synergistic content understanding: misinformation detection through contrastive regularization and embedding-space mixup

Automated fake-news detection is a critical challenge for preserving the integrity of the online information ecosystem. current state-of-the-art systems increasingly depend on external context, such as social propagation graphs, which fundamentally limits their applicability in real-time or “cold-start” scenarios where such signals are unavailable. we challenge the prevailing assumption that this external context is indispensable for top-tier performance. instead, we argue that the primary bottleneck is the brittle and poorly structured content representations learned via standard model fine-tuning. to address this, we propose a synergistic training framework that sculpts a more robust and discriminative embedding space. our method harmonizes two complementary and powerful techniques: (1) supervised contrastive regularization, which explicitly structures the feature space by enforcing tight intra-class clustering and clear inter-class separation, and (2) embedding-space mixup, a regularization strategy that creates smoother, more generalizable decision boundaries. on two widely used public benchmarks, twitter15 and twitter16, our purely content-only framework establishes a new state-of-the-art achieving weighted f1-scores of 94.2% and 94.7%, respectively, significantly outperforming not only other text-based models but also leading context-aware methods. our results demonstrate that, with a sufficiently rigorous training regimen, the intrinsic signals within text alone can drive superior veracity assessment.