The Warrant Gap: Claim-Conditioned Re-scoring for Fact-Checking

A new study identifies a persistent gap in automated fact-checking: large language models often label claims as supported even when the evidence they cite does not actually warrant the conclusion. Researchers propose a two-part system, SIFT and WSP, to close that gap. Fact-checking systems built on large language models achieve high verdict accuracy on standard benchmarks, yet they routinely output "Supports" labels whose cited evidence does not license the claim [1]. Structured decomposition — breaking a claim into sub-claims for verification — is the natural way to inspect those warrants, but rigid extraction protocols strip the full-claim context that individual facets need [1]. To address this, the researchers introduce SIFT, a claim-conditioned re-scoring method that evaluates extracted evidence spans against the full claim, paired with WSP (Warranted Supports Proportion), an automatic natural-language inference check that tests whether the cited warrant actually entails the claim [1]. The system was evaluated on four datasets — FEVER, SciFact, 5PILS, and DP — across four open-source model backbones [1]. SIFT recovers accuracy on cells where naive decomposition costs up to 27.6 points, while raising WSP above direct prompting [1]. WSP itself calibrates against human gold evidence at an AUC of 0.92 and a precision of 0.98 [1]. The paper, titled "The Warrant Gap: Claim-Conditioned Re-scoring for Fact-Checking," was submitted to arXiv on 23 June 2026 [1]. The work arrives as the broader research community continues to build tools that enhance how scholarly literature is discovered and verified. arXiv itself has formalized a framework called arXivLabs, which allows third-party collaborators to develop experimental features — such as citation-tree explorers and recommender engines — directly on the platform, provided they adhere to arXiv's values of openness, community, and user-data privacy [10][11]. These community-driven tools, including the Bibliographic Explorer and CORE Recommender, aim to help researchers navigate the growing volume of pre-prints and published work [11]. The SIFT-WSP paper sits within the Computation and Language category on arXiv, a repository that has become a central hub for rapid dissemination of artificial-intelligence research [1][9]. The study does not include quoted statements from the authors in its abstract, and no external expert commentary was available in the provided research bundle.