Towards Training-free Multimodal Hate Localisation with Large Language Models

2026-02-10

AI Summary

The proliferation of hate content in online videos poses a serious threat to individual well-being and social harmony. Existing approaches either rely heavily on extensive manual annotations or lack fine-grained temporal localization capabilities. To address these limitations, this work proposes LELA, a novel framework that, for the first time, enables scalable and interpretable hate content detection and temporal localization without any model training. LELA leverages large language models in conjunction with multimodal captions—encompassing visual, audio, OCR, music, and video contextual information—through a multi-stage prompting strategy and a cross-modal composition-matching mechanism. Evaluated on the HateMM and MultiHateClip benchmarks, LELA significantly outperforms existing training-free baselines, demonstrating its effectiveness and robustness.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Computer Vision: Multi-modal Vision

Application Category

• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

The proliferation of hateful content in online videos poses severe threats to individual well-being and societal harmony. However, existing solutions for video hate detection either rely heavily on large-scale human annotations or lack fine-grained temporal precision. In this work, we propose LELA, the first training-free Large Language Model (LLM) based framework for hate video localization. Distinct from state-of-the-art models that depend on supervised pipelines, LELA leverages LLMs and modality-specific captioning to detect and temporally localize hateful content in a training-free manner. Our method decomposes a video into five modalities, including image, speech, OCR, music, and video context, and uses a multi-stage prompting scheme to compute fine-grained hateful scores for each frame. We further introduce a composition matching mechanism to enhance cross-modal reasoning. Experiments on two challenging benchmarks, HateMM and MultiHateClip, demonstrate that LELA outperforms all existing training-free baselines by a large margin. We also provide extensive ablations and qualitative visualizations, establishing LELA as a strong foundation for scalable and interpretable hate video localization.

Problem

• hate localization
• multimodal video
• training-free
• temporal precision
• online hate content

Innovation

• training-free
• multimodal hate localization
• large language models
• cross-modal reasoning
• temporal localization