When a language model answers from retrieved documents, a common setup called retrieval-augmented generation (RAG), we judge its faithfulness with detectors trained on human labels that mark which parts of the answer are "hallucinated", meaning not supported by the sources. Those labels are not as objective as they look. Some claims are plainly contradicted by a single source sentence. Many others are supported only if you combine several paragraphs, or accept a reasonable inference from the document as a whole. For these, annotators split. A lenient annotator accepts document-level inference, and a strict one demands a sentence that says it outright. Current detectors train on a single merged label and treat this split as noise. That caps their accuracy on exactly the hard cases and hides where the real ambiguity lies.
We propose a detector that keeps the two things apart. For each claim in a generated answer, it measures how strongly the source supports the claim at three levels of the document: single sentence, paragraph, and whole document. Separately, it learns each annotator's personal threshold, that is, how high up this ladder they are willing to go before calling a claim supported. The content side yields a rater-independent judgment of how grounded a claim is and at what level. The rater side explains and predicts disagreement. Put in two sentences: disagreement about hallucination is largely disagreement about which level of evidence counts. If we model the level of evidence explicitly and model each rater's tolerance for it, we get a cleaner detector and a principled prediction of where humans will disagree.
Core Claim
The label "hallucinated" mixes a property of the content (the lowest level of the source at which support exists) with a property of the rater (the highest level of support they accept). We claim three things. First, a detector that factors labels this way fits held-out annotator judgments better than one trained on majority or averaged labels. Second, its rater-independent score ranks claims better against consensus labels. Third, its predicted disagreement correlates with actual annotator disagreement, which neither flat nor annotator-embedding models can explain mechanistically. If the hierarchy does no work, meaning sentence-level support alone explains disagreement just as well, the central hypothesis is false. We treat that as the main risk.
What Prior Work Establishes
Two benchmarks motivate the proposal by exposing the gap. FaithBench collects 750 hard summaries labeled by 11 expert annotators in a four-way scheme. Automated detectors there reach only 50–60% balanced accuracy, the best being RAGAS with GPT-4o at 62.31%. Agreement on the "Questionable" class is low, and chain-of-thought judges do worse than zero-shot ones. The authors identify subjectivity as the obstacle but propose no model of it, and their contexts are short and unstructured. RAGTruth annotates 17,790 RAG responses at the span level. It finds that unsupported additions are far more common than outright contradictions, and that hallucinations cluster at the end of responses (the "tail effect"). Both findings fit our reading: additions and late-response synthesis are exactly the claims whose support, if any, sits at the paragraph or document level. RAGTruth itself treats its span labels as absolute.
The method's rater side comes from rating-bias work. MBNet splits speech-quality scores into an objective mean predicted by one network and a judge-specific offset predicted by another network fed the judge's ID. This raised system-level rank correlation from 0.922 to 0.949 on VCC 2018, and removing the bias branch dropped it to 0.934. We borrow the decomposition but not the free-form offset. Our rater parameter is a threshold over evidence levels, which can be interpreted and estimated with few labels. Are We Modeling the Task or the Annotator? constrains the design. It shows that models can identify annotators from text (F1 up to 0.91) and collapse by up to 23 points on unseen-annotator splits. This makes clear that any rater-aware model must be evaluated on annotators it never saw, and must not let annotator identity become a shortcut.
The content side builds on hierarchical document encoding. GraphMPA recursively summarizes and clusters document chunks into a multi-layer graph, and reports that two layers are the best trade-off. It treats preference labels as monolithic, and it uses hierarchy to improve answering rather than to measure the level at which a claim is supported. We reuse its two-layer construction as one way to build the paragraph and document levels. Cross-level encoders such as Tan et al.'s document-context feedback for translation show that higher-level context can reinterpret local text. That is the intuition behind measuring support above the sentence level.
The strongest detection method, RL4HS, trains a 7B model with group-relative reinforcement learning to reason toward hallucinated spans. It reaches 55.9 span F1 on RAGTruth, ahead of supervised fine-tuning (50.1) and o3 (51.2). It is flat in its treatment of both the evidence and the labels. Our proposal is meant to sit on top of such a detector, not to replace its reasoning.
Research Program
The system has two parts that are trained jointly. The content part splits a response into atomic claims, or uses the annotated spans directly. For each claim $c$ it computes a support profile $s_c = (s_c^{\text{sent}}, s_c^{\text{para}}, s_c^{\text{doc}})$: entailment strengths of the claim given the best single source sentence, the best paragraph window, and a document-level representation built with GraphMPA-style two-layer summaries. These levels are estimated by a fine-tuned Qwen2.5-7B, the same backbone as RL4HS, so that comparisons are clean. We use separate heads per level, and the document head sees only the summary tier, which keeps the levels distinct. The rater part gives annotator $j$ a weighting $a_j$ over the three levels, constrained to sum to one, and a threshold $\theta_j$. The annotator's label is modeled as
$$P(y_{cj} = \text{hallucinated}) = \sigma\big(\theta_j - a_j^\top s_c\big),$$
where $\sigma$ is the logistic function. The term $a_j^\top s_c$ is the support this annotator perceives, and $\theta_j$ is how much support they require. Strict annotators put their weight on the sentence level, and lenient ones spread it toward the document level. Following MBNet's tolerance idea, ordinal labels such as FaithBench's four classes use cumulative thresholds rather than a binary cut.
At test time with an unseen annotator, the rater parameters fall back to a population prior. For deployment we report the population-mean rater together with the spread of predictions across the learned rater distribution, which serves as the predicted-disagreement score. Each annotator has only a handful of parameters. This is deliberate. It addresses the cold-start and long-tail problems that free annotator embeddings suffer from, and it limits the shortcut risk documented by the annotator-bias study. A new rater can be calibrated from roughly twenty labeled claims, and we test this directly.
Data is the practical bottleneck. We use FaithBench's per-annotator labels if they are released at that granularity; otherwise we request them. Because RAGTruth reports a single label per span, we will re-annotate a stratified subset of about 600 RAGTruth responses, covering all three tasks and weighted toward the "Subtle" categories, with six annotators each. We will also record, for every claim an annotator accepts, which part of the source they relied on. That evidence-location field lets us check the support profile against human judgments directly, rather than only through label fit.
Baselines
The naive reference point is a flat detector fine-tuned on majority labels with the same Qwen2.5-7B backbone. It sees the full source as a flat context and has no rater model, and it represents what the field does by default. Above it sit the two published detectors. The first is RAGTruth's fine-tuned Llama-2-13B, which reports 78.7% response-level F1 and 52.7% span-level F1. The second and strongest is RL4HS at 55.9 span F1 on RAGTruth. On FaithBench the strongest reported result is RAGAS with GPT-4o at 62.31% balanced accuracy, and we rerun it as the zero-shot judge baseline. We retrain RL4HS ourselves on the same training split. It is the strongest competitor because its reasoning traces could, in principle, learn to look across paragraphs without any explicit hierarchy.
For fairness, every trained model uses the same backbone, the same training examples (including our new multi-annotator data, which baselines receive as majority or soft labels), and the same hyperparameter search budget measured in GPU-hours. We report means over three seeds. Two further baselines isolate our assumptions. The first is soft-label training, where the detector predicts the fraction of annotators who flag each claim. The second is an MBNet-style model with a free annotator embedding added to the flat detector. Two ablations isolate the key ingredient: removing the hierarchy, so that the rater model sees only sentence-level support, and removing the rater model, so that the hierarchical support profile is fed straight to a majority-label classifier. On the original RAGTruth test set we only need to match RL4HS's 55.9 span F1 using the population-mean rater. The win we claim is on the multi-annotator evaluations: higher held-out-annotator log-likelihood, higher consensus balanced accuracy than 62.31% on FaithBench, and higher correlation between predicted and observed disagreement than both soft-label and embedding baselines. Each comparison is tested by paired bootstrap at the claim level.
Why not just…?
Why not just train on soft labels? Soft labels capture how often annotators disagree on a claim, but they cannot say whether a new annotator will be strict or lenient, and they say nothing about why. This alternative is a real threat on the population-level metric, and it may match us there. The decisive test is held-out annotators: given twenty calibration labels from a new rater, our model updates two or three interpretable parameters, while a soft-label model cannot personalize at all. If soft labels match us on consensus accuracy and on disagreement correlation, the practical case narrows to rater calibration and interpretation, and we will report that plainly.
Why not just add an annotator ID embedding, as MBNet does? MBNet's bias branch works on VCC 2018, where 267 judges each rated many clips. FaithBench has 11 annotators, and our re-annotation will have six. The annotator-bias study shows that free annotator features become shortcuts and fail on unseen raters, with drops of up to 23 points. Our threshold-over-levels model makes a falsifiable structural claim that a free embedding does not. The unseen-annotator split and the small-calibration curve compare the two directly. If the free embedding wins with equal calibration data, our structural assumption is wrong.
Why not just use a stronger flat detector such as RL4HS with the full context? A reasoning detector might implicitly learn to combine evidence across paragraphs. The ablation that feeds RL4HS-style sentence-only support into the rater model tests this, and so does the comparison between our support profile and the evidence locations annotators recorded. If flat RL4HS predicts disagreement as well as the hierarchical profile does, the hierarchy is unnecessary and the contribution reduces to the rater model.
Decisive Checks
Three results would settle the proposal. First, on held-out annotators in FaithBench and our RAGTruth subset, the full model must beat soft-label, annotator-embedding, and no-hierarchy variants on per-annotator log-likelihood. Second, claims that annotators disagree on should be concentrated where sentence-level support is low but document-level support is high. The no-hierarchy ablation should lose most of its ability to predict disagreement, and the support level each annotator relied on should agree with the level our profile assigns. Third, as a secondary analysis on RAGTruth, claims late in a response should more often have support only at the document level. If this holds, it gives a mechanism for the tail effect: late sentences tend to synthesize. If the first two checks fail, we conclude that hallucination disagreement is not organized by evidence level. That negative result would still tell benchmark builders that rater modeling, not document structure, is where the remaining signal lies.
References
- FaithBench: A Diverse Hallucination Benchmark for Summarization by Modern LLMs
- RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models
- MBNet: MOS Prediction for Synthesized Speech with Mean-Bias Network
- Are We Modeling the Task or the Annotator? An Investigation of Annotator Bias in Natural Language Understanding Datasets
- A Comprehensive Graph Framework for Question Answering with Mode-Seeking Preference Alignment
- Learning to Reason for Hallucination Span Detection