Reasoning models that write out long step-by-step chains before answering do well on clean inputs. When their context contains a misleading passage, they use that chain to build a plausible argument for the wrong answer. A retrieved news article that was true in 2018 and is false today is exactly such a passage, because it is relevant, fluent, and confidently stated. We propose a cheap admission step that sits between reading the retrieved documents and reasoning over them. A small model turns each passage into dated claims of the form "subject, relation, value, as of date". A deterministic rule then discards claims that have probably expired or been superseded. The expensive reasoning mode runs only on whatever survives, and only when something survives.
The bet is that stale context should be treated as an admission problem, not a reasoning problem. A reasoning chain is a poor place to notice that a premise has expired: once the premise is in context, more thinking tends to rationalize it rather than reject it. Interval arithmetic on dates, by contrast, is trivial for a symbolic rule and error-prone for a language model. The proposal splits the work this way: neural reading handles the messy text, symbolic logic handles validity over time, and the reasoning budget is spent only where a consistent set of premises exists.
Core Claim
On time-sensitive questions whose retrieved context mixes current, outdated, and irrelevant passages, gating evidence by predicted fact lifespan before reasoning will reduce stale-answer errors more than three alternatives: instructing the model to check dates, sorting by recency, or pruning by semantic relevance. Gating should also shorten the chains that reasoning models produce when they rationalize. The specific new ingredient is per-claim temporal admission that feeds a gate on reasoning effort. Prior work has used fact lifespans to decide whether to retrieve at all. Prior work has also pruned documents by topical relevance. No prior work in the file has used lifespans to decide which retrieved premises a reasoning model may see.
What Prior Work Establishes
Three papers motivate the gap. Thinking in a Crowd builds SciAux, a set of multi-hop science questions with helpful, irrelevant, and misleading snippets. It reports that Qwen3-8B falls from 96.22% on clean input to 77.03% with thinking on and 74.87% with thinking off when given two misleading snippets, and that rationalizing responses are the longest. Its own numbers show that thinking does not protect against misleading premises, although they do not show that thinking is strictly worse. Our claim therefore rests on the drop and the token inflation, not on thinking-on being worse than thinking-off. RealTime QA finds that retrieval failures account for 34% of errors on weekly news questions, that models then fall back on outdated knowledge, and that simply prepending the current date hurts closed-book models. Neither paper offers a remedy that operates on time.
One paper supplies the core method. Mitigating Temporal Misalignment by Discarding Outdated Facts trains a BERT model to predict how long a fact stays true, using 13 duration classes from seconds to centuries. It retrieves only when the elapsed time exceeds the predicted lifespan, which reaches 20.5 exact match (EM) on SituatedQA while retrieving on 45% of queries. We reuse its duration predictor, but apply it to each retrieved claim rather than to the model's internal answer. We also add a supersession rule, which addresses that paper's acknowledged blindness to when a fact began: a newer observation of the same subject and relation overrides an older one.
Two lines of work constrain the design.
- Relevance-based pruning. Superposition Prompting processes each document on its own parallel attention path and prunes low-saliency paths. This gives large speed and accuracy gains on Natural Questions, but saliency is purely semantic, so an outdated article that matches the question well survives. Distraction-aware answer selection (DAS) adds an "answer unanswerable" instruction and reweights documents by how likely they make the query. It reports a 64% average EM gain across open-domain QA sets, but it also has no notion of time.
- Routing and time-sliced models. SymRAG routes queries among symbolic, neural, and hybrid paths by query complexity over a static knowledge graph, and its pure symbolic path is almost never chosen (0–0.1%). That is a warning that a symbolic component must do something the neural path cannot, which is why ours does only date arithmetic. TiMoE masks out language models trained on time slices later than the query. It operates on whole two-year model slices, not on individual passages in context.
Research Program
The system has three parts.
The reader. A small instruction-tuned model reads each retrieved passage together with the question. It emits zero or more claims as tuples $(s, r, v, t_e)$: subject, relation, value, and the date at which the passage asserts the claim. That date comes from an explicit date in the text if present, otherwise from the publication timestamp. This is the perception side of the split. We train it on TimeQA-style distant supervision and measure its extraction error separately, so that perception failures are never confused with reasoning failures.
The admission rule. A duration model $\hat d(s,r)$, the Zhang and Choi predictor fine-tuned on TimeQA and TimePre, estimates each relation's lifespan. With query time $t_q$, a claim $c$ is admitted when
$$ t_q - t_e(c) \le \hat d(c) \quad\text{and}\quad \nexists\, c' : (s',r') = (s,r),\ t_e(c') > t_e(c),\ v' \neq v , $$
That is, the claim has not outlived its predicted lifespan and no newer passage reports a different value for the same subject and relation. Passages are kept only if they contribute at least one admitted claim or contain no time-bound claims at all. The second condition keeps static background material, such as the science content in SciAux.
The compute gate. The surviving context decides how much reasoning to spend.
- If no relevant claim survives, the system abstains, using DAS-style abstention, rather than letting the model fall back on stale memory.
- If survivors agree, the model answers with thinking off.
