Does Long-Form Uncertainty Survive the Training Cutoff?

When a language model writes a multi-paragraph answer, some sentences are right and some are invented. Methods that flag the invented ones usually sample the model several times and treat claims that keep reappearing as trustworthy. These methods have been tested almost entirely on public, static prompts such as biographies of well-known people. For those prompts, the model may have memorized both the right answer and its own consistent way of stating it. So the reported ability to separate true claims from false ones may partly measure memorization rather than any real sense of what the model does not know. This proposal tests that possibility directly. We build a long-form evaluation from technical-domain topics whose reference documents are dated before or after each model's training cutoff. We then compare how well uncertainty scores rank claim correctness on each side of that line.

A second problem makes the comparison unfair if ignored. On new or specialized topics, models drift between synonyms for the same concept. One remote-sensing corpus contained seven Estonian renderings of "reflectance." Consistency-based methods read that drift as disagreement. After the cutoff, terminology is less settled, so uncertainty would look higher for reasons unrelated to facts. We therefore map each extracted claim's technical terms onto a domain termbase before measuring agreement. We treat this mapping as a control for the cross-cutoff comparison, not as a separate contribution.

Core Claim

Sampling-based long-form uncertainty methods are claimed to rank claim correctness well. We hypothesize that this is inflated on pre-cutoff, likely contaminated topics. Our prediction is that after the cutoff, the area under the ROC curve (AUROC: the probability that a random false claim gets a higher uncertainty score than a random true one) drops more for consistency-based scores than for token-probability scores. The reason is that models hallucinate new material consistently across samples. We also predict that part of any apparent post-cutoff rise in raw uncertainty comes from lexical sprawl, and that it disappears once terms are canonicalized. If both predictions hold, the field's long-form uncertainty numbers need temporal qualification. A further consequence concerns the "convincingly wrong" behavior of reasoning models, which is currently known only from synthetic tasks. Our design can show whether that behavior persists on fresh real-world prose, which would make it intrinsic, or shrinks there, which would make it partly a contamination artifact.

Relation to Prior Work

Several papers supply the uncertainty methods we audit.

  • Graph-based Uncertainty Metrics for Long-form Language Model Generations links sampled responses to the atomic claims they entail. It shows that self-consistency equals degree centrality in this graph, and that closeness centrality works better.
  • GAUSS aligns graphs of facts and their dependencies across samples using a Fused Gromov-Wasserstein distance, which scores how costly it is to match both the facts and their relations.
  • Beyond Semantic Entropy replaces hard clustering of whole responses with soft pairwise similarity (SNNE, and WSNNE, which also weights by token probability). The motivation is that cluster counts explode with length.
  • LUQ scores sentence-level consistency without access to model internals.

All of these evaluate on static benchmarks and compare claims with generic embeddings or entailment models. We neither change their scoring rules nor propose a new one. We change the conditions under which they are evaluated.

Two papers motivate the gap. Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth (SALT) avoids contamination by generating synthetic tasks such as matrix math and DNA translation. Its findings are that log-probability confidence beats verbalized confidence, that ranking collapses at fine granularity, and that reasoning models are confidently wrong. It explicitly leaves natural-language prose untested. FActScore provides the claim decomposition and retrieval-backed verification we reuse. It also documents a steep accuracy penalty for rare entities, which we must control for because post-cutoff topics are rarer by construction.

Temporal benchmarks supply the isolation method. LiveCodeBench shows a "contamination cliff" in pass rates at model cutoffs. LiveBench refreshes objective questions monthly but excludes long-form tasks. CausalProbe-2024 shows that models near ceiling on static causal benchmarks fall sharply on post-cutoff news. TripleFact adds entity substitution to separate entity familiarity from reasoning. None of these measures uncertainty quality.

One constraint shapes our sourcing. FreshBench found that arXiv text became unusually predictable after March 2024 because of LLM-written prose. We therefore avoid relying on recent preprints as references.

The terminology method comes from the Estonian-English remote sensing termbase work. It extracts terms by combining grammatical patterns with a distinctiveness statistic against a reference corpus, and it records variants and replacements of terms over time. The NFPA 1 translation study and TICO-19 show that technical meaning is fragile under rewording and translation. TICO-19 also illustrates the reverse risk: public termbases are themselves likely to be in training data.

Research Program

Benchmark. We build a dated long-form benchmark in two technical domains that have usable controlled vocabularies: remote sensing and public-health/clinical guidance. Each prompt asks for an expository account of a concept, instrument, method, or regulatory provision. Each topic carries a reference document with a verifiable first-publication date, taken from sources such as standards revisions, agency guidance, and dated encyclopedia article creation.

Matching. For every model, topics are binned by their date relative to the model's stated cutoff. Post-cutoff topics are matched to pre-cutoff topics on three properties:

  • reference length,
  • domain subfield,
  • a popularity proxy (count of independent web mentions at collection time).

Without this matching, a post-cutoff drop would simply reproduce FActScore's rarity effect.

Entity-swapped control. A third arm takes pre-cutoff topics and substitutes fictional entity names in the style of TripleFact, keeping the technical content fixed. This separates two possible causes of failure: unfamiliar names versus unfamiliar knowledge.

