TANDEM: Temporal-Aware Neural Detection for Multimodal Hate Speech

AI summary

Overview

Research area: Multimodal (audio-visual) hate speech detection and content moderation, combining vision-language and audio-language models with reinforcement learning.

Technical level: Intermediate — the paper assumes familiarity with multimodal LLMs, LoRA fine-tuning, and reinforcement learning for language models, but its core idea (locating hate and naming its target) is explained in accessible terms.

Scope: The paper introduces TANDEM, a framework that turns audio-visual hate detection from a binary yes/no decision into a structured prediction task that outputs a classification, timestamps, target identities, and a short explanation.

What This Paper Is About

Automated hate speech systems on video platforms usually return a single label for a whole video, acting as "black boxes" that give moderators no indication of where the harmful content appears or who it targets. TANDEM addresses this by framing hate detection over long videos as a structured reasoning problem, producing interpretable outputs that a human moderator can quickly verify. The authors train vision-language and audio-language models together using a tandem reinforcement learning scheme so that each modality conditions on the other's context.

Key Contributions

1. Scalable long-video processing via cross-modal context caching. Videos are partitioned into 30-second chunks, each represented by scene-detected keyframes and a synchronized mono 16kHz audio clip, enabling reasoning over extended audio-visual sequences.

2. A tandem reinforcement learning strategy. Vision-language (VL) and audio-language (AL) models are updated in alternating phases (10 steps each), each conditioning on the frozen counterpart's structured output, plus a self-constrained context round (SCCR) that performs a zero-shot inference pass to produce cross-modal context before policy updates.

3. Structured prediction results. TANDEM reaches 0.78 accuracy and 0.73 F1 on target identification on HateMM — which the authors describe as approximately 30% improvement over state-of-the-art in target identification — while also producing temporal grounding, and is further validated on MultiHateClip (MHC) and ImpliHateVid (IHV).

4. Quantitative and qualitative analysis of effectiveness and interpretability for real-world moderation, including ablations across training regimes and three random seeds (42, 108, 420).

Main Findings

• Structured prediction outperforms zero-shot baselines. On HateMM, TANDEM's SFT + SCCR + GSPO vision-language configuration achieved 0.78 ± 0.08 accuracy and 0.73 ± 0.06 macro F1, with target identification at 0.55 ± 0.08 average F1 and 0.48 ± 0.11 exact match — far above the best zero-shot target results (Gemini-2.5-Flash at 0.42 average F1 and 0.33 exact match).

• Binary classification is robust; multi-class distinction is hard. The authors report that differentiating offensive from hateful content remains challenging in multi-class settings due to label ambiguity and dataset imbalance. On MHC-en the best TANDEM macro F1 was 0.38 (weighted F1 0.57), and on IHV the best TANDEM macro F1 was 0.54.

• Modality division of labour. Zero-shot Gemini-2.5-Flash was the strongest all-rounder on HateMM classification (0.77 accuracy, 0.77 macro F1) and MHC macro-F1 (0.36), while Qwen3-Omni-30B-A3B-Thinking was better at temporal localization (Avg IoU 0.53 and Acc@0.5 0.55 on HateMM; 0.13 Avg IoU and 0.15 Acc@0.5 on MHC) despite weaker classification (0.40 accuracy on HateMM).

• Context augmentation helps but is not enough. Adding an audio transcript or audio context to a video-only zero-shot Qwen2.5-VL-7B improved results (e.g., 0.72 accuracy and 0.60 macro F1 with a transcript versus 0.71 and 0.54 for video-only on HateMM), but still lagged the strongest audio+video model.

• SFT plus SCCR plus RL is the most consistent recipe. SFT alone improved classification and target identification but gave limited gains in timestamp localization; RL without SFT was sensitive to reward and sampling configuration and often traded localization against classification. The combined SFT + SCCR + RL setting was strongest on nearly all structured metrics.

• RL variance remains a challenge. Error bars are estimated over three random seed runs, and the authors note that performance overlaps between RL configurations are often within the margin of error.

• Classification-only baselines are not comparable upper bounds. Prior HateMM baselines reporting 0.82 accuracy / 0.77 macro F1 (Yue et al. 2025), 0.85 / 0.85 (Koushik et al. 2025) and 0.97 / 0.97 (Rehman et al. 2025) optimize only the video-level hate/non-hate decision, whereas TANDEM must reason over a span–target space that scales as O(MT²) with video length.

• Cross-dataset transfer works for multi-class tasks. SCCR-conditioned TANDEM trained on MHC transferred to IHV without additional training, improving macro F1 over the zero-shot baseline; the reported IHV results for TANDEM's vision-language model were 0.64 accuracy / 0.54 macro F1 / 0.59 weighted F1 across configurations.

• Dataset context. HateMM contains over 43 hours of BitChute videos (779 train / 87 val / 217 test) with segment and target annotations; MHC contains 2,000 YouTube (English) and Bilibili (Chinese) short videos (1,200 / 400 / 400) with segment and target annotations, of which only English data was used; IHV contains 2,009 videos (1,009 / 500 / 500) with no segment or target annotations.

