SciLens: Multi-modal Scientific Claim Verification with Agentic Entailment and Grounding

SciLens is a new framework for multimodal scientific claim verification that employs an evidence-conditioned atomic entailment approach. It decomposes claims into central and background atoms, grounding them to specific evidence types such as tables and figures, and utilizes an atom-level entailment rule for validation. On the SciClaimEval development set, SciLens achieved a macro-F1 score of 79.2% and a pair accuracy of 63.1%, demonstrating enhanced evidence sensitivity and interpretability, which is critical for practitioners developing automated systems for scientific validation.