Claim pipeline. Generations are split into atomic claims with the FActScore procedure. Each claim is verified against its dated reference by retrieval plus an evaluator model. Because the verifier always sees the reference text, the "judging gap" LiveBench describes (models judging problems they cannot solve) is reduced but not eliminated. We therefore have domain experts label a stratified subset in each time bin to estimate verifier error. SALT shows that label noise of this kind can distort AUROC and calibration, so we report verifier error per bin and check that it does not itself shift across the cutoff.

Canonicalization. Before any consistency score compares claims across samples, technical spans are mapped to termbase entries:

  • Existing controlled vocabularies are used where they exist.
  • Elsewhere, termbases are induced with the unithood/termhood extraction method and then expert-verified.
  • To avoid leakage, induction for post-cutoff bins uses only post-cutoff documents.

Spans that match no termbase entry are kept unmapped and flagged. This step is the only thing that differs between the "raw" and "canonical" versions of each uncertainty method.

Baselines

The naive reference point is verbalized confidence: the model is asked to rate its confidence in each sentence. SALT reports that information-based scores consistently beat it on long-form targets. Including it shows whether that ordering holds on fresh prose.

The second baseline is self-consistency, meaning the fraction of samples that support a claim (degree centrality). It is the simplest sampling method, and every graph method is a refinement of it.

The strongest consistency-based competitor is closeness centrality from Graph-based Uncertainty Metrics. It reports a 6.8% average relative AUPRC gain over baselines for identifying false claims on FActScore-style biographies. We also run GAUSS, for which no reported result was found on Lacuna (the file lists its benchmarks but no scores). We run WSNNE, for which Beyond Semantic Entropy reports an average QA AUROC of 0.83 against 0.79 for semantic entropy on Llama-3.1-8B. On the white-box side, we run SALT's Logprobs Sum, the best AUROC ranker in that study. No reported number for it on natural prose was found on Lacuna.

All methods get the same conditions:

  • the same 20 samples per prompt (GAUSS's convergence analysis uses N=20),
  • the same claim extractor,
  • the same entailment model,
  • the same verifier labels,
  • hyperparameters (such as GAUSS's structure–semantics weight) tuned once on a held-out pre-cutoff split and frozen.

The key ablation runs every consistency method with and without termbase canonicalization, holding everything else fixed. The proposal "wins" if either of two results appears. The first is a statistically reliable cutoff-by-method interaction in AUROC, with consistency methods losing more after the cutoff than log-probability methods, beyond the matched-popularity and entity-swap controls. The second is that canonicalization recovers a measurable share of the post-cutoff AUROC loss for consistency methods while leaving pre-cutoff AUROC roughly unchanged.

Why not just…?

Why not just run FActScore-style evaluation on post-cutoff biographies and skip terminology? This is the most serious threat. It already covers the temporal half of the question cheaply. If our canonicalization ablation shows no interaction with time bin, we will report that and drop the terminology component. The dated benchmark and the uncertainty audit stand on their own. We keep the component only because biographies use little specialized vocabulary. Lexical sprawl is a technical-domain phenomenon, as the seven-variant "reflectance" case shows, so a biography-only study cannot detect it.

Why not rely on a strong generic entailment model, which should already treat synonyms as equivalent? It may, for English terms that are common on the web. For newly coined or replaced terms, which are exactly the post-cutoff case, the entailment model has the same knowledge gap as the generator. We test this directly: on the expert-labeled subset, we measure how often the raw entailment model splits a canonical concept, broken down by time bin. If that error rate is flat across time, canonicalization adds nothing and we say so.

Why not label contamination per item with a membership-inference detector such as Min-K%++ or PAC instead of using dates? Item-level detectors are complementary, and we will report Min-K% scores per item as a covariate. They cannot be the primary split, for three reasons. The survey of contamination detection methods notes that paraphrase defeats such detectors. On The Fragility of Benchmark Contamination Detection in Reasoning Models reports that they become fragile on reasoning-tuned models, which are central to our question. And PAC itself notes that APIs increasingly restrict log-probability access. Dates give a split that does not depend on model internals.

Why not just have a strong model judge each long answer? LiveBench shows judges fail on problems they cannot solve, and SALT shows that judge noise corrupts calibration metrics. We use an evaluator model only for checking claims against a supplied reference. We audit it with experts in every time bin.

Proposed Experiments and Decisive Checks

We evaluate open-weight models with known cutoffs, plus reasoning-tuned variants, across monthly bins spanning roughly a year on each side of each cutoff. Reported metrics are claim-level AUROC, AUPRC, and expected calibration error for each method, together with factual precision and sample-agreement rates for each bin.

The first decisive check looks for an "uncertainty cliff." It asks whether consistency-method AUROC drops at the cutoff after conditioning on popularity, while the entity-swapped pre-cutoff arm does not drop. That pattern would locate the effect in knowledge rather than names.

The second check asks whether high-agreement false claims, meaning consistent hallucinations, become more frequent after the cutoff. This is the proposed mechanism behind the first effect.

The third check compares reasoning and non-reasoning variants of the same model family. If SALT's confident-wrong gap is intrinsic, it should appear in both bins. If contamination drives it, it should be wider before the cutoff.

The fourth check is the canonicalization ablation, read as an interaction with time bin. If none of the first three effects appear after controls, the conclusion is still useful: existing long-form uncertainty numbers transfer to fresh technical prose. Each benchmark refresh would then serve as a recurring audit rather than as the foundation of a claimed failure.