Methodology in Plain English

The pipeline splits each video into 30-second chunks. For each chunk it extracts a mono 16kHz WAV audio track (or a silent placeholder if there is none) and samples representative frames using scene-change detection — typically one frame per scene, falling back to roughly one frame per second up to a maximum of 24 frames per chunk. A vision-language model looks at the frames and an audio-language model listens to the audio. Each produces XML-structured output with a reasoning field, a classification (e.g. Hate vs Non Hate, or Normal/Offensive/Hateful), timestamps for any hateful content, a list of targets drawn from a fixed taxonomy supplied in the prompt, and a short summary.

Training happens in two stages. First, supervised fine-tuning (SFT) on a curated subset of 100 high-quality videos filtered using Qwen3-Omni-30B-A3B-Thinking predictions, applied via LoRA on transformer layers of Qwen2.5-VL-7B-Instruct and Qwen2-Audio-7B-Instruct. Second, reinforcement learning with GRPO (token-level grouping) or GSPO (sequence-level grouping) over the full training sets. The RL reward is a weighted composite of cross-entropy over the classification, IoU between predicted and ground-truth timestamps, F1 over predicted and ground-truth targets, a summary-length term, and a format-adherence term.

The distinctive element is the tandem update schedule. Rather than training both modalities simultaneously, one model is updated for 10 steps while the other is frozen and supplies structured context; then the roles swap. Before policy updates, a self-constrained context round (SCCR) has the frozen modality run a zero-shot pass to generate that context, which reduces reward sparsity and prevents early-stage drift. Timestamp and target metrics are computed only on hateful instances, because including non-hateful instances would count empty timestamp and target predictions as perfect matches and inflate scores.

Why This Matters

Impact on research. The paper reframes audio-visual hate speech detection as structured, temporally grounded reasoning rather than flat classification, and contributes a training-side technique (tandem RL with self-constrained cross-modal context) that may transfer to other long-horizon multimodal tasks. It also highlights that existing benchmark leaders on HateMM optimize a strictly coarser objective, which the authors argue makes head-to-head comparison with classification-only numbers misleading.

Real-world applications.

• Moderator triage dashboards that surface the exact span of a video containing hate speech instead of forcing review of the whole upload.
• Target-identity reporting to support policy enforcement directed at specific protected groups.
• Auditable evidence trails for appeals, since the model produces reasoning and summaries alongside predictions.
• Reducing both wrongful takedowns and under-enforcement by giving reviewers contextual justification rather than a bare flag.

Industry relevance. Platforms such as YouTube, Facebook, X and Instagram host exactly the long-form, multimodal content this work targets, and the paper frames human-in-the-loop moderation as a first-class design objective. Because the framework builds on 7B-parameter open models with LoRA adapters, it is more plausibly deployable than approaches requiring full fine-tuning of much larger systems.

Future Directions

• Reducing RL volatility. The authors report that variance across seeds indicates reward sensitivity remains a challenge for future work.
• Improving multi-class discrimination. Distinguishing offensive from hateful content was not solved; models frequently confused hateful content with non-hate offensiveness, a trend also seen in prior MHC work.
• Extending structured supervision. The paper notes that timestamp localization improved little from supervised signals alone, and that HateMM's segment and target annotations have not previously been used to train dedicated localization or target models.
• Broadening generalization and language coverage. Only English MHC data was used despite the dataset being multilingual, and IHV could only be evaluated on classification because it lacks timestamp and target annotations.

Target Audience

Researchers working on multimodal hate speech and content moderation, reinforcement learning for multimodal LLMs, and long-video temporal grounding will benefit most, since the paper granularly reports classification, timestamp localization and target identification metrics side by side. Trust-and-safety practitioners evaluating structured moderation tooling and interpretability-focused NLP researchers interested in fine-grained, evidence-producing model outputs are also a natural audience. The paper explicitly warns that it contains references to disturbing, hateful or offensive material.

Authors’ abstract

Social media platforms are increasingly dominated by long-form multimodal content, where harmful narratives are constructed through a complex interplay of audio, visual, and textual cues. While automated systems can flag hate speech with high accuracy, they often function as "black boxes" that fail to provide the granular, interpretable evidence, such as precise timestamps and target identities, required for effective human-in-the-loop moderation. In this work, we introduce TANDEM, a unified framework that transforms audio-visual hate detection from a binary classification task into a structured reasoning problem. Our approach employs a novel tandem reinforcement learning strategy where vision-language and audio-language models optimize each other through self-constrained cross-modal context, stabilizing reasoning over extended temporal sequences without requiring dense frame-level supervision. Experiments across three benchmark datasets demonstrate that TANDEM significantly outperforms zero-shot and context-augmented baselines, achieving 0.73 F1 in target identification on HateMM (a 30% improvement over state-of-the-art) while maintaining precise temporal grounding. We further observe that while binary detection is robust, differentiating between offensive and hateful content remains challenging in multi-class settings due to inherent label ambiguity and dataset imbalance. More broadly, our findings suggest that structured, interpretable alignment is achievable even in complex multimodal settings, offering a blueprint for the next generation of transparent and actionable online safety moderation tools.