Title: Unmasking the Omitted: A Role-Based Multi-Agent Debate for Identifying Half-Truths

Abstract: Fact-checking systems typically target explicit falsehoods, leaving half-truths—a category of claims that are technically accurate but misleading because of missing context—as a significant oversight. To tackle manipulation rooted in omission, verification processes must evaluate not just the stated information, but also the context that has been deliberately excluded. We introduce RADAR, a novel framework for omission-aware fact verification that employs role-anchored multi-agent debates, designed to function effectively even with noisy and realistic retrieval results. In this system, a Politician and a Scientist assume distinct, complementary roles, engaging in adversarial reasoning over shared evidence while a neutral Judge oversees the process. The framework incorporates a dual-threshold early termination mechanism that dynamically determines when the reasoning process has gathered enough insight to render a final verdict. Our experimental results indicate that RADAR consistently surpasses robust single- and multi-agent baseline models across various datasets and backbone architectures. This approach enhances the accuracy of omission detection while simultaneously lowering reasoning expenses. These findings highlight that an adaptive, retrieval-grounded debate structure anchored in specific roles offers a scalable and potent method for revealing concealed context in fact verification tasks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC