What the audit examined
Jian Yang and six co-authors present an empirical audit of evaluation artifacts and possible data leakage in AI systems for rare-disease diagnosis. The work is available as a preprint on arXiv, submitted on October 24, 2026. The arXiv record identifies it as version 1; the available information does not indicate that it has undergone peer review.
The abstract describes a repository of more than 60,000 cases from public and independent clinical cohorts. The authors reconstructed an agentic diagnostic framework and compared results across different evaluation setups. The size of the repository alone does not reveal how many cases were included in each analysis, how cases were distributed among cohorts, or how independent those cases were of each system’s training data.
Why data separation matters
A benchmark is intended to estimate how a system might perform on cases it has not seen. If information related to a case also appears in the candidate set or in data used to build or evaluate the system, the score may partly reflect that contamination rather than the system’s ability to generalize to new cases.
For systems that rank diagnoses among several alternatives, Recall@1 measures the share of cases in which the system’s top-ranked answer matches the reference answer. It is a classification metric for a particular setup: it is not equivalent to the rate of correct diagnoses in real clinical consultations, and does not by itself show whether a system is safe, useful, or suitable for informing medical decisions.
The reported change in the agentic system
According to the abstract, mean cross-dataset Recall@1 fell from 54.3% to 40.3% when the evaluation shifted from a conventional one-vs-rest setup to strict separation between datasets. That is a difference of 14 percentage points. The authors also report a Hedges’ g effect size of 0.94, a standardized measure describing the size of the observed difference in their analysis.
The finding supports a limited conclusion: in the authors’ reconstructed evaluation, candidate-set configuration and data separation were associated with a substantial difference in score. It does not establish that one cause accounts for the entire difference, nor does it justify generalizing the percentages to other models, databases, diseases, or clinical settings. The abstract does not break down the results by cohort or provide uncertainty intervals here.
Reported quantitative results
Comparisons as stated in the preprint abstract; these figures should not be interpreted as clinical performance.
| Analysis | Setup or dataset | Reported result |
|---|---|---|
| Agentic system | Conventional one-vs-rest evaluation | Mean cross-dataset Recall@1: 54.3% |
| Agentic system | Strict separation between datasets | Mean cross-dataset Recall@1: 40.3% |
| Supervised classifiers | Internal mean | Mean accuracy: 42.9% |
| Supervised classifiers | External cohorts | Mean accuracy: 16.7% |
Supervised classifiers and other sources of variation
The same abstract reports that the mean accuracy of supervised classifiers fell from 42.9% in an internal evaluation to 16.7% across external cohorts. The authors link this difference to spatial overfitting and report a Cohen’s d effect size of 2.01. The comparison indicates that observed performance can change substantially when evaluation moves beyond an internal dataset, but it is not enough to determine how much of the difference reflects overfitting and how much reflects differences between cohorts, data, or protocols.
The paper also says it examined retrospective treatment cues in clinical narratives and temporal misalignment in the knowledge bases used by models. According to the abstract, these factors can affect model inferences or lead to performance fluctuations. It does not provide separate figures for each mechanism there, so the contribution of each to the reported declines cannot be quantified from the abstract.
What to check when reading a score
A benchmark score is meaningful only in the context of how the data were prepared and what the comparison measures.
- 01Identify the metric: Recall@1 and accuracy describe different outcomes and depend on the evaluation setup.
- 02Check whether the evaluation is internal or external and how the datasets were separated.
- 03Consider whether clinical information includes retrospective clues that might reveal the diagnosis.
- 04Keep experimental scores distinct from any conclusion about clinical usefulness or safety.
What the findings do—and do not—show
The authors’ main conclusion is that some evaluation protocols may include leakage or other artifacts and therefore give an overly favorable picture of rare-disease diagnostic systems. They propose combining strict data separation with standardized safeguards for benchmarks. This is a methodological recommendation; the abstract does not demonstrate that applying it guarantees bias-free evaluations or clinically valid results.
The evidence remains that of an audit described in a preprint. The available information does not allow independent verification of the data, code, and protocol, or establish that the cohorts were entirely independent of all original training and evaluation datasets. Nor does it show that the audited systems improve or worsen patient outcomes. The reported declines therefore matter when interpreting the evaluated benchmark scores, but they do not replace clinical validation studies designed to measure performance and risk in care settings.
Open questions
- The available information does not establish whether every cohort was independent of the original training and evaluation data.
- The abstract does not provide enough disaggregated results to attribute a specific share of the declines to each leakage or misalignment mechanism.
- Independent replication and clinical validation of the findings have not been established here.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction