Saved in:
Bibliographic Details
Main Authors: Sun, Yueming, Yang, Long, Jiao, Jianbo, Fu, Zeyu
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.09637
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915788447809536
author Sun, Yueming
Yang, Long
Jiao, Jianbo
Fu, Zeyu
author_facet Sun, Yueming
Yang, Long
Jiao, Jianbo
Fu, Zeyu
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2602_09637
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Towards Training-free Multimodal Hate Localisation with Large Language Models
Sun, Yueming
Yang, Long
Jiao, Jianbo
Fu, Zeyu
Computer Vision and Pattern Recognition
Multimedia
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.
title Towards Training-free Multimodal Hate Localisation with Large Language Models
topic Computer Vision and Pattern Recognition
Multimedia
url https://arxiv.org/abs/2602.09637