- If survivors conflict, or the question needs multi-hop composition, thinking is turned on, and the remaining conflict is stated explicitly in the prompt.
This is conditional computation at the query level. Its expected payoff is fewer thinking tokens on easy cases and no rationalization fuel on hard ones.
The analysis traces every error to one of three causes: the reader extracted the wrong claim or date, the duration model mispredicted the lifespan, or the reasoner erred given correct premises. We separate them by substituting oracle claims and oracle durations one at a time.
Baselines
Naive retrieval. The do-nothing reference is plain retrieval-augmented generation: all retrieved passages are concatenated, and Qwen3-8B answers with thinking on and with thinking off. This matches the Thinking in a Crowd setup, so the SciAux numbers above (96.22% clean, 77.03% thinking on, 74.87% thinking off with two misleading snippets) anchor our reproduction. Beating it means recovering most of that gap on stale-distractor conditions.
Retrieval gating by fact lifespan. This is the adaptive inference method of Mitigating Temporal Misalignment, reported at 20.5 EM on SituatedQA with retrieval on 45% of queries. It is the strongest existing temporal method because it already uses lifespans. We run it with the same retriever, reader model, and duration model, so the only difference is where the lifespan test is applied. Beating it means higher EM on SituatedQA under injected outdated passages, which its design does not filter.
Relevance pruning. Superposition Prompting and DAS are the strongest non-temporal competitors. We run both with the same retrieved sets and Qwen3-8B backbone, with their thresholds tuned on the same development split and the same tuning budget as ours. No reported result was found on Lacuna for either method on TSQA, SituatedQA, or SciAux, so all numbers on these sets are ours.
Time-sliced model. TiMoE is the reference on TSQA. The file reports only relative figures: up to 15% fewer future-knowledge errors and roughly 3% above 3B models. No absolute TSQA accuracy is reported in the file. We will report TSQA results by distractor type and compare against Llama 3.2 3B and Qwen 2.5 3B, run ourselves.
Ablations. Three ablations isolate the key ingredient.
- Neural admission: the symbolic rule is replaced by the same small model prompted to judge each claim "still valid?".
- Recency only: lifespans are dropped and only the supersession test is kept.
- Always think: the compute gate is removed.
Why not just…?
Why not just tell the reasoning model to ignore outdated documents? This is the cheapest alternative and a real threat. It also matches the failure mode documented by Thinking in a Crowd: models that rationalize misleading premises are not reliably talked out of them by their own chains. We run it with the exact date of each passage and the current date in the prompt. If it matches our system on TSQA past-distractors and SituatedQA, the symbolic stage is unnecessary, and we will report that.
Why not just keep the most recent document? Recency gets supersession right but fails on two fronts. It discards static facts that remain true from old sources. It also keeps fresh but irrelevant or misleading passages, which are the bulk of TSQA's "irrelevant" and SciAux's "misleading" conditions. The recency-only ablation measures exactly how much the lifespan term adds beyond this heuristic.
Why not just turn thinking off whenever retrieved context is present? SciAux shows this does not help: 74.87% with thinking off is no better than 77.03% with thinking on. Turning thinking off also forfeits the 10–13% clean-context gain that thinking provides. Our gate keeps thinking where premises conflict, and the always-think and never-think arms of the experiment show whether the gate improves on either fixed setting.
Why not just use a time-sliced model like TiMoE? It handles the model's own knowledge, not stale passages placed in context. It is also limited to two-year resolution, so it cannot help with weekly news in RealTime QA.
Proposed Experiments and Decisive Checks
The evaluation uses four benchmarks:
- TSQA, whose distractors are already labeled past, future, or irrelevant.
- SituatedQA, into which we inject the earlier-year answers as outdated passages.
- RealTime QA, for organic weekly drift; the reported anchor is 66.5% multiple-choice accuracy and 34.6 EM for GPT-3 with web search.
- SciAux (billhdzhao/SciAux), as a non-temporal control where admission should rarely fire and should not hurt.
The metrics are:
- accuracy or EM;
- the stale-answer rate, meaning the fraction of answers that equal the injected outdated value;
- abstention calibration, measured as expected calibration error;
- thinking tokens per query;
- the error attribution described in the research program.
Three results would decide the question. First, on TSQA past-distractors and injected SituatedQA, the stale-answer rate should fall clearly below prompted date-checking, recency-only, and lifespan-based retrieval gating. Second, the symbolic rule should beat neural admission at matched extractor quality. If it does not, the contribution narrows to "admit before reasoning", and the date arithmetic is not essential. Third, with gating on, thinking-on accuracy under stale distractors should move most of the way back toward clean-context accuracy, while mean thinking tokens drop. That would show the rationalization trap is closed by controlling what enters the context, not by thinking harder.
References
- Thinking in a Crowd: How Auxiliary Information Shapes LLM Reasoning
- RealTime QA: What's the Answer Right Now?
- Mitigating Temporal Misalignment by Discarding Outdated Facts
- Superposition Prompting: Improving and Accelerating Retrieval-Augmented Generation
- Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering
- SymRAG: Efficient Neuro-Symbolic Retrieval Through Adaptive Query Routing
- TiMoE: Time-Aware Mixture of Language Experts