Dharini24/DiagRare_Bench

DiagRare-Bench

DiagRare-Bench evaluates how strongly a language model changes its diagnostic ranking when patient evidence changes after a disease-prior feature is accounted for. The public panel contains general-purpose and medically specialized open-weight checkpoints evaluated under the same protocol.

For case $x$ and diagnosis $d$, the primary rank model is

$$U_m(d,x)={\gamma }_{\mathrm{p}\mathrm{r}\mathrm{i}\mathrm{o}\mathrm{r}}^{(m)}p(d)+{\gamma }_{\mathrm{e}\mathrm{v}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}}^{(m)}e(d,x)\mathrm{.}U\_m(d,x)=\backslash gamma\_\{\backslash mathrm\{prior\}\}^\{(m)\}p(d)+\backslash gamma\_\{\backslash mathrm\{evidence\}\}^\{(m)\}e(d,x).$$

The fitted $\gamma_{\mathrm{evidence}}$ is an output-level measure of evidence responsiveness. The paper validates its interpretation by changing the data distribution, assigning evidence independently in a randomized experiment, and measuring sequential diagnostic revision.

Paper

DiagRare-Bench: Separating Disease Priors from Patient Evidence in Language Models for Clinical Diagnosis
Dharini Raghavan, Amritpal Singh
ICLR 2027 Conference Submission

• Paper: https://openreview.net/forum?id=0Eua4EkYht
• Project page: https://diagrare-bench-dhariniraghavan2001-2901.vercel.app
• Code: https://github.com/rdharini2001/DiagRare
• Leaderboard: https://diagrare-bench-dhariniraghavan2001-2901.vercel.app/leaderboard.html

Benchmark at a glance

• 984 primary diagnostic cases
• 82 diseases across nine organ systems
• 25 open-weight checkpoints in the released model panel
• 27 target/confounder pairs in the randomized 3 x 3 experiment
• 82 target/confounder pairs in the sequential revision task
• Negative-control cases with clinically irrelevant text
• QLoRA supervised adaptation examples used in the paper

Main results

Across the 21 checkpoints with identifiable baseline estimates, evidence responsiveness is associated with top-1 accuracy ($r=0.843$). The cross-model ordering transfers to CUPCase ($r=0.662$) and RareArena ($r=0.790$) under a different evidence representation. Baseline evidence responsiveness predicts the finite-difference effect of randomized evidence ($r=0.689$, $p=5.5\times10^{-4}$) and sequential recovery ($r=0.860$, $p=9.8\times10^{-6}$).

The prior coefficient does not receive a symmetric validation claim; it is used as a conditioning coordinate for the specified disease-prior feature.

Files

Path	Description
structured/vignettes.csv	984 primary diagnostic cases
structured/ontology.csv	disease-frequency metadata and disease-finding relations
structured/vignettes_negative_control.csv	irrelevant-text negative-control cases
interventions/causal_grid.csv	randomized evidence/stated-prevalence prompts
interventions/sequential_anchoring_cases.csv	two-step sequential revision cases
metadata/model_metadata.csv	checkpoint identities, families, parameter counts, upstream HF repositories, specialization flag
finetuning/debias_sft.jsonl	supervised adaptation examples used in the QLoRA experiment
DATASHEET.md	construction, intended use, and limitations

Download

pip install -U "huggingface_hub>=0.34"
hf download Dharini24/DiagRare_Bench --repo-type dataset --local-dir ./DiagRare_Bench

External case-report evaluations

The paper also evaluates all 3,562 CUPCase reports and all 22,901 RareArena cases. Those datasets remain in their original repositories and retain their original licenses and attribution; they are not repackaged in this dataset.

Intended use and safety

DiagRare-Bench is a research benchmark for studying diagnostic decision behavior. It does not evaluate clinical safety and does not establish that any language model is suitable for patient care. A high evidence-responsiveness value should not be interpreted as a deployment recommendation or safety certificate.

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}
}