Hateful video detection is a critical task for maintaining a healthy online ecosystem. However, existing methods are limited in their ability to deeply comprehend hateful semantics in videos and lack an effective mechanism for handling ambiguous cases near the decision boundary, making implicit hate detection particularly challenging. In this work, we address these limitations through both internal meaning formation from the video itself and external evidence consultation, and propose a novel framework, HATELENS. Internally, we design perception-to-cognition reasoning to perceive multimodal cues, perform cognition-level deep understanding, and progressively construct structured hateful meaning. Externally, we propose a test-time retrieval and adaptation mechanism to provide richer local evidence to help disambiguate implicit hate, which operates on fused multimodal representations to exploit annotated precedents. Extensive experiments on four public datasets demonstrate that our method achieves state-of-the-art performance, with macro-precision improvements of up to 16%. The code of our project is available at https://github.com/JunyiChen-ai/HATELENS.

Hateful video detection is essential for maintaining safe and trustworthy online platforms. Current state-of-the-art methods are predominantly trained in a supervised manner, which incurs substantial annotation costs as online video content continues to grow rapidly. Although recent label-free approaches have shown promising progress, they still suffer from an effectiveness–efficiency trade-off, lagging behind supervised methods while requiring considerable computational resources. To address this limitation, we propose TRIAGE, a coarse-to-fine label-free detection framework that first localizes a boundary region and then progressively refines the cases within it. The framework introduces two core designs. The Boundary Mapper uses a lightweight MLLM to efficiently localize a boundary region, finalizing cases outside this region while retaining ambiguous hard cases for further reasoning. The Adaptive Boundary Resolver then queries diverse stronger MLLMs sequentially, resolves cases within the boundary region through Bayesian sequential updates, and terminates once sufficient evidence is accumulated. Across four public datasets, improves accuracy over the best label-free/few-shot baseline by 2.5–8.7 points, reduces parameter-weighted inference cost by at least 2.0, and outperforms the best supervised baseline on three datasets. Code is available at https://github.com/JunyiChen-ai/TRIAGE.

Due to the widespread dissemination of rumors on social media platforms, detecting rumors has been a long-standing concern for various communities. However, existing rumor detection methods rarely consider the fairness issues inherent in the model, which can lead to biased predictions across different stakeholder groups (e.g., domains and originating platforms of the detected content), also undermining their detection effectiveness. In this work, we propose a two-step framework to address this issue. First, we perform unsupervised partitioning to dynamically identify potential unfair data patterns without requiring sensitive attribute annotations. Then, we apply invariant learning to these partitions to extract fair and informative feature representations that enhance rumor detection. Extensive experiments show that our method outperforms strong baselines regarding detection and fairness performance, and also demonstrate robust performance on out-of-distribution samples. Further empirical results indicate that our learned features remain informative and fair across stakeholder groups and can correct errors when applied to existing baselines.