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

Portuguese

No entanto, as questões científicas frequentemente se preocupam se uma intervenção altera o efeito de outra. Essas interações podem ser de interesse primário enquanto explicam pouco da variação total, permitindo que os modelos apareçam precisos em geral sem recuperar o efeito científico relevante. Ainda não está claro como esses modelos aprendem interações alucinadas e quando estes se extrapolam com sucesso para condições invisíveis. Portanto, estudamos a recuperação de interações em modelos alucinados de fluxo usando um simulador de série de tempo fisiológico, dados de interação de dose de drogas, perturbações de células combinatórias e gerações de conformadores moleculares, onde separamos a recuperação da estrutura de interação alucinada. Na configuração controlada, os efeitos individuais são aprendidos muito antes das interações e modelos

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