When a language model answers a factual question, it increasingly reads documents fetched from outside its weights, a setup called retrieval-augmented generation (RAG). If those documents are out of date, the model repeats the stale fact with full confidence. This note starts from one observation: different language editions of the web go stale at different speeds, and the pattern is predictable. English Wikipedia may record a new Bangladeshi cabinet minister weeks after Bengali Wikipedia does. Bengali Wikipedia may still list a European club's coach from two seasons ago. Systems that route non-English questions through English evidence, or that trust whichever passage is longest, therefore produce a specific and avoidable kind of hallucination: confidently giving the old value of a fact that has already changed.
We propose to measure this cross-lingual update lag directly from the revision histories of Wikipedia language editions. We then use it to decide, at answer time, which retrieved passage is most likely current, and to abstain when no passage probably is. In two sentences: the probability that a retrieved passage is stale depends on both how often that kind of fact changes and how quickly that language edition tends to absorb changes about that region. A reader that weighs multilingual evidence by this probability should hallucinate stale answers less often than one that trusts English, trusts the newest timestamp, or ignores time altogether.
Core Claim
Current work treats time and language separately. On the time side, Mitigating Temporal Misalignment by Discarding Outdated Facts predicts how long a fact stays true and lowers the model's confidence when the evidence is older than that lifespan. It reports a 50–60% reduction in calibration error on SituatedQA. Its limits are that it works only on English, it ignores when a fact began to be true, and it assumes a lifespan that does not depend on cultural context. On the language side, XOR QA: Cross-lingual Open-Retrieval Question Answering shows that searching English Wikipedia raises answerability for Bengali questions by 40%. That result motivates cross-lingual retrieval, but it uses static dumps, so it cannot see which edition is fresher. XRAG: Cross-lingual Retrieval-Augmented Generation comes closest: it tests post-cutoff news in five languages. It finds that the main difficulty is combining evidence scattered across languages, not generating non-English text. XRAG does not, however, model which language's evidence is current.
Our claim is that staleness is a joint property of the fact type, the language edition, and the region the entity belongs to. Modeling all three together gives better answer-or-abstain decisions than modeling any one alone. The new pieces are three: a measured lag distribution for each (edition, region) pair, a validity estimate for each passage built on that distribution, and an arbiter that uses those estimates to choose among conflicting multilingual passages or to decline to answer.
Research Program
The first component is a measurement study and benchmark. We will mine revision histories for facts that change over time and can be aligned across editions through Wikidata identifiers. Examples are office holders, team coaches and rosters, company leaders, and record holders. For each change event we record the date the new value first appears in each of roughly eight language editions. The languages will include those used by XOR-TyDi QA and XRAG, so the benchmark connects to existing evaluations. From these timestamps we estimate a lag distribution for each pair of edition $\ell$ and entity region $g$.
The first scientific question is purely descriptive: does lag vary enough across (edition, region) pairs to matter? If lags turn out roughly uniform, the project reduces to language-blind duration modeling, and we will report that result. If lags vary substantially, we build queries pinned to time points at which editions disagree. Each query comes with a gold current answer and a gold "stale" answer, so we can count stale hallucinations separately from other errors.
The second component is the validity model. Let $t_q$ be the query time, and let $t_a$ be the time the value stated in a retrieved passage $x$ became true. That start time is the quantity the earlier duration work ignored. Let $h_r(s)$ be the rate at which facts of relation type $r$ change at time $s$, estimated from the same revision data. Let $F_{\ell,g}(\Delta)$ be the probability that edition $\ell$ absorbs a change about a region-$g$ entity within delay $\Delta$. The probability that the passage is stale is the probability that the fact changed at some moment after $t_a$ and the edition has not yet caught up:
$$ P_{\text{stale}}(x, t_q) = \int_{t_a}^{t_q} h_r(s)\,\big[1 - F_{\ell,g}(t_q - s)\big]\,ds . $$
At test time the system does not see fact-level edit dates. It sees only page-level metadata and the passage text. We will therefore train a small multilingual predictor to estimate $t_a$, $r$, and $g$ from the passage, and plug those estimates into the equation. To estimate the fact-change rate $h_r$, we reuse the cloze-style duration predictor from Mitigating Temporal Misalignment by Discarding Outdated Facts, which serves as our method source here. Facts that rarely change, such as birthplaces, get a change rate near zero, so their passages are treated as safe regardless of lag. This directly addresses the tension raised by KDD Cup Meta CRAG 2024 Technical Report: Three-step Question-Answering Framework: that system answers from internal knowledge first and retrieves only when the model's self-consistency is low. RealTime QA: What's the Answer Right Now? shows that this self-confidence is unreliable on facts that evolve. The validity model offers an outside signal for when the model's own answer cannot be trusted.
The third component is the arbiter. We retrieve from all editions with a multilingual dense retriever, group passages by the value they assert, and score each value by its passages' validity mass, $\sum (1 - P_{\text{stale}})$. The reader is shown the top value's supporting evidence and generates its answer in the query language. If the top value's validity mass falls below a threshold, the system abstains. This targets the failure documented by RealTime QA: What's the Answer Right Now?, where models fall back on outdated knowledge when retrieval fails instead of admitting uncertainty.
