Dharini24/DiagRare_Bench

DiagRare-Bench

Paper · Code · Project page · Leaderboard

DiagRare-Bench studies a simple question in diagnostic model evaluation: when the patient evidence changes, how much does the model's diagnostic ranking move with it after a disease-prior feature is accounted for?

The benchmark contains 984 structured diagnostic cases spanning 82 diseases and nine organ systems, together with randomized and sequential experiments that test the interpretation of the fitted evidence coefficient under independent changes to the input.

Tasks and files

Configuration	Unit	Rows	File
primary	diagnostic case	984	structured/vignettes.csv
ontology	disease profile	82	structured/ontology.csv
randomized_intervention	assigned evidence/prior prompt	243	interventions/causal_grid.csv
sequential_revision	evidence step	328	interventions/sequential_anchoring_cases.csv
negative_control	diagnostic case with irrelevant text	984	structured/vignettes_negative_control.csv
finetuning	supervised adaptation example	72	finetuning/debias_sft.jsonl
model_metadata	checkpoint	25	metadata/model_metadata.csv

The paper also evaluates CUPCase and RareArena. Those datasets are not redistributed here; obtain them from their original repositories and follow their original licenses.

Loading the data

from datasets import load_dataset

repo_id = "Dharini24/DiagRare_Bench"
primary = load_dataset(repo_id, "primary")
randomized = load_dataset(repo_id, "randomized_intervention")
sequential = load_dataset(repo_id, "sequential_revision")

print(primary["test"][0])

For a local clone, the files can also be read directly with pandas:

import pandas as pd

cases = pd.read_csv("structured/vignettes.csv")
ontology = pd.read_csv("structured/ontology.csv")

Primary benchmark

Each primary case records a target disease, organ system, disease-frequency attribute, observed findings, explicitly absent findings, and difficulty. The cases are generated from the released disease ontology rather than patient records. Each disease contributes 12 cases across three difficulty levels.

For candidate diagnosis $d$ in case $x$, the evaluation represents disease prior and case-specific evidence separately and fits a Plackett-Luce model to the model's ranked differential. The resulting $\gamma_{\mathrm{evidence}}$ is interpreted comparatively under a fixed prompt, candidate set, feature scale, and decoding rule.

Randomized intervention

The randomized experiment uses 27 target/confounder disease pairs. Evidence strength takes three levels (weak, medium, strong), while stated target prevalence independently takes three levels (rarer, equal, more common), giving 243 prompts per model. This experiment supports causal statements about changing the presented prompt evidence and the model output within the experiment; it is not a claim about causal structure inside the network.

Sequential revision

The sequential task contains 82 target/confounder pairs presented in both evidence orders. The final information is the same across orders. The task measures whether a model recovers from an initial competing diagnosis when target-specific evidence arrives and whether the final diagnosis depends on presentation order.

Negative control

structured/vignettes_negative_control.csv adds a clinically irrelevant sentence to the primary cases. It is used to distinguish responsiveness to diagnostic evidence from generic sensitivity to textual perturbations.

Model panel

The released metadata lists 25 open-weight checkpoints spanning general-purpose and medically specialized models. The same diagnostic protocol is used for both groups. Model-specific eligibility differs by analysis; missing estimates are not imputed.

Fields

The primary cases include vignette_id, target_disease, organ_system, is_rare, prevalence, difficulty, positive_findings, and negative_findings, together with counts and split metadata. The ontology records disease frequency and required, supporting, and excluding findings. Intervention files record target/confounder identity, assigned evidence or order condition, and the findings shown to the model.

Version and provenance

This release corresponds to the benchmark version used in the ICLR 2027 submission. The structured cases and interventions are benchmark-owned data generated from the released ontology and experimental construction code. CUPCase and RareArena are external resources used only for transfer evaluation and are not included here.

Scope and limitations

DiagRare-Bench is a research benchmark for model behavior in clinical diagnosis. The primary ontology is finite and intentionally simplified so that prior and evidence can be represented separately. The evidence score is an evaluation construct rather than a calibrated patient likelihood, and results are protocol-dependent. A high evidence-responsiveness value is not evidence that a model is safe or appropriate for clinical deployment.

License and external resources

Benchmark-owned data in this repository are released under Apache-2.0. CUPCase, RareArena, and model checkpoints are external resources and remain governed by their respective licenses and terms.

Citation

@misc{raghavan2026diagrarebench,
  title  = {{DiagRare-Bench: Separating Disease Priors from Patient Evidence in Language Models for Clinical Diagnosis}},
  author = {Raghavan, Dharini and Singh, Amritpal},
  year   = {2026},
  note   = {ICLR 2027 Conference Submission},
  url    = {https://openreview.net/forum?id=0Eua4EkYht}
}