acceptodds
Under review as a conference paper at ICLR 2027

DIAGRARE-BENCH: SEPARATING DISEASE PRIORS FROM PATIENT EVIDENCE IN LANGUAGE MODELS FOR CLINICAL DIAGNOSIS

Abstract

Diagnostic benchmarks tell us whether a language model selected the correct disease, but they do not identify what drove that decision. A diagnosis can rank highly because the findings in the case support it, because the disease is common, or because both signals agree. We introduce DIAGRARE-BENCH to separate these effects at the level of the model's diagnostic ranking. Given a ranked differential, we fit a Plackett-Luce choice model with two interpretable features: a disease-prior feature and a case-specific evidence feature. The fitted coefficient measures how strongly pairwise diagnostic log odds change with an evidence contrast when the prior contrast is held fixed. The estimator is model agnostic and uses ranked outputs only. We evaluate 25 open-weight checkpoints, spanning both general-purpose models and models specialized for medicine. The primary evaluation contains 984 diagnostic cases covering 82 diseases and nine organ systems; 21 checkpoints yield identifiable baseline coefficients. Evidence responsiveness is associated with top-1 accuracy (), but our main tests ask whether the coefficient predicts behavior when the evidence is changed independently of the rankings used to estimate it. The cross-model ordering transfers to all 3,562 CUPCase reports () and all 22,901 RareArena cases () under a different evidence representation. In a randomized experiment that varies evidence strength and stated prevalence independently, baseline predicts the change in target choice caused by stronger evidence (, ). In a sequential diagnosis task, it predicts whether a model revises an initial competing diagnosis after target-specific evidence arrives (, ). QLoRA adaptation, self-consistency, and test-time reasoning experiments further show that diagnostic accuracy and evidence responsiveness can change by different amounts. These results support evidence responsiveness as a distinct and measurable property of diagnostic decision behavior.

Primary area:
datasets and benchmarks
Keywords:
large language models, clinical diagnosis, differential diagnosis, model evaluation, benchmarks, evidence responsiveness

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_o8eSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Traditional diagnostic evaluations are limited as they only study accuracy, ranking quality, or validate reasoning steps. This does not disentangle the LLM’s responsiveness to evidence and the role of the disease prior / prevalence. The authors propose to fit a linear model with two features over a set of diagnosis tasks, where one feature contains the disease prevalence and the other feature contains the evidence. The coefficients are fit using a Placket-Luce likelihood using the observed diagnosis rankings from a given model. This would disentangle whether the model is responding to evidence or just diagnosing based on prior disease prevalence. The paper provides various diagnostics to show that the evidence sensitivity coefficient is correlated with accuracy, the ability to incorporate evidence of varying strength independent of the stated disease prior, and the ability to revise a diagnosis.

Critical strengths
  • Investigating whether LLMs rely on disease priors versus actual clinical evidence is an essential question for safe medical deployment.

  • The decomposition and insights are helpful for understanding a primary benchmark based on the specific featurization of evidence level. The authors evaluate sensitivity to evidence-level featurization and data distribution using two different benchmarks, showing that model ordering based on the coefficient can translate.

  • The authors evaluate across 25 LLMs and demonstrate that the proposed evidence responsiveness coefficient is strongly correlated with various model properties such as diagnostic accuracy with statistical significance. There is also strong experimental validation regarding the sensitivity of the coefficient to various design decisions.

Critical weaknesses
  • The main motivation of the paper is that standard accuracy metrics fail to reflect evidence responsiveness. However, the empirical results seem to contradict this premise. Specifically, γevidence\gamma_{evidence} appears highly correlated with Top-1 accuracy, while the prior coefficient shows no correlation. While the paper's proposed decomposition was valuable for uncovering this relationship, doesn’t this suggest that Top-1 accuracy already serves as a strong proxy for evidence responsiveness in practice? Or if certain models actually have a high γprior\gamma_{prior} and low γevidence\gamma_{evidence}, this would be an interesting finding and also useful to substantiate the practical necessity of this decomposition in uncovering weak models that standard accuracy metrics miss.

  • While evidence responsiveness measures whether a model’s diagnostic ranking is sensitive to evidence, it does not evaluate whether this sensitivity translates to accurate or well-calibrated disease diagnosis. Other diagnostic methods can be used to validate whether evidence is incorporated in the wrong direction or the model is miscalibrating its confidence relative to evidence strength. For example, one can directly assess evidence responsiveness and diagnostic accuracy across varying evidence strengths (e.g., via a similar setup to Section 4.3 / Δ_E) without needing to fit a Plackett-Luce coefficient. While the various diagnostics proposed in the paper are useful for understanding aspects of evidence sensitivity, it remains unclear how practitioners should interpret or act upon the sensitivity coefficient itself.

Questions for authors
  • To better highlight the problem, were there combinations of models and diagnostic benchmarks that the decomposition caught where the γprior\gamma_{prior} was high and γevidence\gamma_{evidence} was low, but top-1 accuracy was high?

  • Could the authors clarify how an aggregate evidence responsiveness coefficient should be preferred over more granular diagnostic evaluations? Specifically, a single global coefficient computed across diverse diagnostic tasks (based on a specific featurization of evidence-level) may mask failure modes. In clinical contexts, evaluating responsiveness at a fine-grained level, such as across specific disease categories, evidence types and strengths, may be more actionable for identifying where a model fails to incorporate evidence. My concern is whether a single aggregate sensitivity coefficient provides actionable insights or safety guarantees to practitioners.

