Large language models are now being tested as assistants for one of the slowest steps in clinical genetics: reading papers that report lab experiments on a DNA variant and deciding whether that evidence supports calling the variant disease-causing or harmless. On the one public benchmark for this task, the results look excellent. But every variant in it, and usually its published interpretation, sits in open databases and papers the models were almost certainly trained on. So we cannot tell whether a model read the experiment or recalled the answer. We propose to find out by editing the evidence. Domain experts, with help from a language model that drafts the edits, rewrite real paper abstracts so that the experimental result points the other way. For example, a variant that "reduced enzyme activity to 12% of wild type" is rewritten to have "retained 95% of wild-type activity." A model that reads should follow the edited text. A model that recalls will repeat the public label.
The test works with any model that returns text, including closed APIs. That matters because existing contamination tests need access to the model's token probabilities. It also creates a hazard that ordinary benchmarks do not have. A set of plausible, expert-polished, false functional evidence for real human variants must never reach the open web. If it leaked, future models would memorize the probe, and worse, they would absorb wrong clinical facts. That hazard is why this benchmark needs encrypted, isolated distribution, not as an add-on but as a basic requirement.
Core Claim
A realistic setting for applying these models to variants of uncertain significance is literature curation. This is the step where functional evidence gets assigned the ACMG/AMP criteria labels PS3 (supports pathogenic) or BS3 (supports benign). In that setting, how much a model relies on memorized variant knowledge can be measured without token probabilities. The measurement is the model's accuracy drop on expert-validated counterfactual versions of the same evidence. The shortfall should be concentrated on well-known variants the model can already classify with no evidence in front of it. The correct behavior here is unambiguous: PS3/BS3 is a judgment about the specific paper, so following the text is right even when it conflicts with ClinVar.
Where This Sits
The motivating evidence comes from VarLitBench and VarLitAgent. That work built 463 ClinGen-derived functional-evidence variants plus negative controls. It reports that o4-mini reaches 0.979 F1 on PS3-versus-BS3 direction, while models struggle to grade strength (Moderate versus Strong). The benchmark is built entirely from public ClinGen and PubMed text and does not test for contamination. We reuse its items and task definitions and ask whether the 0.979 reflects reading. AIVARI Agent, which reasons over OMIM and ClinVar for 300 clinical cases, has the same blind spot.
On the contamination side, three lines of work constrain the design. PaCoST compares a model's confidence on original versus paraphrased questions with a paired test. It needs white-box $P(\text{True})$, and its paraphrases keep the answer fixed. Our edits deliberately change the answer, which is what separates reading from recall. A Careful Examination of LLM Performance on Grade School Arithmetic built GSM1k from scratch with human writers and reported drops of up to 8% against GSM8k. It also found that some RL-trained models overfit without high likelihood scores, which warns against relying on probability audits alone. Writing clean items from scratch is not an option in genomics, because the underlying variants and papers are the scarce resource. Editing existing items is the feasible analogue. On the Fragility of Benchmark Contamination Detection in Reasoning Models reinforces that passive detectors are easy to evade. Genomics has its own version of this problem: Guidelines for releasing a variant effect predictor names train/test circularity through ClinVar as pervasive, but offers only voluntary disclosure as a remedy.
On distribution, P⁴Bench uses zero-knowledge proofs, which let a model provider prove it solved a problem without revealing the answer. The method only works when correctness can be checked inside an arithmetic circuit (SAT, discrete log). A PS3/BS3 label is an expert judgment, not a checkable constraint, so that route is closed here. Stop Uploading Test Data in Plain Text recommends password-protected archives and no-derivatives licenses. It also documents a 4.2× ratio of papers that used a leaderboard's test set to models that officially submitted to it, which shows how often such sets circulate outside official channels. We borrow its concrete measures but do not rely on voluntary compliance. On the synthesis side, we borrow the author–reviewer pattern from Building a Role Specified Open-Domain Dialogue System, where a model drafts and humans filter, with a reported 13-fold efficiency gain. BrokenMath warns that model-drafted content can be persuasive and wrong, which is why expert review is necessary rather than cosmetic.
Research Program
The first component is the probe set. For each VarLitBench positive item, a drafting model proposes a minimal edit that reverses the direction of the functional result. The edit changes only the assay readouts, the comparison with controls, and the authors' concluding sentence. Gene, variant nomenclature, assay type, and writing style stay the same.
A second family of edits leaves direction alone and shifts evidence strength. These edits add or remove replicate counts, validation controls, and benchmark variants of known effect, so that an item moves between the Moderate and Strong levels. This targets the weakness VarLitBench reports.
Two clinical curators independently review each draft for three things: internal biological consistency, whether the new label is unambiguous under ClinGen's functional-evidence guidance, and whether the edit left visible traces. Items survive only if both curators agree. We budget for about 300 direction flips and 150 strength shifts.
The second component is a pair of per-variant memorization measures. The first is a closed-book query that gives only the gene and variant and asks for the published functional conclusion. The second is a public-prominence score, based on counts of ClinVar submissions and PubMed mentions. We will stratify every result by both measures.
The third component is distribution. The probe set is never published in plain text. Open-weight models are evaluated by sending the model to the data: they run inside a sandbox that holds the items and returns only scores, in the style of TIREx. API models are queried under zero-retention terms. For every item we publish a cryptographic hash, which is a public fingerprint that proves an item existed without revealing it, together with an embedded canary string. With these, any later leak can be detected, and scores can be re-audited if items are revealed to vetted curators. This is weaker than a zero-knowledge proof. Unlike a zero-knowledge proof, it can actually be applied to judgment-based labels.
