Hallucination at the Edge of a Burst

Language models most often invent facts about real papers when the paper sits on a research topic that was just starting to surge in attention at the moment the model's training data was frozen. At that point the model has seen the topic's name and a few early papers, but not the later literature that settled what those papers actually showed. This note proposes to test that claim directly. We would locate each topic's attention surges, or "bursts", in the citation record with a standard burst detector. We would then ask whether a topic's burst state at the model's training cutoff predicts citation hallucination. The key requirement is that it predict hallucination beyond the two obvious explanations, which are publication year and how often the paper had been cited.

If the claim holds, it gives a cheap, black-box risk signal. The signal needs only bibliographic metadata, and it tells a verifier where to spend its effort. It also gives a principled way to build the hard negative examples that hallucination detectors need. The negatives are "temporal counterfactual" claims: statements of a finding written the way the field described it before the burst consolidated. The work treats burst detection as the core measurement. Temporal drift in word meaning supplies the mechanism and the negative-example generator. Hallucination is the outcome we measure.

Core Claim

Burst detection models a paper or keyword as a stream of events with a hidden low-rate or high-rate state. A burst is declared only when the elevated rate lasts long enough to pay a penalty for switching states. For each topic and each model with a known training cutoff year $c$, we assign one of four burst phases. The phases are "no burst by $c$", "burst onset within two years before $c$ and still ongoing at $c$", "burst finished before $c$", and "burst begins after $c$".

The central prediction is that citation hallucination on real pre-cutoff papers peaks in the onset phase. By hallucination we mean two things: stating findings the paper does not contain, or substituting a different paper for the correct one. We further predict that this onset effect survives controlling for publication year and for citations accumulated by $c$.

The test is a logistic regression of a per-item hallucination label $h$ on burst phase $b$, log citations at cutoff $n_c$, years before cutoff $\Delta y$, and a word-drift score $d$:

$$\operatorname{logit} P(h=1) = \beta_b + \gamma \log(1+n_c) + \delta\, \Delta y + \eta\, d,$$

Here $\beta_b$ is a per-phase intercept, and $d$ is the change in the topic keyword's embedding across the years around $c$. The claim is falsified if the onset-phase coefficient is not significantly above the "burst finished" phase once the other terms are in the model.

Why this is not already done

Several papers come close, and each plays a different role here.

Three papers motivate the question. Large Language Models Reflect Human Citation Patterns with a Heightened Citation Bias masks citations in 166 recent cs.LG papers and asks models to fill them in. It finds that suggested references skew toward "superstar" papers, with a median of 1,326 more citations than the true references. It measures only cumulative counts, never the timing or acceleration of attention. Sibyl partitions astrophysics literature by year and documents two failure modes: real papers cited for fabricated findings, and modern framing leaking into pre-cutoff hypotheses. It uses time only as a fixed train/validation split. Harnessing Temporal Databases for Systematic Evaluation of Factual Time-Sensitive Question-Answering in LLMs shows a 21.7% average drop when models must also get the timeframe right, but it works on structured facts rather than scientific text.

One paper supplies structure but not the method. SurveyG layers papers into foundation, development and frontier tiers using a heuristic trending score and a publication-year percentile. It does not use a state-change burst model, and it never asks where hallucination concentrates.

Two papers supply methods we reuse. DRIFT aligns word embeddings across years (Temporal Word Embeddings with Compass) on cs.CL abstracts from 1994 to 2021, which gives us $d$ and the neighbor terms used to build counterfactuals. FACTUM scores internal pathway signals at the step where a citation is generated. It reports that stronger feed-forward "parametric force" accompanies correct citations, on a corpus treated as static.

That last finding is the one this proposal most directly challenges. For onset-phase topics, what the model has memorized is precisely the unsettled early picture of the field. We therefore predict the sign of parametric force reverses in that stratum.

Two papers constrain the design. (Im)possibility of Automated Hallucination Detection in Large Language Models proves that detection needs explicit negative examples, so any trained detector we build must be fed deliberate falsehoods. SciFact-Open shows that verification degrades on realistic corpora, from 72.5 to 52.4 F1 for MultiVerS, with 44% of evidence pairs mismatched in specificity. Our labels therefore cannot rest on automatic entailment alone.

Research Program

The first component is a burst-annotated hallucination probe set. We start from the cs.CL corpus used by DRIFT and extend it with citation records from Semantic Scholar, the same source Algaba et al. used for verification. We run the burst detector on two streams: keyword streams from titles and abstracts, and incoming citation streams per paper.

For each sampled real paper we pose two tasks. The first is an Algaba-style masked citation task, where the model must name the cited work from its citing context. The second is a findings task, where the model is asked "what did paper X report about Y?" The labels are whether the returned reference exists, whether it is the correct paper, and whether the stated finding is supported by the abstract. Support is judged by a model verifier, with a human-annotated stratified subset of about 1,500 items used to estimate verifier error in each burst phase. Sampling is stratified so each burst phase has matched distributions of publication year and citations-at-cutoff. This matching is what lets the regression separate burst timing from simple exposure.

The models are Llama-3.1-8B-Instruct, which gives white-box access, and GPT-4o, accessed through its API. Using each model's documented cutoff shifts which topics fall in the onset phase, so the same paper serves as its own control across models.

