Cross-modal Evidential Rupture Detection and Localization in Scientific Research

Despite the growing success of AI systems in scientific research, hallucinations and unsupported outputs remain a critical concern, particularly when plausible errors propagate across stages of scientific analysis and lead to consequential conclusions. In particular, AI systems increasingly generate heterogeneous scientific artifacts, including text, mathematical formulations, code, and experimental results. While individual artifacts may appear plausible, the evidential relations connecting them can lose support, making such cross-artifact inconsistencies difficult to detect. We propose a novel notion named cross-modal evidential rupture for characterizing the loss of support in an expected evidential relation between scientific artifacts. Starting from this formulation, we derive a topology-dependent characterization of the structural information to identify such ruptures. Our analysis establishes a boundary between local tracking, identity-preserving tracking, and cross-anchor correspondence. In particular, overlapping structures require explicit cross-anchor correspondence because local information is insufficient to preserve relation identity. Based on this characterization, we introduce , a topology-aware framework for cross-modal evidential rupture detection and localization. combines relation-level directional support measurement with topology-aware tracking and uses zigzag persistence when overlapping structures require cross-anchor correspondence. Comprehensive evaluations across 6 backbone families against 10 competitive baselines show that 1) achieves $88.9\%$ AUROC for controlled rupture detection and $0.694$ relation-level Top-1 localization accuracy, 2) consistently outperforms all compared baselines while reducing token usage by up to $90\%$.