agreement. Under change, that agreement signal would actively penalize a correct historical attestation.
The motivation for the temporal side comes from three results. SituatedQA shows that about 16.5% of information-seeking questions change answers over time. Its models stay 15 points less accurate on updated facts even when given a newer corpus. RealTime QA shows that when retrieval is incomplete, models fall back on stale memorized answers. Mitigating Temporal Misalignment by Discarding Outdated Facts discounts memorized confidence using predicted fact duration. None of these asks what a system should cite when the only authoritative evidence is outdated.
The closest competitor is EvoTrustRAG. It classifies conflicts among retrieved passages as evolution, intervention, or unresolved, and it preserves historical versus current states. It assumes that the conflicting passages are already in one context window, and it has no notion of which corpus is authoritative for citation. Our proposal differs on both points. The hard step is finding the old-state passage in the archive when the new answer string is absent. Citation correctness is also judged against a designated provenance corpus. We will run EvoTrustRAG on our retrieved evidence to test whether its conflict attribution alone closes the gap.
On the hallucination side, Self-RAG trains a model to critique whether its output is supported by retrieved passages. However, it judges support without time: a 2018 passage saying A does not "support" B, and Self-RAG has no way to say so usefully. FACTUM detects hallucinated citations from internal signals. It finds that valid citations carry higher feed-forward "parametric force," meaning stronger activation in the layers thought to store memorized facts. That finding constrains our design. It suggests that internal-signal detectors may be blind to laundering driven by stale memory, and we test this directly.
Research Program
We build three things.
The first is the benchmark. We use the temporal questions from SituatedQA, whose answers carry validity intervals, together with the 2018 and 2021 Wikipedia snapshots from its temporal experiments. We designate one snapshot as the authoritative provenance corpus and the other as the discovery corpus. This yields four instance types. In the stable type, both corpora give the same answer. In the updated type, the archive holds the old value and discovery holds the new one; this is where laundering of the first kind can occur. In the stale-memory type, the archive is current but the base model's memorized answer is old; this is where laundering of the second kind can occur. In the unattested type, the archive is silent. We add a live-freshness split from RealTime QA. There the provenance corpus is the 2018 Wikipedia DPR index used in that paper, and discovery is the weekly news and search results. The gold label for each instance is a triple: the current value, the archive-attested value with its date, and the archive document that attests it. The last two elements are recovered by annotators through slot-based search, with a second annotator on a sample.
The second is the method, and it is intentionally light. Discovery-corpus retrieval proposes a candidate current value. Archive retrieval uses a query built from the question's entity and relation, not the candidate string. For example, it searches for "head of the ministry, country X" rather than "B." Each retrieved archive passage is then labeled by an entailment-style classifier conditioned on the candidate value and the passage date. The three labels are supports the value, attests another value for the same slot (with that value and date extracted), or silent. The generator is constrained to emit each archive citation together with its label. It can therefore write "B [discovery]; archive last attests A as of 2018 [archive doc]" but cannot attach an archive citation to B unless the label is "supports." When the archive is silent and discovery is weak, the output is an abstention.
The third is the mechanistic study. On open 8B instruction models, we compute FACTUM's four scores at each citation token. We compare the scores' ability to separate valid from laundered citations across the updated and stale-memory types.
Baselines
The naive reference point is standard RAG over the provenance corpus alone, with inline citations. It shows how much staleness and laundering a system produces when it never sees fresh evidence. RealTime QA reports 13.3 exact match (EM, the share of answers that match the gold string exactly) for GPT-3 with a 2018 Wikipedia DPR index, against 34.6 EM with Google Custom Search. That gap is the freshness cost this baseline pays.
The second baseline is classic answer projection in the style of Katz et al.. It discovers a candidate value in the fresh corpus, searches the archive for that string plus question keywords, cites the best hit, and otherwise marks the answer unsupported. This baseline isolates what slot-based retrieval adds.
The third baseline is two-corpus RAG with Self-RAG-style support critique. Passages from both corpora go into the prompt, and a supported/not-supported check filters citations. This is the strongest general-purpose grounded generator in the file. The file gives no per-benchmark Self-RAG numbers, only the qualitative claim that it surpasses retrieval-augmented ChatGPT and Llama2-chat.
The strongest competitor is EvoTrustRAG given the same passages our method retrieves from both corpora. On benchmark-native conflicts it reports 81.4% average accuracy against 80.2% for its strongest baseline, and it raises historical accuracy from 74% to 79.1%. Because it explicitly models evolution, it is the method most likely to already produce correct "current versus historical" answers.
