Debating Which Views Have Gone Stale: Time-Aware Pluralistic Answers for Contested Questions

Language models freeze at a training cutoff. The field already treats this as a problem for facts: models keep naming the old office holder or the old score. It has not treated it as a problem for opinions. When a model is asked a contested social question, it describes the range of reasonable views as they stood when its training data was collected. Public debate keeps moving. A position that was mainstream can become marginal, a new position can enter respectable discussion, and many positions rest on factual premises that have since expired, such as a law that was repealed or a statistic that was revised. A model that presents a stale snapshot of public opinion as the current one misleads users quietly. The risk grows when the user already holds the outdated view and the model, trained to please, tells them it is still the mainstream position.

We propose to measure this "normative obsolescence" and to test one specific remedy. The remedy is a debate among several language-model agents. Each agent defends one perspective. A challenger agent cross-examines each perspective on two separate fronts: whether its factual premises still hold, and whether there is dated evidence that people still argue for it today. The output does not pick a winner. It returns every perspective with a status label: current, waning, emerging, or historical. Any perspective whose supporting facts have expired is flagged. In two sentences: existing debate-over-retrieval systems throw away disagreeing views as noise, and existing pluralism benchmarks treat public opinion as fixed. We build a benchmark where the range of reasonable opinion moves over time, and a debate protocol that labels views as stale rather than deleting them.

Core Claim

The claim is that facts inside a perspective and the value stance of that perspective should be handled differently when time passes. Factual premises can expire and should be checked against current evidence. Value stances do not become false, but they can lose or gain public standing, and that standing should be reported rather than adjudicated.

A single model reading retrieved documents tends to merge these two operations. It may drop a view whose premise expired without saying why, or keep presenting a view as current because nothing in the retrieved context contradicts it. A structured cross-examination that forces the two checks apart should produce better-calibrated status labels and fewer silent deletions. It should also be harder to sway by a user's stated stance, because the perspective agents and the judge never see that stance.

The claim is falsifiable. If a single retrieval-augmented model, given the same evidence and the same labeling instructions, matches the debate on status accuracy and premise flagging, then the debate adds nothing. In that case the contribution reduces to the benchmark.

Where This Sits in Prior Work

Several papers motivate the gap. Benchmarking Overton Pluralism in LLMs scores how many of the reasonable perspectives in public discussion a model actually covers. It finds that single frontier models cover 35–41%. The authors acknowledge that their perspective sets are static snapshots that must be refreshed as discourse shifts. Pluralistic Alignment Over Time formalizes how preferences change and how welfare should be aggregated across time, but it is purely conceptual and does not involve language models, retrieval, or training cutoffs. Who's in Charge? Disempowerment Patterns in Real-World LLM Usage shows two things from real consumer conversations: users give higher ratings to interactions that distort their sense of reality, and these interactions concentrate in personal and value-laden domains. That is the mechanism by which a stale, flattering picture of public opinion would be rewarded rather than corrected.

Two lines of work on factual staleness supply methods we reuse but limit to facts. RealTime QA shows that live search raises weekly-news accuracy from 44.9% closed-book to 66.5%. It also finds that 34% of remaining errors are retrieval failures, and that models fall back on outdated memory instead of abstaining. Mitigating Temporal Misalignment by Discarding Outdated Facts predicts how long a fact stays true and triggers retrieval only when that lifespan has likely elapsed. We adopt its idea of expiry-aware checking for the factual premises inside perspectives. We do not apply it to value stances, which have no truth lifespan.

The debate literature supplies the mechanism and also constrains it. WinnowRAG clusters retrieved documents, gives each cluster to an agent, and uses a critic to eliminate losing clusters. That design is right for single-answer factual questions and wrong for contested ones, because elimination erases legitimate minority views. We keep its cluster-to-agent assignment and replace elimination with labeling. Gradual Vigilance and Interval Communication uses debate for alignment, but only along a single helpful-versus-harmless axis with fixed criteria. A Diagnostic Study of Multi-Agent LLMs for Real-World Debates is a warning we take seriously. It finds that final consensus is a poor indicator of debate quality on subjective topics, and that more diverse agent teams debate more healthily but reach agreement less often. We therefore never ask the agents to agree. We evaluate the debate's output against an external, time-indexed reference set.

Examining Inter-Consistency via Debate adds a second constraint: a stubborn weaker agent can pull a stronger one off a correct answer. This is one reason the challenger agent must cite dated evidence rather than simply assert a view.

Finally, Position: Align AI to Our Aspirations, Not Our Flaws separates legitimate value tradeoffs from a non-negotiable floor that includes accuracy. Our split between premise checking and stance labeling operationalizes only the accuracy part of that floor. We deliberately do not ask the system to decide which moral positions fall below it, since that boundary is itself contested and shifts over time.

Research Program

Benchmark. The first artifact is a backtesting benchmark for normative obsolescence. We take contested questions from OVERTONBENCH and the PRISM alignment data, both used in Benchmarking Overton Pluralism in LLMs. For each question, we build the set of reasonable perspectives at two or more dates, using a dated corpus of opinion writing and news.

An LLM annotator clusters stances by year. We use an automated judge here because that paper reports a rank correlation of 0.88 between its LLM judge and human ratings. Human annotators then validate the clusters and assign each perspective a status at each date. Annotators also mark the factual premises a perspective depends on, and record whether each premise still held at each date.

We keep the questions where the perspective set actually changed, along with an equal number of stable controls. The controls let us detect a system that simply labels everything as waning. Following the human-panel design of the Overton benchmark, which relied on U.S.-representative raters, we limit the first version to U.S. discourse and state that limitation openly.