Clarification requests and missing details:

  • Sm(x,d)S_m(x, d) is defined in the methodology section, but it is never explicitly used or referred to in the subsequent equations or text. Could the authors clarify its role or remove it if redundant?
  • Are the set of ranked candidate diagnosis (or just the number of candidates) fixed across all tasks xix_i? My understanding is that a single set of coefficients is fit across all diagnostic tasks irrespective of confounding factors such as task difficulty. Did you consider a more granular alternative of fitting coefficients separately across tasks?
  • How is the evidence feature standardized across heterogeneous diagnostic tasks? In Section 4, the paper mentions that the evidence feature is based on a log-likelihood ratio score, but it remains unclear how this is computed in practice. Specifically, how are evidence strengths normalized so that they exist on a comparable scale across different diseases and clinical domains?
Reviewer_BdsoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a measure of evidence responsiveness for diagnostic LLMs. It fits a Plackett–Luce model to diagnostic rankings using disease-prior and patient-evidence features. The resulting evidence coefficient correlates with accuracy and predicts responses to stronger evidence and later diagnostic revision. The authors evaluate it on synthetic cases, two external datasets, and several model and inference interventions.

Critical strengths

I like that the authors test whether a coefficient fitted to static rankings predicts responses to separate evidence interventions. This is more informative than simply showing a correlation with accuracy on the same cases. The approach also only requires ranked outputs, which makes it relatively easy to apply across models.

Critical weaknesses
  1. My main concern is the role of medical knowledge. Models receive case findings and candidate diagnoses, but must already know the relevant disease finding relations. I did not find an independent check of this knowledge. A model that does not know a relation and one that knows it but fails to use it can both have low evidence responsiveness. The behavioral measure remains meaningful, but it cannot distinguish these failures. The unclear sources and clinical validation of the disease profiles make this harder to assess.
  2. The coefficient’s added value over accuracy remains unclear. The authors report strong correlations with intervention effects and recovery, but do not compare their predictive value against baseline accuracy. The existing robustness checks address other explanations, but do not answer this question. Given the strong correlation between accuracy and the coefficient, better overall diagnostic ability could account for both. The adaptation experiments do not establish what additional information the coefficient provides.
Questions for authors
  1. Where do the disease profiles come from, and how were they clinically validated?
  2. Is there evidence that the models know the medical relations required by the benchmark, independently of their diagnostic performance?
  3. Does adding the coefficient improve prediction of intervention effects or recovery over accuracy alone?
Reviewer_TUJGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DIAGRARE-BENCH, a benchmark designed to disentangle the effects of disease priors and patient-specific evidence on language models’ diagnostic rankings. The authors fit a two-feature Plackett–Luce model using prior and evidence features, and define the estimated coefficient gamma_evidence as evidence responsiveness. The evaluation covers 25 open-weight models and validates the proposed metric through cross-dataset transfer, randomized evidence interventions, and sequential diagnostic revision. The results suggest that evidence responsiveness is correlated with diagnostic accuracy, but the two do not necessarily improve in tandem under model adaptation and inference-time interventions.

Critical strengths
  1. The paper addresses an important problem and offers a novel evaluation perspective by going beyond conventional diagnostic accuracy to examine whether models base their decisions on patient-specific evidence rather than disease prevalence alone.
  2. The proposed method is simple, interpretable, and model-agnostic. The Plackett–Luce estimator requires only ranked diagnostic outputs, without access to hidden states or logits, while gamma_evidence admits a clear behavioral interpretation.
  3. Cross-dataset analyses, randomized evidence interventions, and sequential revision experiments jointly support the validity of the proposed evidence-responsiveness metric, complemented by extensive robustness checks.
Critical weaknesses
  1. The two-factor linear decomposition may be too simplistic. Diagnostic rankings may also depend on disease familiarity, training frequency, candidate position, prompt wording, and patient-specific priors, whose effects could be absorbed by gamma_evidence. A single linear coefficient also cannot capture thresholds, saturation, prior–evidence interactions, or asymmetric responses to conflicting evidence.

  2. The Plackett–Luce independence-of-irrelevant-alternatives assumption may be inappropriate for clinically related diagnoses, such as disease subtypes or alternative etiologies of the same syndrome. The observed variation in coefficient magnitude across candidate sets further indicates sensitivity to candidate-set composition.

  3. The evaluation is largely closed-set: the primary task fixes 82 diseases, and the external tasks provide four candidates. It therefore evaluates ranking among predefined diagnoses rather than open-ended diagnosis, leaving candidate generation, diagnostic granularity, missing correct candidates, abstention, and uncertainty underexplored.

  4. The benchmark only partially reflects real clinical complexity. The primary data are synthetically generated, while TF-IDF similarity and PubMed counts are imperfect proxies for evidence and prior. The evaluation also excludes multimodal and longitudinal information, comorbidities, measurement error, and source reliability, limiting real-world generalizability.

  5. The incremental value of gamma_evidence over simpler baselines is unclear. The paper does not compare it sufficiently against measures such as evidence–diagnosis rank correlation, top-k inclusion of the strongest-evidence diagnosis, or answer-flip rates. Given its strong correlation with accuracy, it may partly reflect general model capability.

  6. High evidence responsiveness does not necessarily imply clinically appropriate evidence use. A model may respond strongly to both valid and misleading evidence. The irrelevant-text control does not test erroneous, adversarial, fabricated, or conflicting findings, so gamma_evidence may measure response magnitude rather than selective and robust use of reliable evidence.

Questions for authors
  1. Does gamma_evidence provide incremental predictive value beyond diagnostic accuracy and simpler baselines, such as evidence–diagnosis rank correlation, top-k agreement with the strongest-evidence diagnosis, or answer-flip rates under evidence perturbations?

  2. How sensitive are the conclusions to violations of the Plackett–Luce independence-of-irrelevant-alternatives assumption, particularly when candidates are closely related diseases or subtypes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.