The second component is the mechanistic check. On Llama-3.1-8B we compute FACTUM's parametric force and pathway alignment scores at the step where the citation is generated, and fit their relation to correctness separately within each burst phase.

The third component uses the effect. We build temporal counterfactual negatives: for a paper in an onset-phase topic, a model rewrites its true finding using the pre-burst sense of the topic term. The pre-burst sense is identified from the term's embedding neighbors in the years before the burst. These negatives are added to CiteCheck-style training of a parameter-efficient-tuned Llama-3.1-8B faithfulness detector. Separately, the burst phase is used as a prior that allocates a fixed budget of expensive verification calls across a model's output.

Baselines

The naive reference point is to assume hallucination is uniform. That means a constant-rate predictor for the diagnostic study, and random allocation of the verification budget for the downstream use. Anything we claim must beat this by a clear margin.

The strongest competitor for the diagnostic claim is exposure. We fit a regression containing only years-before-cutoff and log citations-at-cutoff, on the same items and labels. This baseline encodes the simple story that models hallucinate about papers they rarely saw, which the Algaba et al. popularity finding makes very plausible. The burst terms must significantly improve held-out log-likelihood and area under the ROC curve (AUC) over this model. The onset coefficient must also be positive and stable across both models. There is no reported result on Lacuna for predicting citation hallucination from bibliometric features, so this baseline is established here.

For the task itself, the reference numbers are Algaba et al.'s existence rates for GPT-4: 64.4% with plain prompting and 86% with iterative prompting. We reproduce their prompting on our items so that our existence rates are comparable before we stratify them.

For detection, the strongest black-box method is a CiteCheck-style Llama-3.1-8B tuned detector (CiteCheck), which reports 90.6% test accuracy. Zero-shot GPT-4o is included as a weaker reference at 71.6% accuracy on negatives. The strongest white-box method is FACTUM, which reports a peak AUC of 0.737 on TREC NeuCLIR 2024.

We retrain the CiteCheck-style detector on the same base model with the same number of training examples and the same tuning budget. The only difference is that a fixed fraction of its standard negatives is replaced by temporal counterfactuals. Beating it means higher accuracy on the onset-phase test stratum without losing accuracy on the rest. FACTUM scores are computed on our items without retuning. Beating FACTUM means our phase-conditioned scoring raises AUC within the onset stratum.

Two ablations isolate the key ingredient. The first replaces burst phase with SurveyG's publication-year percentile tier, which tests whether a formal burst model matters at all. The second replaces temporal counterfactuals with equally many generic perturbations, such as entity swaps, which tests whether the temporal content of the negatives is what helps.

Why not just…?

Why not just filter or flag by publication year relative to the cutoff? Year captures only whether the model could have seen the paper, not what state the field was in when it did. ChronoQA (A Question Answering Dataset for Temporal-Sensitive Retrieval-Augmented Generation) finds that implicit, event-relative time cues cause 69% of failures, which a date window cannot express. Still, year is a real threat, and it is built into the exposure baseline. If the onset coefficient vanishes once year is included, the claim is dead.

Why not just use citation count at cutoff as the risk signal? This is the most dangerous alternative. Rarely cited papers are plausibly the ones models misremember, and the substitution effect Algaba et al. document looks like a popularity effect. The matched-exposure sampling and the regression test exactly this. The distinguishing prediction is that among papers with equal citations-at-cutoff, onset-phase topics are worse than topics whose burst had already finished. A popularity-only account predicts no difference between them. If we find none, we will report that burst timing adds nothing beyond exposure.

Why not just ask a strong model to verify every citation? Zero-shot GPT-4o reaches only 71.6% on unsupported citations in CiteCheck. SciFact-Open shows verifiers disagree heavily, with Jaccard similarity below 0.2 between models. Exhaustive checking with retrieval also costs money and time. Our budgeted-verification experiment asks a narrower question: at a fixed number of verifier calls, does allocating them by burst phase catch more fabricated findings than random or citation-count allocation? If verification is cheap enough that exhaustive checking wins, the burst prior is useful only for diagnosis, and we will say so.

Why not just use word-drift scores alone? Drift and bursts are correlated, since new senses often arrive with surges in attention. The drift term $d$ sits in the same regression so the two can be separated. Drift is also slower and costlier to compute than burst detection on citation counts. If drift alone explains the effect, the practical recommendation changes to using drift, and the counterfactual generator still stands.

Decisive Checks

Three results decide the proposal.

First, the onset-phase coefficient must be positive and significant after controlling for exposure, in both Llama-3.1-8B and GPT-4o. It must also shift with each model's cutoff: the same paper should be riskier for the model whose cutoff falls in that topic's onset years. This cross-model shift is the strongest evidence against confounding by topic difficulty.

Second, FACTUM's parametric force should correlate positively with correct citations outside the onset stratum and weakly or negatively inside it. That would show "memory corroborates retrieved context" holds only for settled topics.

Third, temporal counterfactual negatives should raise detector accuracy on the onset stratum beyond both the generic-perturbation ablation and the standard CiteCheck training mix. The other strata serve as a check that general accuracy is not traded away.

A failure of the first check ends the work as a clean negative result on burst timing. A failure of the second or third, with the first intact, still leaves a useful black-box risk map for scientific assistants.