Conditional generative models are becoming common tools for scientific prediction, especially when researchers want answers beyond the direct support of their training data. But many scientific questions are about interactions: does one intervention alter the effect of another? A new preprint on arXiv addresses this specific problem.
The paper, titled "Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects," argues that flow models—a type of conditional generative model—can produce misleading results when asked to predict such interactions. According to the abstract, the models "hallucinate" non-additive effects, meaning they generate apparent interaction patterns that are not grounded in the data.
The abstract is brief and does not yet provide details on the experimental setup or the magnitude of the problem. Still, the claim is a cautionary note for anyone using generative models to extrapolate to new combinations of interventions. The preprint is available on arXiv under identifier 2610.20288.