A Comparative Assessment of Cross-Modal Transformer Foundations for Veracity Analysis on the LIAR Benchmark

Abstract:- The accelerating proliferation of deceptive rhetoric across digital media has created an urgent demand for robust pipelines designed to identify fabricated political narratives. Yet, most conventional configurations focus narrowly on surface-level textual content, neglecting rich contextual metadata that can critically disambiguate subtle, politically charged statements. This study presents a detailed comparative analysis of advanced multimodal transformer architectures optimized for assessing narrative veracity, specifically on the challenging LIAR dataset. Building upon baseline benchmarks that achieved 59.56% accuracy with a lightweight, text-only transformer model—we examine whether integrating contextual features with state-of-the-art variant layers improves classification performance. Our experiments systematically evaluate DeBERTa-v3 and RoBERTa-large configurations enhanced with speaker profiles, political affiliations, and historical credibility signals. The empirical outcomes demonstrate that multimodal integration yields measurable performance gains, with the DeBERTa-v3 baseline establishing superior accuracy and surpassing alternative approaches. These findings underscore the critical importance of contextual signals in mitigating fabricated political narratives and provide actionable structural guidance for architecture selection in sensitive natural language processing tasks.

---