We also run the adaptive trigger of Zhang and Choi, which reports 20.5 EM on its SituatedQA subset while retrieving for 45% of queries. It is the representative of duration-based staleness handling.
Fairness rules apply to every comparison. All systems use the same generator backbones, the same two indexes, the same number of retrieved passages per corpus, and the same prompt-tuning budget on a shared development split. EvoTrustRAG is not scored on our retrieval quality alone. It receives our slot-retrieved passages in one arm and answer-string-retrieved passages in another.
Two ablations isolate our ingredients. One replaces slot retrieval with answer-string retrieval while keeping the typed labels. The other keeps slot retrieval but uses a plain supports/does-not-support check with no "attests another value" label.
Our headline metric is provenance-correct accuracy: the current value is right and every archive citation is labeled with the value it truly supports. No reported result was found on Lacuna for this metric or for laundering rate on any benchmark, so all systems are measured fresh. Beating the field means two things: lower laundering than EvoTrustRAG given identical passages, and higher archive-attestation recall than answer-string projection. Current-value EM must stay within noise of the best baseline.
Why not just…?
Why not just cite the fresh source and drop the authoritative corpus? In many settings this is forbidden, not merely suboptimal. The citation requirement exists because only the archive counts as proof. This is the premise of the TREC provenance rules and of enterprise and legal retrieval. Where it is allowed, it is the right answer. Our updated-type instances report discovery-only accuracy separately so readers can see what is lost and gained. The proposal does not claim that archive citations make answers more accurate. It claims they are frequently misrepresented.
Why not just update the authoritative corpus? Authoritative collections are often frozen by policy, and re-indexing does not by itself fix temporal errors. SituatedQA finds models remain 15 points worse on updated facts even with a newer snapshot. AURORA makes updates cheaper but does not tell the generator which of two dated states a citation supports. Our stale-memory instance type is exactly the "updated corpus" case. If laundering there stays high for baselines, updating alone is shown to be insufficient.
Why not just run an entailment filter on citations? This is the real threat. An off-the-shelf entailment model would mark "2018 archive says A" as contradicting "B" and drop that citation. That removes laundering at almost no cost. What it does not do is find the archive passage in the first place when retrieval searches for "B." It also discards a correct historical attestation instead of reporting it. The ablation that keeps slot retrieval but uses a binary support check tells us how much the third label adds. If binary filtering plus a fixed template matches the full method, we will report that the contribution is the retrieval change and the benchmark, not the verifier.
Why not just use EvoTrustRAG? If EvoTrustRAG, given our slot-retrieved passages, matches our laundering and attestation numbers, then conflict attribution already solves verification. Our remaining contribution would be the slot-based projection step and the evaluation. The matched-passage arm settles this directly.
Decisive Checks
Three results would confirm or refute the proposal. First, we measure laundering rates on the updated and stale-memory types for all baselines. If laundering is rare (say, under a few percent of archive citations) for the naive two-corpus RAG baseline, the problem is not worth a method, and we would publish the benchmark as a negative finding. Second, on the updated type we compare archive-attestation recall for slot retrieval against answer-string retrieval. This is the direct test of the projection claim, and it should show a large gap, since answer-string search cannot by construction retrieve a passage lacking the new value. Third, we compute FACTUM's AUC separately on laundered citations of each type and compare it to the 0.737 peak reported on TREC NeuCLIR. If parametric force separates valid from laundered citations on the updated type but drops toward chance, or inverts, on the stale-memory type, that confirms that internal-recall detectors mistake stale memory for grounding. Such a result would limit their use in any system facing knowledge change. Citation judgments throughout are made by humans on a stratified sample and by an LLM judge on the rest, with agreement reported in the same way FACTUM reports its kappa of 0.68.
References
- Integrating Web-based and Corpus-based Techniques for Question Answering
- Extracting Answers from the Web Using Data Annotation and Knowledge Mining Techniques
- FDUQA on TREC 2007 QA Track
- FDUQA on TREC2003 QA task
- The University of Amsterdam at QA@CLEF2003
- IBM's Statistical Question Answering System-TREC 11
- Dublin City University at QA@CLEF 2008
- SituatedQA: Incorporating Extra-Linguistic Contexts into QA
- RealTime QA: What's the Answer Right Now?
- Mitigating Temporal Misalignment by Discarding Outdated Facts
- EvoTrustRAG: Evolution-Aware Conflict Attribution and Evidence Handling for Reliable Retrieval-Augmented Generation
- Self-RAG: Self-reflective Retrieval Augmented Generation
- FACTUM: Mechanistic Detection of Citation Hallucination in Long-Form RAG
- AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems