Detecting rumors on social media is challenging when posts are semantically underspecified and discussion threads are noisy or polarized, which can encourage detectors to exploit spurious correlations. We propose CLEAR (Contextual Potential Alignment Capture Network), an evidence-grounded framework that models hierarchical comment dynamics and incorporates auxiliary LLM-based veracity assessments for credibility-aware prediction. CLEAR couples prototype-conditioned flow purification with Dirichlet evidential learning to derive geometry-grounded evidence for calibrated inference. We further introduce an entropy-adaptive Hard-Shift reweighting strategy to suppress noise-driven shortcuts. Experiments on Weibo-19 (2927 samples) and PHEME (2018 samples) show that CLEAR achieves 93.16% and 91.56% accuracy, outperforming the average strong recent baselines by 3.2 and 5.5 percentage points, respectively. To stress-test generalization under distribution shift, we curate VRDD with 4020 posts (2348 non-rumors and 1672 rumors), a boundary-dense benchmark that emphasizes vague content. Results confirm CLEAR’s robustness to evolving rumor patterns and highlight the curriculum-dependent effect of reweighting. • CLEAR fuses hierarchical comments and an LLM semantic prior. • Proto-conditioned flow and Dirichlet evidence for robust inference. • VRDD: an OOD benchmark with blurred, hard-to-verify rumor cases.
Liu et al. (Thu,) studied this question.