Conclusion Fidelity of AI Agents in Synthetic Omics Environments.

Testbed design

1. A real cohort as the reference. The study uses the TCGA breast cancer cohort (TCGA-BRCA) as distributed by UCSC Xena: RNA-seq expression, PAM50 subtype calls, receptor status, stage, histology, age and somatic mutation calls for five frequently mutated genes. It is restricted to the 844 primary tumours that have a PAM50 subtype call and to 219 genes, 69 with an established role in breast cancer and 150 that are little studied.
2. Generators with planted flaws. The baseline generator is a Gaussian copula fitted within each PAM50 subtype, and two deep tabular generators, CTGAN and TVAE, are trained with default settings. Flaws are introduced as deterministic edits to the sampled data, so an edited cohort differs from the baseline only in the edited gene, and the strength of each edit is varied in steps. One edit weakens and then reverses the correlation of GRB7 with ERBB2. GRB7 is co-amplified with ERBB2 and is a textbook marker of HER2-enriched tumours. The other edit gives an unrelated, little-studied gene (C8orf84) an artificial difference in HER2-enriched tumours. In the real cohort, that gene’s difference between HER2-enriched and other tumours is −0.09 standard deviations.
3. Agents on three open-weight models. The agents are built on three open-weight models from different families: Qwen3.6-35B-A3B with its reasoning mode enabled, which is used in all experiments, and Gemma-4-26B and Nemotron-3-Super-120B, which are evaluated on a subset of the conditions. Each model is served locally with vLLM and given a Python tool, a research brief, the cohort as a CSV file and a data dictionary. The agents are not told whether the data are real or synthetic. They choose their own analyses and finish with a written report and a list of structured claims. Unless stated otherwise, the results below are for Qwen3.6.
4. Scoring against the real cohort. Claims are scored automatically, without an LLM judge. A claim holds if its standardised effect (Cohen’s d or Spearman’s ρ) has the stated direction and clears a fixed threshold. Each claim is tested on the real cohort, on the synthetic sample the agent analysed and on a large sample from the same generator, which separates correct claims from artefacts of the generator, artefacts of sampling and errors made by the agent. Correct here means agreement with the real cohort, which is an empirical reference and not causal truth. Precision is the share of claims that are correct, and recall the share of real findings that are reported.

