Does Prompt Variation Track Where Clinicians Disagree?

When a language-model assistant reads a clinical signal and changes its answer after a small rewording of the question, two very different things could be going on. The model may be brittle: a stray space or a reordered option pushes it across a decision line. Or the patient's signal may be genuinely ambiguous, so that trained experts would also split between readings. Today these two cases look the same from outside. Both show up as "the model gave different answers." A clinician looking at a copilot's explanation has no way to tell whether the variation is noise to distrust or disagreement worth hearing.

We propose to separate the two, using data where the truth about human disagreement is known. The testbed is sleep staging, where each 30-second window of an overnight recording has been scored independently by four to seven experts. We perturb the model's prompt in two families. One family is nuisance edits that should change nothing. The other is clinically legitimate reframings that a real scorer might adopt. We then ask three questions. Does the model stay put under nuisance edits? Does it move when the evidence moves? And when it moves under legitimate reframings, does it land on the stages the experts actually chose? We route the model's answer through a guideline-defined intermediate layer of clinical feature judgments, which lets us see which judgment moved whenever the answer moves.

Core Claim

A trustworthy clinical explanation should have calibrated variability, which has three parts. It is invariant to phrasing that carries no clinical content. It responds to changed clinical evidence. And where the underlying signal is perceptually ambiguous, its spread under legitimate reinterpretation matches the spread of expert opinion. Prior work measures at most one of these at a time.

We claim three things. First, current models fail the invariance and responsiveness tests in exactly the way the text-only literature predicts. Second, raw prompt sensitivity, the uncertainty signal used by existing methods, mixes brittleness with genuine ambiguity. Third, forcing the answer through explicit guideline-level feature judgments, with the final stage computed by a fixed rule from those judgments, both reduces nuisance flips and makes the remaining variation line up with expert disagreement. If the third claim fails, that result is itself informative: it means prompt variation is not a usable proxy for clinical ambiguity. Such a proxy is implicitly assumed whenever sensitivity-based abstention is proposed for clinical copilots.

Why this gap is real

Two literatures motivate the split. The first is prompt sensitivity. The Butterfly Effect of Altering Prompts shows that a trailing space or a "thank you" flips classifications. When Benchmarks are Targets shows that reordering answer options moves leaderboard ranks by up to eight places. ProSA attributes such flips to low decoding confidence, where a flat output distribution lets small edits tip the answer. On the Relation between Sensitivity and Accuracy in In-Context Learning turns this into SENSEL, a method that abstains when answers are unstable under paraphrase and demonstration reordering, and it beats softmax-confidence abstention.

All of this treats instability as a sign of model ignorance. The perceptual-subjectivity literature says instability can also be a property of the stimulus. The MediaEval 2013 Emotion in Music task shows that averaging split listener ratings produces a "neutral" label no listener gave. In sleep staging, Evaluation of Automated Hypnogram Analysis on Multi-Scored Polysomnographies finds that expert agreement is only 84.1% and that disagreement clusters at stage transitions. SENSEL-style signals have never been checked against such a distribution, so nobody knows how much of a model's instability is "correct" instability.

The closest clinical work constrains the design. LLM Sensitivity Evaluation Framework for Clinical Diagnosis introduces DiagnosisQA, 4,603 MedQA-derived vignettes. It splits perturbations into a Same Answer Subset, where the diagnosis should not change, and a Different Answer Subset, where it must. GPT-4 drops from 78.95% to 3.08% accuracy when pivotal symptoms change, and it updates correctly only 50% of the time when age changes. Models are therefore both too sensitive to form and too numb to evidence. That paper evaluates only output labels, on text, against a single answer, so its framing covers our first two conditions but not the third.

Towards Explainable Multimodal Depression Recognition for Clinical Interviews shows that routing predictions through standardized symptom items can help rather than hurt: a 13.78-point absolute F1 gain. Towards Explainable Music Emotion Recognition via Mid-level Features shows that a perceptual bottleneck costs almost nothing (Pearson 0.75 versus 0.76). Both supply the method we reuse, an interpretable intermediate layer. Neither tests whether that layer affects robustness to prompts, and the depression work compresses psychiatrist disagreement into consensus.

