When a language model writes a biography and states that its subject "graduated from Harvard in 1985," it may sound fully confident and still be wrong. Today's entity-level hallucination detectors flag a name, date, or place when the model was uncertain while producing it. They therefore miss exactly these confident errors, and the authors of the leading method say so themselves. This proposal tests a simple diagnosis of many confident errors. The model is not recalling a fact about this person. It is filling a sentence slot with whatever is typical for that kind of person. The check is a counterfactual: we replace the subject with several made-up people of the same type, keep only the clause that holds the entity, and ask whether the model would have written the same entity anyway. If it would, the entity is subject-blind. The model's confidence then comes from the sentence frame, not from knowledge of the subject, and the entity should be treated as suspect.
Two of the other directions play supporting roles. Punctuation does real work in the method. Commas and periods mark off the clause around each entity, and that clause becomes the context for the counterfactual. The same structure also lets us test whether a known fact can be explained by slot filling: entity hallucination rates climb from 9% at the start of a sentence to 36% at the end. Hyperparameter optimization constrains the evaluation. Every detector, ours and the baselines, gets the same tuning budget, so any gain cannot be blamed on the baselines' thresholds being set by hand.
Core Claim
Our central claim is that a large share of high-confidence entity hallucinations are subject-blind. The model assigns them nearly the same probability when the subject is swapped for a random surrogate of the same type, in the same punctuation-bounded clause. Let $e$ be a generated entity span and $c$ the clause containing it, bounded by the nearest enclosing commas, semicolons, or sentence-ending marks. Let $x$ be the true subject and $\tilde{x}_1,\dots,\tilde{x}_K$ be $K$ surrogate subjects of the same type and similar popularity. We define the subject-dependence score as
$$ s(e) = \log p(e \mid x, c) \;-\; \frac{1}{K}\sum_{k=1}^{K} \log p(e \mid \tilde{x}_k, c), $$
where $p(e \mid x, c)$ is the model's probability of the entity's tokens, given a minimal frame naming the subject followed by the clause prefix up to $e$. A low $s(e)$ means the subject did not matter, so the entity was produced from the frame alone. This should hold even when $\log p(e \mid x, c)$ is high and entropy is low, which is the regime where current detectors fail.
The claim makes two predictions that existing work does not. First, on hallucinated entities the model produced with low entropy, $s(e)$ should separate true from false entities well above chance, while entropy cannot by construction. Second, subject-blindness should be more common in clause positions that punctuation marks as slot-like: trailing prepositional phrases, comma-delimited lists, and non-restrictive appositives. That excess should account for much of the end-of-sentence trap after controlling for entity type and popularity.
Where This Sits Relative to Prior Work
The proposal grows out of three findings. HalluEntity annotates 18,785 entities in GPT-3.5 biographies and finds that 15% are hallucinated. Proper nouns and numbers fail at 30.9% and 36.2%, and hallucination rates rise from 9% to 36% across a sentence. That benchmark is diagnostic only and offers no explanation for the positional effect. DRAD uses entity probability and entropy to decide when to retrieve evidence and regenerate. It cuts retrieval calls on 2WikiMultihopQA from 9.61 to 1.40, but its authors state that it cannot see high-confidence hallucinations. FActScore documents that factual precision collapses on rare entities. Its biography task gives us a mitigation benchmark.
The closest use of masking is FActScore's nonparametric-probability check, which masks content inside the evaluator to test whether the evaluator relies on retrieved evidence. We instead mask the subject inside the generator to test whether a generated entity depends on who is being described. The surrogate step borrows its method from mention randomization in the entity-masking literature, for example EMBRE. That work replaces names during training so that relation extractors cannot exploit co-occurrence shortcuts. We apply the same logic at inference time, as a probe rather than a training regularizer.
Three punctuation papers supply the clause segmentation:
- Labeling the Semantic Roles of Commas classifies commas into nine roles at 83.6% accuracy with automatic parses.
- Annotating and Predicting Non-Restrictive Noun Phrase Modifications detects comma-enclosed non-restrictive modifiers at 0.66 F1 with predicted parses.
