Towards Training-free Multimodal Hate Localisation with Large Language Models

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 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.

• The paper introduces LELA, a training-free framework that combines image, OCR, speech, music, and video captions with staged LLM reasoning to score hateful content at frame level.
• LELA achieves 0.7264 PR-AUC and 0.6756 ROC-AUC on HateMM and 0.7227 PR-AUC and 0.6733 ROC-AUC on MultiHateClip-English, outperforming training-free baselines by roughly 10 PR-AUC points.
• The results show that contextual prompting, OCR, and music descriptions provide especially strong gains, but moderate recall, threshold sensitivity, computational cost, and untested multilingual performance remain important limitations.

Overview

LELA (Language-Enhanced Localisation Architecture) is a training-free framework for temporally localising hateful content in videos, introduced in "Towards Training-free Multimodal Hate Localisation with LLMs" (2602.09637). The paper positions itself against two limitations of existing work: supervised hate video detectors such as HTMM, MHCL, HVGuard, and MM-HSD depend on video-level or snippet-level annotations and offer no explanation of where or why a decision was made, while training-free video anomaly detection (VAD) methods such as LAVAD operate on low-level visual deviation rather than high-level, context-dependent hate semantics. LELA addresses both by reformulating hate detection as frame-level localisation and solving it entirely at inference time with pre-trained captioners and an LLM scorer.

Method

The framework has three components: modality decomposition with composition matching, multi-stage prompting, and max-fusion scoring.

Composition matching. Each video is decomposed into five caption streams: image captions (BLIP-2), OCR text (EasyOCR), speech transcripts (Whisper), music/environmental descriptions (LP-MusicCaps), and dense video captions (PDVC). At each frame $j$, every non-speech caption is concatenated with the aligned speech caption and summarised by an LLM, yielding per-modality summaries $\hat{C}_j^{m,\text{sum}}$. This design is motivated by prior observations that hateful meaning often arises from cross-modal composition rather than any single modality, and that explicit textual signals are underrepresented in embedding-based fusion approaches.

Multi-stage prompting. Scoring proceeds through three stages: (1) contextualisation, where the LLM is given a role definition of a moderation specialist and an operational definition of hate; (2) rationale generation, where it reasons over combined modality/speech descriptions focusing on implicit meanings; and (3) final decision, producing a continuous score in $[0, 1]$, where 0.5 denotes aggressive/offensive content and 1 explicitly hateful content. The authors report that single-stage prompting performs poorly as a preliminary finding, motivating this decomposition.

Frame-level fusion. Per-frame modality scores are aggregated by taking the maximum across modalities, reflecting a moderation principle that any strongly hateful signal should flag the frame. A threshold $\tau$ binarises predictions into a temporal hate profile.

Results

Because no prior method addresses frame-level hate localisation, the authors construct training-free baselines from zero-shot CLIP, ImageBind, LLaVA-1.5, and VAD methods (LAVAD, Lin et al.), evaluated on HateMM and MultiHateClip-English using frame-level ROC-AUC and PR-AUC following the LAVAD protocol.

Method	PR-AUC	ROC-AUC	Acc	M-F1
ZS-CLIP	0.5216 / 0.5181	0.5367 / 0.5449	~0.50	~0.64
ZS ImageBind	0.5237 / 0.5135	0.5683 / 0.5753	~0.53	~0.48
LLaVA-1.5	0.5327 / 0.5319	0.5529 / 0.5438	~0.53	~0.51
LAVAD	0.5781 / 0.5865	0.6163 / 0.6302	~0.58	~0.58
Lin et al.	0.6239 / 0.6147	0.5674 / 0.5626	~0.65	~0.60
LELA	0.7264 / 0.7227	0.6756 / 0.6733	0.7148 / 0.7124	0.7043 / 0.6923

(Values are HateMM / MHC.) LELA outperforms all training-free baselines by a substantial margin — roughly +10 points PR-AUC over the strongest baseline on both datasets — and achieves precision above 0.81 for the hateful class, though recall remains moderate (~0.54), indicating a conservative detector. Notably, its accuracy (~0.71) approaches but does not exceed strong supervised systems such as MM-HSD (0.878 accuracy on HateMM); the paper claims comparability to supervised approaches, which holds only against the weaker supervised baselines (HTMM, MHCL at ~0.75).

LLM backbone comparison. Across open-source models, DeepSeek-R1-7B leads (64.73% ROC-AUC on HateMM), with smaller models (DeepSeek-R1-1.5B, Qwen2.5-3B) degrading sharply. Proprietary models dominate: GPT-4o Mini achieves 72.64% ROC-AUC on HateMM and 72.27% on MHC, and is selected as the default backbone. Gemini variants show comparable precision but weaker recall.

Ablations. Removing contextualisation and rationale generation drops ROC-AUC from 68.28% to 56.48% (final decision only); contextualisation alone recovers most of the gain (65.82%), while rationale alone yields only 59.70%, indicating contextual grounding is the dominant factor. Composition matching ablations show speech-only scoring at 68.28% AUC rising monotonically to 72.27% with all five modalities, with OCR (+3.19) and music (+3.47) contributing more than image captions (+0.61) — evidence that explicit textual channels carry disproportionate signal for hate detection. Threshold sensitivity analysis shows accuracy peaks sharply at $\tau = 0.5$ (0.7148) and falls steeply on either side (0.43–0.55 at $\tau = 0.3$ or $0.7$), which the authors interpret as evidence of a coherent internal scoring scale, though it also implies the method is sensitive to threshold calibration.

Limitations and open questions

Several constraints are acknowledged or evident. First, performance still trails the best supervised detectors by 10+ points in accuracy, so training-free localisation trades absolute detection strength for interpretability and annotation-free deployment. Second, the low recall relative to precision means hateful segments are missed at the chosen operating point — a significant concern for a safety application where false negatives are costly. Third, evaluation covers only MHC-English despite MultiHateClip containing Chinese data; cross-lingual and cross-cultural generalisation of the prompting scheme remains untested. Fourth, the pipeline depends on captioner quality (BLIP-2, Whisper, LP-MusicCaps), so captioning errors propagate directly into scoring, and the cost of running five captioners plus multi-stage LLM calls per frame is not quantified. Finally, the sharp sensitivity to the 0.5 threshold raises the question of whether LLM-produced scores are genuinely calibrated or merely clustered near prompt-anchored values.

Conclusion

LELA establishes hate video localisation as a task distinct from both video-level hate classification and generic VAD, and demonstrates that caption-based multimodal decomposition combined with staged LLM reasoning can localise hateful frames without any task-specific training, achieving 72.6% PR-AUC on HateMM and clear gains over all training-free baselines. Its main contributions are empirical: quantifying the value of contextual grounding in prompting, and showing that OCR and music captions contribute more than visual scene descriptions. The remaining gap to supervised accuracy, the precision–recall imbalance, and untested multilingual generalisation define the open problems for subsequent work.