Long answers from language models tend to start out accurate and then go wrong. The opening sentences state easy facts. Later sentences drift into specific claims where errors pile up, and each error then becomes context that the rest of the answer builds on. The best current defense checks every sentence as it is written. It does this with an expensive consistency test: sample alternative versions of the sentence and see whether they agree in meaning. When the test fails, it rewrites the sentence. This works, but it pays the full cost of the test at every sentence, even though most early sentences are safe.
Clinical statistics has a close analogue. A doctor tracks a noisy, irregularly measured biomarker and wants to know when a patient is about to have an event. Here the expensive lab test is the consistency check, the cheap bedside vitals are token-level confidence signals, and the event is the first false claim. Joint longitudinal-survival models were built for this setting. They combine two parts: a model of a hidden trajectory seen through noisy, occasional measurements, and a model of the moment-to-moment risk of an event (the "hazard"). We propose to fit such a model to the sentence-by-sentence course of a generation. Its estimated hazard then decides three things: when the expensive check is worth running, when a sentence should be resampled, and when the model should stop talking.
Core Claim
We claim that treating "first unsupported claim" as a time-to-event outcome over sentence index lets a decoder match the factual precision of always-on checking while running far fewer checks. The hazard model is fed by a filtered latent trajectory that combines cheap dense signals with expensive sparse ones. The joint model is not there for elegance. It is there because the measurement problem matches the clinical one. The informative signal (semantic entropy) is costly, so in practice it can only be sampled intermittently. The cheap signal (token entropy) is noisy enough that, used raw, it has been reported to be near-useless for long-form text. Generations also end on their own, and that ending may be correlated with risk. These are the three conditions under which, in clinical data, joint models beat plug-in two-stage pipelines: measurement noise, irregular sampling and informative dropout (Yan et al.; Akhavan Masouleh et al.).
The claim is also falsifiable. If a model has an unlimited checking budget, checking every sentence is already optimal, and our method should reduce to it. All of the proposal's value must therefore show up in the budget-constrained regime and in the calibration of the stopping rule. We evaluate it there.
Where This Sits
The closest work is SHARS (Building Reliable Long-Form Generation via Hallucination Rejection Sampling). It scores each sentence with a semantic-entropy detector called HalluSE, resamples sentences that fail, and abstains after a fixed number of consecutive failures. It reports roughly 20–26% higher factual precision than greedy decoding on FActScore. SHARS supplies our detector and our strongest competitor. We differ from it in three ways: we do not check every sentence, we use how uncertainty changes across sentences rather than a per-sentence threshold, and we replace the failure-count stopping rule with a survival-probability rule.
The stopping rule adapts the decision-theoretic abstention region of J-LTM (Scalable Joint Models for Reliable Uncertainty-Aware Event Prediction). J-LTM withholds an ICU alert when the posterior of the event probability is too uncertain. We reuse that construction on text, which J-LTM never touched.
Two linking ideas come from clinical joint models. The first is linking hazard to the slope of the latent marker as well as its level, from work on kidney disease trajectories (Futoma et al.). The second is linking hazard to each subject's own residual variance, from work on hormone trajectories (Chen et al.). Both found signal that level-only models miss. Transformer joint models such as TransformerLSR, and the source-grounded EHR note pipeline of A Multi-Agent Pipeline for Source-Grounded Synthetic Note Generation from Longitudinal Structured EHR, sit nearby but do not model the generation process itself as a survival process.
Two results constrain the design. RATE-FT (Learning Auxiliary Tasks Improves Reference-Free Hallucination Detection) reports that raw token confidence is close to random for long-form text. That rules out a cheap-signal-only design. FActScore (Min et al.) documents both the decay of precision over the course of an answer and a sharp penalty for rare entities. That tells us the baseline hazard must depend on sentence position and on entity prominence, which play the roles that time and baseline covariates play in a clinical model.
Research Program
The model works at the level of sentences. For response $i$ at sentence $k$, a latent "epistemic state" $z_i(k)$ evolves as a slowly drifting random walk with drift. It is observed in two ways. At every sentence we see noisy cheap measurements $x_{ik}$: mean token entropy and a linear probe on hidden states. At sentences the policy chooses, we see an expensive semantic-entropy measurement $s_{ik}$. Each measurement has its own noise level, so token-level stylistic variation is absorbed as measurement error instead of being read as a change in the state. The event is the first sentence containing an atomic claim that the knowledge source does not support. The hazard is
$$ h_i(k) = \sigma\big(\alpha(k) + \eta^\top c_i + \beta_1 z_i(k) + \beta_2 [z_i(k) - z_i(k-1)] + \beta_3 \log \tau_i^2\big), $$
where the terms are defined as follows. $\sigma$ is the logistic function. $\alpha(k)$ is a spline in sentence position that captures factuality decay. $c_i$ holds prompt-level covariates such as entity frequency bin. $\beta_1$ links hazard to the level of the latent state, $\beta_2$ to its velocity and $\beta_3$ to its volatility. $\tau_i^2$ is the response's own residual variance, estimated as a random effect.
A response that ends without an unsupported claim is censored at its final sentence. We also test whether ending is informative by adding end-of-sequence as a competing risk.
Inference runs a Kalman-style filter with a Laplace or particle approximation for the hazard term. The cost is linear in the number of sentences, which avoids the cubic and quadrature bottlenecks that have limited clinical joint models (Li et al.). Parameters are fit offline, on generations whose atomic facts are labeled by the FActScore pipeline (reported under 2% error against humans) or by SAFE (Long-form factuality in large language models). Retrieval is used only to create training labels. At test time the system is reference-free.