Evaluation setup. We use open-weight models with documented training cutoffs. Each is queried "as of" a later date, and its retrieval is restricted to corpus documents dated up to that query date. This reproduces the deployment situation, frozen weights plus a live but imperfect search, without leaking future information.

Method. The protocol runs in four steps. First, perspective clusters are formed from both the model's own generations and the retrieved documents, so that views appearing only in the model's frozen memory are included and can be tested. Second, each cluster is assigned to an advocate agent. Third, a challenger agent cross-examines every advocate with two separate demands: show that each factual premise still holds, citing dated sources, and show that the stance is still argued for, citing recent dated sources. Fourth, a judge that sees only the transcript assigns a status label to each perspective and flags expired premises. The judge is not allowed to remove any perspective.

Analyses. We study where staleness originates (the model's memory versus the retrieved documents), how it depends on the gap between training cutoff and query date, and whether users' stated stances shift the status labels.

Baselines

The naive reference point is the closed-book model. It is prompted with the query date and asked to list perspectives with status labels. RealTime QA found that adding the date can hurt closed-book models, so we run this baseline both with and without the date. It shows how much normative staleness the frozen weights carry on their own.

The first strong baseline is a single model with date-restricted retrieval and an explicit labeling prompt. It sees exactly the documents the debate sees and receives the same status-label instructions. It gets the same token budget, spent through multiple self-revision passes, and the same number of prompt-tuning trials on a held-out development split. This is the most direct test of whether the debate structure matters at all.

The second strong baseline, and the strongest competitor on coverage, is pooling. We collect perspectives from several models and many persona-conditioned samples, add retrieval, and have an aggregator deduplicate and label the result. On the static benchmark, Benchmarking Overton Pluralism in LLMs reports that pooling the best responses of eight frontier models reaches an OVERTONSCORE of 0.69, against 0.35–0.41 for single models. We therefore expect pooling to match or beat the debate on raw coverage. The comparison we care about is status accuracy and premise flagging at equal coverage. Pooling is run with token usage matched to the debate.

The third baseline is a WinnowRAG-style protocol following WinnowRAG, with the same clusters and the same critic model, but eliminating losing clusters as the original does. No reported number was found on Lacuna for WinnowRAG on any normative or pluralism benchmark. This baseline measures how much legitimate minority perspective is lost when elimination is applied to contested questions.

For the factual-premise sub-task, we compare against expiry-based adaptive retrieval from Mitigating Temporal Misalignment by Discarding Outdated Facts. That paper reports an exact match of 20.5 on SituatedQA while retrieving for only 45% of queries. We apply its approach to premise sentences pulled from each perspective.

Two ablations isolate the key ingredient. The first merges the challenger's two demands into a single "is this view still valid?" question, which tests whether separating facts from stances matters. The second drops the requirement that the challenger cite dated evidence, which tests whether grounding in dated sources is what protects against the distraction effect.

The debate counts as beating the baselines only if three conditions hold at matched compute and matched coverage. It must improve status-label accuracy on questions where opinion shifted. It must keep false-waning labels on stable control questions no higher than the best baseline's rate. And it must delete or omit fewer legitimate perspectives than the WinnowRAG-style protocol.

Why not just…?

Why not just retrieve recent documents and prepend today's date? This is the first strong baseline and is run exactly that way. The evidence against relying on it comes from RealTime QA: when search fails, models revert to outdated memory rather than abstaining, and search fails often enough to cause a third of errors. Our concern is that a stale view surviving in memory will be presented as current whenever no retrieved document contradicts it. The memory-versus-retrieval breakdown in our analysis tests this directly. If this baseline ties the debate on status accuracy, we report that the debate is unnecessary.

Why not just pool many models and personas? Pooling is the real threat to our coverage numbers, and we concede coverage to it in advance. Pooling widens the set of perspectives but has no step that tests whether a perspective is still current or whether its premises still hold. A pooled set drawn from models with similar cutoffs could share the same stale views. The decisive comparison is whether pooling plus an aggregator labels status as well as cross-examination does when both see the same evidence. If it does, the cheaper method wins.

Why not just let a single model play several personas, as in SAGE or JANUS? Simulated populations of the kind proposed in For Questions of Ought, AI Could Use Some SAGE Advice sample and summarize perspectives but never challenge them. That work also reports that personas built from demographic labels alone collapse into stereotypes. A persona has no reason to admit that its own premise has expired. The merged-challenge ablation and the single-model baseline together measure how much the adversarial check contributes beyond persona diversity.

Decisive Checks

The main experiment plots status-label accuracy against the gap between training cutoff and query date, separately for questions where opinion shifted and for stable controls. If debate helps, its advantage should grow with that gap, because longer gaps leave more stale views in model memory for the challenger to catch.

The second experiment tests sycophancy. Each query is repeated with a user who states an outdated stance and with a user who states a current stance. We measure how much the status labels move between the two conditions. The prediction is that the debate's labels move less than the single model's, because the debate agents and judge never see the user's stance. The single-model baseline is run both with and without the stance hidden, so that simple stance-hiding is credited where it is responsible.

The third check counts silently dropped perspectives across all systems. This tests the core claim that labeling a view as stale protects minority perspectives that elimination would erase.

Every system is scored with the time-indexed version of OVERTONSCORE, status accuracy, false-waning rate, and expired-premise precision and recall. A human audit covers a stratified sample to confirm that the automated judge remains reliable on dated perspective sets.