We tune the abstention threshold on a rolling recent window rather than a fixed calibration set. The reason is that time drift breaks the exchangeability assumption behind the conformal guarantees in TRAC: Trustworthy Retrieval Augmented Chatbot. TRAC wraps RAG answers in prediction sets with coverage guarantees on static Natural Questions; our setting is non-stationary. Multimodal entity grounding, of the kind studied by EntRAG: Entity-Centric Retrieval-Augmented Generation for Knowledge-based Visual Question Answering, is a natural later extension, but we keep it out of scope so that the time-and-language effect can be isolated.
Baselines
The naive reference point is closed-book answering in the query language with the current date in the prompt. RealTime QA: What's the Answer Right Now? reports 44.9% multiple-choice accuracy for closed-book GPT-3 on current news, against 66.5% with Google Search. It also finds that adding the date can hurt closed-book accuracy, so we will run this baseline both with and without the date. A second simple point is native-only retrieval, the monolingual constraint that XOR QA argues against.
The English-pivot baseline follows the idea behind CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generation: reason and answer in English, then translate. That paper reports hallucination reductions of up to 62% in Chinese, Malay, and Tamil, but it uses no retrieval. To make it competitive here, we add English retrieval. This baseline is the one most exposed to regionally stale English evidence.
The strongest existing approach to compare against is all-language retrieval read by a strong LLM given the query date, which is the multilingual setting of XRAG. On its own benchmark, XRAG: Cross-lingual Retrieval-Augmented Generation reports 62% accuracy for GPT-4o in the English setting, with further drops cross-lingually. Two further baselines are a newest-timestamp-wins heuristic and a language-blind version of our validity model, in which the edition-and-region lag factor is removed. That version is effectively the duration method from Mitigating Temporal Misalignment carried over to multilingual retrieval.
The language-blind version is the ablation that isolates our key ingredient. If removing the lag factor costs nothing, the core claim fails. All systems share the same retriever, index snapshot, reader model, and hyperparameter search budget, and every abstaining system is compared at matched coverage. To beat the strongest baseline, our method must reduce the stale-answer rate at equal or higher selective accuracy, by a margin larger than the variance across random query-time samples. This must hold both on our benchmark and on XRAG's time-sensitive subset.
Why not just…?
Why not just take the most recently timestamped passage? This is the real threat, and the experiments will settle it rather than argue around it. The heuristic fails exactly when document dates do not match fact dates. A Wikipedia page edited yesterday to fix formatting may still list a coach who left two seasons ago. A recent news article may recap an old office holder. Our benchmark records fact-level change dates, so we can split queries by whether the page timestamp is informative about the fact. We will report the results honestly. If newest-wins matches our method on news-style evidence such as XRAG, we will narrow the claim to encyclopedic sources, where page timestamps and fact timestamps come apart.
Why not just retrieve from every language and let a strong model sort it out? XRAG's main finding is that this kind of cross-lingual evidence synthesis is precisely where strong models fail: GPT-4o at 62% in English with lower scores cross-lingually, and relative drops above 50% for Arabic in some models. Our arbiter does not replace the reader. It gives the reader a smaller, agreeing set of evidence. The XRAG-style baseline measures directly whether that helps.
Why not just pivot through English? English often has more facts but not newer ones about local entities. XRAG also reports failures of response-language correctness, with Mistral-Large answering in English on over 12% of German cases. Our time-pinned queries about regional entities whose native edition updated first are designed to expose this baseline. If English pivoting wins even on those queries, then measured lag must be small, which the descriptive study will have shown first.
Why not just apply conformal abstention as in TRAC? TRAC's guarantees assume calibration and test queries are exchangeable, and its prediction sets already grow to 20.4 unique answers at $\alpha=0.1$ on static Natural Questions. We include conformal abstention, calibrated once at the start, as a baseline. We expect its realized coverage to drift as query time moves away from the calibration window, and we will measure that drift.
Proposed Experiments
We will run on four benchmark groups: the new revision-history benchmark; XRAG's post-cutoff news in Arabic, Chinese, German, Spanish, and English; XOR-TyDi QA, where the best reported end-to-end XOR-Full system reaches 18.7 average F1; and SituatedQA. On SituatedQA, Mitigating Temporal Misalignment reports 20.5 Exact Match while retrieving for only 45% of queries, which serves as an English-only sanity check that our language-blind ablation reproduces prior behavior. No reported result exists on Lacuna for any method on the new cross-edition benchmark, since it does not yet exist.
The metrics are accuracy, a stale-answer rate counted against the gold outdated value, selective accuracy against coverage, expected calibration error of the validity scores, and response-language correctness.
Three checks decide the project. First, the measurement study must show that lag differs meaningfully across (edition, region) pairs, judged by comparing the fitted lag distributions; otherwise the language factor has nothing to explain. Second, the full model must beat its language-blind ablation on stale-answer rate, specifically on queries where editions disagree. Third, it must beat newest-timestamp-wins on the subset where page timestamps do not reflect fact changes. If all three hold, staleness in multilingual retrieval is a modeled quantity rather than noise. If the first fails, that is still a useful negative result: it means English-pivot systems are not systematically stale on local facts.
References
- Mitigating Temporal Misalignment by Discarding Outdated Facts
- XOR QA: Cross-lingual Open-Retrieval Question Answering
- XRAG: Cross-lingual Retrieval-Augmented Generation
- KDD Cup Meta CRAG 2024 Technical Report: Three-step Question-Answering Framework
- RealTime QA: What's the Answer Right Now?
- TRAC: Trustworthy Retrieval Augmented Chatbot
- EntRAG: Entity-Centric Retrieval-Augmented Generation for Knowledge-based Visual Question Answering
- CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generation