- Punctuation: Making a Point in Unsupervised Dependency Parsing shows that text between punctuation marks tends to form a unit governed by a single head. This is the "Loose" constraint, and it justifies treating comma-bounded fragments as self-contained contexts.
Two other punctuation papers set constraints on the design.
- Punctuation and Predicates in Language Models finds that punctuation tokens act as information bottlenecks in GPT-2 but not in Gemma or DeepSeek. That information is spread across tokens in the newer models. We therefore do not read hidden states at punctuation positions, which would likely fail on modern models. We use punctuation only to segment the text.
- Punctuation-aware Hybrid Sparse Attention uses punctuation as attention anchors for long-context efficiency. It never measures factuality, and we do not modify attention.
For the evaluation protocol, Quantifying the Hyperparameter Sensitivity of Neural Networks and MolRep both show that uneven tuning reverses model rankings. TRAC supplies the tuning method: Bayesian optimization of a pipeline's thresholds on a held-out split.
Research Program
The work has three parts.
The probe. For each entity span in a generated biography, we parse the sentence and take the smallest punctuation-bounded fragment containing the entity, using comma-role labels to decide where a fragment ends. A comma inside a "City, State" location does not end a fragment; an appositive comma does. We then build a minimal frame: "[Name] is a [occupation]." followed by that fragment up to the entity. Here [Name] is either the true subject or one of $K$ surrogates. We draw surrogates from Wikipedia people with the same occupation, gender, and popularity bin, so that cues such as pronoun agreement and occupation stay constant while identity changes. This avoids the loss of meaning that blank [MASK] tokens cause.
We compute $s(e)$ with the scoring model's log-probabilities summed over the entity's tokens. Because restricting context to the clause is a design choice, we ablate it directly. The variants condition on the full preceding biography, the current sentence, and the clause alone. The full prefix should leak the subject's identity through earlier facts and weaken the contrast. The clause should give the cleanest signal.
The structural analysis. We fit a regression of hallucination labels on entity type, entity popularity, token position, and clause role (main clause, trailing phrase, list item, appositive, non-restrictive modifier). We then ask how much of the end-of-sentence effect remains after adding $s(e)$.
The mitigation. We reuse DRAD's retrieve-and-regenerate step with a different trigger. Retrieval fires when an entity is either uncertain or subject-blind, and we measure FActScore precision against the number of retrieval calls.
Baselines
The naive reference is to flag nothing, which on HalluEntity means accepting a 15% entity hallucination rate. Alongside it we report raw token likelihood and raw entropy, which HalluEntity shows over-flag informative true entities. These set the floor that any structural signal must clear.
The strongest existing detector is Focus, the context-aware uncertainty method evaluated in HalluEntity. It is the best reference-free entity-level method in the file, reporting AUROC 0.78 and AUPRC 0.40 on Llama3-8B. We run it with the same scoring model and the same biographies, split by biography. Unlike in the original paper, its keyword-selection and weighting settings get the same 100-trial Bayesian optimization budget on the development split that our method gets. Beating it means higher AUPRC on the full test set without lower AUROC. More importantly, it means a clear margin on the low-entropy hallucination subset, where we expect Focus to sit near chance. AUPRC is the primary metric because positives are rare.
The strongest mitigation baseline is DRAD, run with the same retriever (BM25 over Wikipedia), the same generator, and its thresholds tuned under the same budget. The file reports no FActScore number for DRAD; the paper reports only that it beats SelfCheckGPT-style baselines on WikiBio GPT-3. As fixed reference points on the biography task we have FActScore's reported 58.3% for ChatGPT and 71.5% for the search-grounded PerplexityAI. Beating DRAD means higher FActScore precision at an equal or lower average number of retrieval calls.
Two ablations isolate the key ingredients. The first replaces true surrogates with blank [PERSON] placeholders, to test whether keeping the type and popularity of the surrogate matters. The second replaces punctuation-bounded clauses with fixed-width token windows of the same average length, to test whether punctuation structure matters or any short context would do.
Why not just…?
