Newest papers

Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects

Sam Christian Jan Gijsen, Moritz Vinzent Seiler, Kerstin Ritter · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.20288 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los modelos generativos condicionales se utilizan cada vez más para la predicción científica más allá del soporte de entrenamiento directo. Sin embargo, las preguntas científicas a menudo se preocupan de si una intervención cambia el efecto de otra. Estas interacciones pueden ser de interés primario al explicar poco de la variación total, permitiendo que los modelos aparezcan precisos en general sin recuperar el efecto científico relevante. Todavía no está claro cómo estos modelos aprenden interacciones alucinadas y cuando estas extrapolan con éxito a condiciones invisibles. Por lo tanto, estudiamos la recuperación de la interacción en modelos de correspondencia de flujo utilizando un simulador de serie de tiempo fisiológico, datos de respuesta a dosis de medicamentos, perturbaciones de células combinatoriales y generaciones de conformadores moleculares, donde separamos la recuperación de la estructura de interacción alucinada. En

Machine translation (Meta M2M100); the English original follows.

Abstract

Conditional generative models are increasingly used for scientific prediction beyond the direct training support. Yet scientific questions often concern whether one intervention changes the effect of another. Such interactions can be of primary interest while explaining little of the total variation, allowing models to appear accurate overall without recovering the relevant scientific effect. It remains unclear how these models learn interactions and when these successfully extrapolate to unseen conditions. We therefore study interaction recovery in flow-matching models using a physiological time-series simulator, drug dose-response data, combinatorial cell perturbations, and molecular conformer generations, where we separate recovery from hallucinated interaction structure. In the controlled setting, individual effects are learned much earlier than interactions and models hallucinate interactions where none exist. Across settings, we find qualitatively different generalization problems, with successful prediction depending on whether condition structure continues from observed to held-out targets. When it does, interactions are recoverable. If not, models can generate plausible responses while missing the interaction. We furthermore find that hallucinations increase as targets become less identifiable, and that successful interaction recovery in output space can fail under nonlinear readouts. Finally, we show that aggregate fidelity metrics can conceal whether models learn and extrapolate interactions.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Hallucinating, Generalization, Artificial intelligence, Overfitting, Computer science, Machine learning

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange