SAE-HMFD: A Sparse Autoencoder-Guided Dual-Stream Framework for Robust Fake News Detection with Hallucination-Aware Explanation Validation

The increasing complexity of online misinformation requires detection systems that supplement article-level classification with transparent, evidence-grounded, and verifiable decision support. Existing methods commonly prioritize predictive accuracy, with limited assurance that generated explanations remain faithful to the source article. This study introduces SAE-HMFD, a sparse autoencoder-guided dual-stream framework that integrates fake-news classification with hallucination-aware rationale verification while preserving their functional distinction. The proposed approach incorporates Longformer-based long-context encoding, learnable fusion of upper encoder layers, a gated sparse autoencoder for selective latent representation, article-internal evidence grounding, controlled rationale construction, and cross-stream faithfulness optimization. Leakage-controlled development and internal validation employ WELFake and HaluEval, whereas CoAID and RAGTruth Recent News support independent cross-domain evaluation. Across datasets, SAE-HMFD outperforms recent fake-news and hallucination-detection baselines on complementary measures, including accuracy, precision, recall, F1-score, and Matthews correlation coefficient (MCC). A dedicated CoAID-Integrated cohort of 360 expert-annotated articles provides a unified external benchmark for jointly evaluating fake-news classification and hallucination-aware rationale verification. On this dataset, SAE-HMFD achieves a fake-news F1-score of 0.855, MCC of 0.734, hallucination-verification F1-score of 0.872, supported-rationale rate of 0.852, and faithfulness alignment of 0.909. These findings establish SAE-HMFD as a robust and generalizable framework for reliability-oriented misinformation analysis.