Why not just add sentence position and entity type as features to an entropy detector? This is the most serious threat. The end-of-sentence trap and the high error rates on numbers might be fully explained by position and type, with no counterfactual needed. We run exactly this model, Focus plus position, type, and popularity features, under the same tuning budget, and report how much $s(e)$ adds over it. If the added AUPRC on low-entropy hallucinations is negligible, the subject-blindness hypothesis fails. We would then report a cheaper, position-aware detector as the result instead.
Why not just resample the generation and check self-consistency? Resampling-based checks such as the SelfCheckGPT variants DRAD compares against measure whether the model gives the same answer twice. A confidently wrong model gives the same wrong answer every time, so consistency should not catch the failures we target. This is a prediction, not established evidence. We include a resampling baseline with the same sampling budget as our $K$ surrogate forward passes and report it on the low-entropy subset.
Why not just retrieve evidence for every entity? Always retrieving does help: PerplexityAI reaches 71.5% versus ChatGPT's 58.3% in FActScore. But 28.5% of its facts remain unsupported, and indiscriminate multi-round retrieval cost 9.61 calls per question where targeted retrieval needed 1.40 in DRAD. We plot precision against retrieval calls for always-retrieve, DRAD, and our trigger. If always-retrieve dominates at the call budgets practitioners can afford, the detection contribution matters only for budgeted settings, and we will say so.
Why not just tune DRAD's thresholds properly? The tuning literature warns that hand-set thresholds understate methods (Wiemerslage et al.). Tuned entropy thresholds might close the gap. Tuning moves a threshold, though, and cannot rank a confidently wrong entity below a confidently right one when both have low entropy. Our equal-budget protocol tests this directly. We also report how robust each detector is to its settings, using the share of tuning trials within 1% of the best result. A detector that wins only at one fragile setting counts as a weaker result.
Proposed Experiments
The primary experiment uses HalluEntity's entity labels with five-fold cross-validation split by biography. Llama3-8B is the main scoring model, matching the reported Focus number, and we replicate with a second open model to check generality. We report AUROC and AUPRC overall and on the bottom-entropy tercile of hallucinated entities.
The secondary experiment is the structural regression. It tests whether clause role and $s(e)$ jointly account for the rise from 9% to 36% along the sentence, beyond what entity type and popularity explain.
The mitigation experiment generates biographies for the FActScore entity list, split across frequency bins. We score them with FActScore and count retrieval calls. Separate reporting on rare entities matters most, because popularity-matched surrogates should be hardest to distinguish from true subjects exactly when the model knows least.
Two limits are known in advance. The probe needs token log-probabilities. For black-box models we test an approximation that counts how often surrogate-conditioned samples reproduce the entity, and we report how much accuracy it loses. The comma parsers lose recall where punctuation is missing (Stanovsky & Dagan report 0.19 recall for commas alone). LLM output is usually punctuated, but we record fragment-detection failures and score affected entities with the sentence-level fallback.
The decisive checks are three.
- Does $s(e)$ beat tuned Focus plus position and type features on low-entropy hallucinations?
- Does the clause-bounded context beat both the full prefix and fixed windows?
- Does using $s(e)$ as a retrieval trigger raise FActScore precision over tuned DRAD at matched retrieval cost?
If the first check fails, the idea is wrong. If the second fails, punctuation has no role and the method reduces to a plain counterfactual probe.
References
- HalluEntity: Benchmarking and Understanding Entity-Level Hallucination Detection
- Mitigating Entity-Level Hallucination in Large Language Models
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- EMBRE: Entity-aware Masking for Biomedical Relation Extraction
- Labeling the Semantic Roles of Commas
- Annotating and Predicting Non-Restrictive Noun Phrase Modifications
- Punctuation: Making a Point in Unsupervised Dependency Parsing
- Punctuation and Predicates in Language Models
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models
- Quantifying the Hyperparameter Sensitivity of Neural Networks for Character-level Sequence-to-Sequence Tasks
- MolRep: A Deep Representation Learning Library for Molecular Property Prediction
- TRAC: Trustworthy Retrieval Augmented Chatbot