Conditional flow models struggle to predict unseen scientific interactions
Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects
Machine Learning
Summary
When scientific models predict how one factor changes the effect of another, they often miss or invent these interactions even if their overall predictions seem accurate. The authors studied models that generate data based on conditions and found that these models learn separate effects first but can hallucinate interactions that don’t exist. Whether models correctly predict interactions depends on how similar the new situation is to what they’ve seen before. They also point out that standard accuracy measures can hide the fact that models might not be capturing the true interaction effects.
What this means in practice
- •For pharmaceutical researchers: Assess model reliability when predicting drug interactions beyond observed data to avoid misleading conclusions in drug development.
- •For bioinformatics teams: Improve interpretation of cellular perturbation models by checking if predicted interactions reflect true biological effects or hallucinations.
Authors
Sam Gijsen, Moritz Seiler, Kerstin Ritter
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.