The decoding policy runs at each sentence boundary. It first computes the posterior predictive hazard of the next sentence. It calls the semantic-entropy check only when the expected reduction in decision loss from observing $s_{ik}$ exceeds a per-check cost $\lambda$. Sweeping $\lambda$ traces out a precision-versus-checks curve. The policy resamples the sentence when the hazard's posterior falls confidently above a threshold. It abstains when the predicted probability of surviving the next $m$ sentences without an error drops below a target. Following J-LTM, the policy defers rather than acts when the posterior is too wide. We use an open-weights model so that logits and hidden states are available.
Baselines
The naive reference point is greedy decoding with no checks. For this we have ChatGPT's reported FActScore of 58.3% on the biography task (Min et al.), and we rerun greedy decoding on our own model.
The strongest existing method is SHARS with HalluSE checks at every sentence and its consecutive-failure abstention rule. Its reported result is a roughly 20–26% gain in FActScore precision over greedy decoding (SHARS). We found no reported LongFact F1@K for SHARS on Lacuna, and no reported check count or compute figure. We rerun SHARS on the same base model, the same prompts and the same HalluSE sampling settings. We tune its threshold and its failure count on the same development split with the same number of configurations as our model. Beating SHARS means matching its FActScore precision within one point using at most half of its semantic-entropy calls, while keeping at least as many supported facts per response. The last condition guards against winning by producing trivially short answers, a trade-off SHARS itself flags.
We add budget-matched versions of SHARS that check every $j$-th sentence or a random subset of sentences at the same call count as our policy. This isolates whether the savings come from choosing when to check. For detection quality we compare to RATE-FT, which reports 79.6% balanced accuracy on LongFact against 76.1% for standard fine-tuning and 69.1% for SelfCheckGPT. We score it on the same LongFact sentences we use for next-sentence hazard prediction.
Four ablations target the key ingredients. A level-only model removes the velocity and volatility terms. A no-latent model feeds raw signals straight into the hazard. A cheap-only model never calls HalluSE. A two-stage model computes per-response slopes and variances with ordinary regression and plugs them into a discrete-time logistic hazard.
Why not just…?
Why not just tune SHARS's threshold and failure count? At full budget this may be unbeatable, and we expect our model to roughly tie it there. That tie is the sanity check that the model reduces correctly. Our claim concerns fewer checks and a calibrated stopping rule. The budget-matched SHARS variants settle whether uniform or random sparse checking already captures the savings. If periodic checking matches our curve, the survival model adds nothing to budget allocation, and we would report that.
Why not just train a sequence classifier on lagged cheap and expensive features? This is the most serious threat. Text generations are plentiful, and the clinical argument against two-stage pipelines comes from small cohorts (Chen et al.; J-LTM, 0.84 versus 0.79–0.80 AUC). A recurrent classifier trained on enough data could learn the filtering implicitly. We give it the same features and training labels. We then compare discrimination, survival-curve calibration, and above all performance when expensive measurements are missing at irregular positions. A classifier has no native way to handle that last case, whereas the state-space model marginalizes over the missing values. If the classifier matches on all three, the contribution narrows to the decision rule, and we would say so.
Why not just verify post hoc with retrieval? SAFE costs about $0.19 per response and depends on public search, which fails in private domains (Wei et al.). More importantly, post-hoc checking cannot stop an early error from contaminating later sentences (Holt et al.). We use retrieval only for training labels and for evaluation.
Why not just threshold token entropy? RATE-FT reports that raw token confidence is near random for long-form text. Our cheap-only and no-latent ablations test directly whether filtering recovers useful signal from that noise. The answer tests the "internal state myth" either way.
Proposed Experiments
We generate responses to FActScore biography prompts and LongFact prompts, label them as described, and hold out the rare-entity bins for testing. FactualBio's minor/major severity labels let us check whether the hazard ranks major errors earlier than minor ones.
Three results decide the proposal. The first is the precision-versus-checks curve against full and budget-matched SHARS, with informativeness held fixed. The second is next-sentence hazard discrimination and calibration against RATE-FT and the classifier baseline under randomly masked expensive measurements. The third is whether the fitted $\beta_2$ and $\beta_3$ are nonzero on held-out data and whether removing them hurts early warning, meaning detection one or more sentences before the first error.
A negative result on the first comparison, together with a positive result on the third, would still show that the velocity and volatility of model uncertainty forecast hallucination. That finding would stand on its own.
References
- The reporting quality and methodological quality of dynamic prediction models for cancer prognosis
- A Flexible Joint Longitudinal-Survival Model for Analyzing Longitudinally Sampled Biomarkers
- Building Reliable Long-Form Generation via Hallucination Rejection Sampling
- Scalable Joint Models for Reliable Uncertainty-Aware Event Prediction.
- Scalable Joint Modeling of Longitudinal and Point Process Data for Disease Trajectory Prediction and Improving Management of Chronic Kidney Disease
- Using subject-level variability to predict time-varying outcomes: investigating the association between hormone variability and BMD trajectories over the menopausal transition
- TransformerLSR: Attentive joint model of longitudinal data, survival, and recurrent events with concurrent latent structure
- A Multi-Agent Pipeline for Source-Grounded Synthetic Note Generation from Longitudinal Structured EHR
- A Multi-Agent Pipeline for Source-Grounded Synthetic Note Generation from Longitudinal Structured EHR
- Learning Auxiliary Tasks Improves Reference-Free Hallucination Detection in Open-Domain Long-Form Generation
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- A flexible joint model for multiple longitudinal biomarkers and a time‐to‐event outcome: With applications to dynamic prediction using highly correlated biomarkers
- Long-form factuality in large language models
- Data-Driven Discovery of Dynamical Systems in Pharmacology using Large Language Models