Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects
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.
Related
- MINT: Predicting Enzyme-Cofactor Specificity by Multimodal Knowledge Transfer from Protein-Ligand Interactions
- Adaptive gating-based fusion of sequential and local features for predicting drug-drug interactions
- Predicting cortical signal mixing in coarse-resolution fMRI: a forward-operator analysis of surface mapping and the conditional benefit of correction
- Host–virus interactions dominate fitness and conceal additive genetic variation
- Predicting Differential Protein-Protein Interactions
BibTeX
@misc{gijsen2026predicting,
title = {{Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects}},
author = {Sam Gijsen and Moritz Seiler and Kerstin Ritter},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.20288},
eprint = {2610.20288},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.20288},
}
RIS
TY - UNPB
TI - Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects
AU - Sam Gijsen
AU - Moritz Seiler
AU - Kerstin Ritter
PY - 2026
DA - 2026/10/14
PB - arXiv
DO - 10.48550/arxiv.2610.20288
UR - https://arxiv.org/abs/2610.20288
AB - 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.
ER -