Baselines
The naive reference point is VarLitBench accuracy taken at face value. It assumes contamination is irrelevant and that the reported 0.979 F1 direction score (o4-mini, from VarLitBench) measures reading. We will re-run the same models and prompts, in both Direct and Agentic modes, on the original items to reproduce that number under our settings before comparing anything.
The strongest existing detector is PaCoST (PaCoST). We will run it on open-weight models only, since it needs token probabilities, using the same variants and the same compute and paraphrase-generation budget. The file reports no PaCoST number on VarLitBench or any genomic set; no reported result found on Lacuna for PaCoST in this domain. Our probe beats PaCoST if it separates deliberately contaminated from clean variants (see the injection experiment below) with higher area under the ROC curve, and if it also works on API models where PaCoST cannot run at all.
The closed-book baseline asks each model to classify the variant from its identifier alone. It is the cheapest possible memorization detector, so our probe must add information beyond it.
The key ablation removes expert review. It runs the full pipeline on automatically drafted counterfactuals that no curator has checked. This isolates whether the human step matters. It does if unreviewed edits either produce spurious drops (the model is correctly rejecting incoherent evidence) or are more detectable as edited text.
Why not just…?
Why not just ask the model closed-book whether it knows the variant's classification? Knowing a fact does not mean using it when evidence is present. A model can recall ClinVar and still read the abstract faithfully. The clinical risk lies specifically in memory overriding evidence, and only the counterfactual gap measures that. Our experiments test the claim directly. If closed-book accuracy predicted counterfactual failure almost perfectly across variants and models, the cheap test would be enough and we would say so. We expect only partial correlation, because models differ in how strongly they defer to context.
Why not just use a temporal split of ClinGen curations published after each model's training cutoff? This is a real competitor, and we include it as a second baseline. Its weaknesses are practical. Cutoffs for closed models are uncertain. Post-cutoff curations are few and often concern variants discussed earlier in other sources. Synthetic training data from an already-contaminated model can pass memorized labels downstream despite a nominal cutoff (Stop Uploading Test Data in Plain Text). If the temporal-split drop matches our counterfactual gap on the same models, the temporal split is the simpler tool for the direction-level question. The probes would still be needed for strength grading, which a temporal split cannot target.
Why not just release the counterfactuals openly with a canary string, or in a password-protected zip? A canary only detects a leak after the fact, and a password-protected zip is only a speed bump (Stop Uploading Test Data in Plain Text). For ordinary benchmarks a leak spoils the measurement. Here a leak also puts fabricated evidence about real human variants into the corpora that clinical tools are trained on. We think that asymmetry justifies isolated execution, although it costs openness.
Why not just generate the counterfactuals automatically? VarLitBench shows models cannot reliably grade evidence strength, and BrokenMath shows model-written content fools reviewers. Unreviewed strength edits are therefore likely to carry wrong labels. The no-review ablation measures how much this happens.
Proposed Experiments
The primary measurement is the reliance gap. For each model we take direction F1 on the original items minus F1 on the flipped items. We also report the prior-adherence rate, meaning the fraction of flipped items where the model outputs the public label. Everything is stratified by closed-book correctness and by prominence. The strength-shift items are scored with Moderate/Strong accuracy.
The decisive validation is a controlled injection experiment, similar to PaCoST's use of the uncontaminated WMDP benchmark. We fine-tune an open-weight model on ClinGen curation text for a random half of the variants and then run the probe. If the probe is valid, the reliance gap and prior-adherence rate should rise sharply on the injected half and stay flat on the held-out half. If this does not happen, the probe is not measuring memorization.
A naturalness check guards against a confound: models might reject flipped items because the edits look artificial, not because of memory. Two curators who did not write the edits, plus a fine-tuned classifier, try to tell original from edited abstracts. We require near-chance discrimination. Unfiltered automatic edits provide the comparison point.
The finding would be clean if three things hold together: the gap is near zero on obscure variants, large on prominent ones, and absent under a no-memory control (flipped items for the same variants with gene and variant identifiers replaced by unrelated ones of the same type). That pattern would show that high curation accuracy partly reflects recall, and it would give genomics benchmarks a contamination test that works through a text-only API.
References
- VarLitBench and VarLitAgent for Benchmarking and Automating LLM-Assisted Functional Evidence Curation in Genomic Variant Interpretation
- AIVARI Agent: An Evidence-Grounded Agentic LLM for Variant Reportability and Interpretation
- PaCoST: Paired Confidence Significance Testing for Benchmark Contamination Detection in Large Language Models
- A Careful Examination of Large Language Model Performance on Grade School Arithmetic
- On The Fragility of Benchmark Contamination Detection in Reasoning Models
- Guidelines for releasing a variant effect predictor
- $\text{P}^4$Bench: Contamination-Proof, Publicly Verifiable, and Privacy-Preserving LLM Evaluation via Zero-Knowledge Proofs
- Stop Uploading Test Data in Plain Text: Practical Strategies for Mitigating Data Contamination by Evaluation Benchmarks
- Building a Role Specified Open-Domain Dialogue System Leveraging Large-Scale Language Models
- BrokenMath: A Benchmark for Sycophancy in Theorem Proving with LLMs
- The Information Retrieval Experiment Platform (Extended Abstract)