Research Program

Input. The primary input is the multi-scored polysomnography set from Van der Plas et al.: 50 full nights, each 30-second window scored by four to seven experts. For each window we compute the guideline features that the paper's pipeline defines under American Academy of Sleep Medicine rules, such as spindle density, the Odds Ratio Product depth index and eye-movement indicators. We render them as structured text, together with the same features for the neighboring windows. Neighbors matter because that paper attributes 45.6% and 35.6% of feature importance to the preceding and following window. A secondary arm shows the model rendered signal plots of each window through a vision-capable model. This tests whether the conclusions survive when the model must perceive rather than read numbers.

Three answer formats. The model answers in one of three ways. Direct asks for the stage outright. Free rationale asks for reasoning and then the stage. Guideline bottleneck asks only for a fixed set of rule-level judgments, such as "spindle or K-complex present," "alpha rhythm over half the window," "slow-wave activity over 20%," and "rapid eye movements present." A deterministic rule then maps those judgments to a stage. Under the bottleneck, every change in the final stage is attributable to a specific named judgment by construction. This is what makes the explanation faithful rather than merely plausible.

Two perturbation families. The nuisance family contains edits with no clinical content: whitespace, option-label symbols, option order, polite tokens, and feature listing order. These are drawn from the perturbation types in the Butterfly Effect and leaderboard papers. The interpretive family contains edits a real scorer might plausibly adopt. Examples are "score conservatively at transitions," "weight continuity with the previous window," and "treat borderline spindle evidence as present." We also rotate among persona framings as expert scorers from different training traditions, in the style of ProSA's role-player templates. Each interpretive template is written and screened by sleep technologists blind to the model outputs. This screening guards against us designing prompts to produce disagreement.

Measurements. For window $i$, let $q_i$ be the empirical distribution of expert stages. Let $p_i^{\mathrm{int}}$ be the distribution of model stages across the $K$ interpretive variants, and $f_i^{\mathrm{nui}}$ the fraction of nuisance variants whose answer differs from the unperturbed answer. The central quantities are these three:

$$B = \mathbb{E}_{i:\,q_i \text{ unanimous}}\big[f_i^{\mathrm{nui}}\big], \qquad A = \mathrm{AUROC}\big(H(p_i^{\mathrm{int}}) \rightarrow \mathbb{1}[q_i \text{ not unanimous}]\big), \qquad D = \mathbb{E}_i\big[\mathrm{JS}(p_i^{\mathrm{int}}, q_i)\big].$$

Here $B$ is brittleness residue: how often the answer flips under meaningless edits on windows every expert agreed on. $A$ asks whether the entropy $H$ of interpretive variation identifies the windows experts disagreed on. $D$ is the Jensen–Shannon divergence between where the model's answers land and where the experts' answers land. Calibrated variability means low $B$, high $A$ and low $D$.

For the bottleneck format we add a localization check. On windows where the stage changes across interpretive variants, we record which rule-level judgment changed. We then ask technologists whether that judgment is the one they consider borderline in that window.

Evidence-responsiveness on text. On DiagnosisQA, the same three answer formats are applied with a bottleneck built from guideline-level findings rather than sleep features. The Same Answer Subset serves as the nuisance test and the Different Answer Subset as the evidence test. This checks whether the bottleneck's effect generalizes beyond sleep. It also checks whether the bottleneck repairs evidence-numbness or only suppresses phrasing noise.

Models. We run open-weight instruction models at two sizes, so that token probabilities are available for confidence baselines, plus one closed model. ProSA's observation that larger models can be more sensitive instance by instance makes model size a variable of interest rather than a nuisance.

Baselines

The naive reference is the direct-answer format under a single fixed template. Its accuracy against the per-window majority vote, and its softmax entropy as a disagreement predictor, set the floor. Any claim about the bottleneck must beat this on $B$, $A$ and $D$ at comparable accuracy.

The strongest existing uncertainty method is SENSEL from On the Relation between Sensitivity and Accuracy in In-Context Learning. It computes instability under instruction paraphrase and demonstration reordering, without separating perturbation families. Its reported gain is +4.1 points of area under the F1–coverage curve over maximum softmax probability on GPT-J-6B, averaged over ten classification datasets. We run it on the same models with the same number of forward passes per window as our interpretive family, and with paraphrase templates drawn from the same template budget. Two results would count as beating it. The first is higher $A$, meaning our interpretive-family entropy predicts expert disagreement better than SENSEL's undifferentiated flip rate. The second is lower $D$ at equal compute, meaning the answers land where experts land and not merely somewhere else. If SENSEL matches us on $A$ and $D$, the family split adds nothing, and we will report that.

The strongest non-language-model reference is the guideline-feature random forest of Van der Plas et al. It reaches 82.7% accuracy against the 84.1% inter-expert agreement rate, and 97.8% accuracy on the 63.3% of windows it marks certain. We retrain it on the same features and splits. Its certainty is a competing predictor of expert disagreement. It bounds how much disagreement is recoverable from these features at all, and we report our $A$ as a fraction of its $A$. For the language model's own stage accuracy on this multi-scored set, there is no reported result found on Lacuna, so the naive reference above establishes it.

The key ablation keeps the bottleneck's intermediate judgments but lets the model state the stage itself instead of applying the rule. A second ablation replaces the interpretive family with temperature sampling at matched sample count. The first isolates whether faithfulness comes from the deterministic mapping. The second isolates whether structured prompt variation carries information beyond ordinary sampling randomness.

On DiagnosisQA, the reference numbers are GPT-4's 78.95% baseline accuracy, 3.08% on the symptom-changed Different Answer Subset, and 50% correct updates under age changes, all from LLM Sensitivity Evaluation Framework for Clinical Diagnosis. Beating them means a higher Different Answer Subset score without a drop on the Same Answer Subset.

Why not just…?

Why not just use SENSEL's flip rate as the ambiguity signal? It is the most serious threat, because on these windows brittleness and ambiguity may co-occur. Low confidence in ProSA's sense is plausibly highest exactly at stage transitions, so a single flip rate might already predict expert disagreement well. The design settles this directly. $B$ measures on unanimous windows, where any flip is pure brittleness. $D$ asks whether flips land on expert-chosen stages, which a family-blind flip rate cannot evaluate. If SENSEL's $A$ and $D$ equal ours, we drop the decomposition claim and keep the bottleneck result.

Why not just use the random forest's certainty and skip the language model? For predicting disagreement, it may well win. It already isolates a highly reliable 63.3% of windows. But the proposal's object is the language-model copilot that clinicians are being offered, such as the interactive assistant studied in Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot. The random forest cannot tell anyone whether that copilot's changing explanations are trustworthy. We use it as the ceiling, not as a rival we must beat.

Why not just sample at higher temperature, as self-consistency does? Sampling variance reflects the decoder's spread under one framing. Clinical disagreement plausibly arises from different framings of borderline evidence. The matched-sample ablation tests this. If temperature sampling gives equal $D$, the interpretive family is unnecessary, and the practical recommendation simplifies.

Why not just tune the prompt, or add few-shot examples, until it is robust? Tuning can reduce $B$. But DiagnosisQA shows that well-formatted prompting leaves models numb to changed evidence, and SENSEL's authors show that demonstration order is itself a new source of instability. Tuning also targets a single consensus answer, which by construction drives $D$ toward the mismatch that averaging produces in the MediaEval findings. We include a validation-tuned best template as an extra arm to show this.

Proposed Experiments and Decisive Checks

The first decisive check is on unanimous windows. If the bottleneck does not lower $B$ relative to direct answering at matched accuracy, the claim that structured intermediate judgments shield against phrasing fails.

The second check is on contested windows. We ask whether interpretive-family variation predicts disagreement better than both SENSEL and softmax entropy ($A$), and whether its answer distribution is closer to the experts' distribution ($D$). We use a permutation test across nights, because windows within a night are correlated.

The third check is localization. On contested windows, the technologist-rated agreement that the changed judgment is the genuinely borderline one must exceed a shuffled-judgment control.

The fourth check, on DiagnosisQA, asks whether the bottleneck raises Different Answer Subset accuracy above GPT-4's 3.08% reference while holding the Same Answer Subset steady. If it only helps the latter, the bottleneck suppresses noise but does not cure evidence-numbness, and we will say so.

The main practical risk is access to the multi-scored recordings. If they are unavailable, the fallback is to collect multi-rater stage labels on a small public subset, with the same analysis at reduced power. Consensus-only sets such as DAIC-Explain cannot substitute, because they lack the expert distribution that the third condition requires.