Key findings

1. Moderately imperfect generators give claims that almost all hold on the real cohort but miss real findings. Although the copula and TVAE differ by an order of magnitude in maximum mean discrepancy (MMD) to the real cohort, nearly all the claims the agent makes on either synthetic cohort hold on the real data. On the brief that asks for HER2 targets, precision is 0.96 on the copula and 0.99 on TVAE, against 0.99 on the real cohort itself. Recall is lower. The share of core findings reported, those claimed in at least half of the runs on real data, falls from 0.76 on real data to 0.59 on the copula and 0.49 on TVAE. With eight runs per cohort these differences are uncertain. Many false claims appear only when a generator fails to learn the structure of the data, as CTGAN did with default settings: it lost all gene–gene correlations and subtype differences, and the precision of the agent’s claims fell to 0.54.
2. A weakened true relationship disappears from the agents’ conclusions and is seldom reported in reverse. As the correlation between GRB7 and ERBB2 is weakened and then reversed, the agent reports GRB7 as a HER2 marker less and less often: in 12 of 16 runs with the data unchanged, in 2 to 5 of 8 as the correlation is weakened to zero and in 1 of 24 once it is reversed. On the real cohort it does so in 7 of 8 runs. GRB7 starts to disappear while it is still strongly enriched in the synthetic data (d = 1.2), because the agent reports a short ranked list of findings, from which a weakened signal drops out early. At full reversal the agent mostly omits GRB7 and states the false direction in only 1 of 8 runs. Gemma-4-26B drops GRB7 under full reversal in all 8 runs.
3. A fabricated signal is reported as a marker once it nears the strength of genuine markers. When an unrelated, little-studied gene is given an artificial difference in HER2-enriched tumours, the agent almost never reports it below about 1.3 standard deviations (2 of 48 runs). From 1.5 standard deviations onwards it reports the gene as a HER2 marker in 20 of 32 runs, against 0 of 16 on the unedited copula, and no run describes it as unexpected or anomalous (0 of 72 runs with a fabricated effect). The threshold depends on the strength of the genuine markers. On the basal-like brief, whose genuine markers reach about 2.4 standard deviations, the same fabrication is reported in 1 of 8 runs at 1.74 standard deviations and in 5 of 8 at 2.08. The false marker also displaces real findings, lowering recall of the core findings from 0.59 to 0.40. Gemma-4-26B reports the fabricated gene in 5 of 8 runs without comment. Nemotron-3-Super-120B relies more on its prior knowledge and reports the fabrications less often (6 of 16 runs), mostly by omitting the gene.
4. When a fabrication contradicts textbook biology, the agent notices the conflict but rarely rejects the finding. On clean data the agent reports ESR1 as lower in HER2-enriched tumours in 23 of 24 runs. When ESR1 is shifted in the synthetic data to the same effect as the little-studied gene, 1.69 standard deviations higher in HER2-enriched tumours, the agent reports it as higher in 29 of 32 runs across both HER2 briefs. It notices the conflict in 23 runs, calling the result “unexpected” or “contrary to the clinical paradigm”, but in 21 of those it reports the finding anyway, typically attributing it to characteristics of the cohort. It rejects the finding in only two runs.
5. Standard fidelity metrics register neither flaw. For every edited cohort the correlation error, the mean absolute difference of gene–gene correlations from the real cohort, is 0.011 to 0.013, and the MMD is at most 0.002, as for the unedited copula. In this setting the standard metrics were enough to reject a generator that failed to learn the structure of the data, such as CTGAN, but they missed the local flaws that decide individual conclusions.
6. Re-evaluating each claim on an ensemble of generators separates the claims that hold from artefacts. Each claim is re-evaluated in the agent’s generator, TVAE and three bootstrap refits of the copula, and its agreement is the share of these generators in which it holds. Across all three models and all briefs, agreement separates the 3,718 correct claims from the 188 artefacts of generator or sampling with an AUROC of 0.9998, and of 0.984 with CTGAN in place of the copula refits. It relies on the generators having different flaws. For the 119 claims about an edited signal, agreement separates the 85 false claims from the 34 true ones perfectly when only one generator has the flaw (AUROC 1.00), but not when all of them do (AUROC 0.46). Flaws that every generator shares can only be found by checking some claims on real data, and when the location of a flaw is known, few claims need checking. Checking only the 5 to 13% of claims that involve an edited gene catches 78 to 97% of the false claims in cohorts with a fabricated signal and raises precision from 0.93–0.94 to 0.98–1.00.

Why it matters

Before an agent is used to analyse a synthetic cohort, its users need to know whether its conclusions would hold on the real data, and fidelity metrics do not measure that. Synthetic data is offered as a way to work with sensitive biomedical data without giving access to the patient records themselves, and patient-level molecular data is among the most restricted. Synthetic versions of such cohorts could open them to far more researchers, and the research is increasingly carried out by AI agents that plan analyses, write code and report findings. Fidelity metrics such as the MMD measure similarity, but they do not measure whether the particular relationships an analysis depends on are preserved. For example, synthetic clinical time series can improve in conditional MMD while still failing to reproduce the relationship between fever and tachycardia, as K01’s earlier work on multimodal alignment found.

The results suggest two practices for using an agent on synthetic data. First, findings that the agent itself describes as unexpected should be checked on real data before they are reported, since the agent usually reports such a finding anyway and attributes it to the cohort. Second, a data holder can fit several independently built generators and re-evaluate each claim in all of them. Disagreement between the generators reveals the flaws specific to one of them, and a check of selected claims on real data is needed for the flaws they share.

Scope and limitations

The study uses one cohort, three agent models and two kinds of introduced flaw, so the thresholds it reports may not transfer, and 8 to 32 runs per combination of brief and cohort leave wide intervals. CTGAN was trained with default settings, and tuning may avoid its failure. The agreement score assumes that the generators’ flaws are independent, which bootstrap refits of a good copula favour, and it cannot detect a flaw that every generator shares. The testbed measures conclusion fidelity under known flaws and does not certify a generator. Three problems remain open: building agents that reject implausible findings instead of explaining them away, detecting shared flaws without real data and locating flaws as precisely as when their location is known.