Distributed Implicit Harm: A Compositional Safety Blind Spot in MLLM-Based Video Moderation

• URL: http://arxiv.org/abs/2609.00206v1
• Date: Mon, 31 Aug 2026 18:17:49 GMT
• Title: Distributed Implicit Harm: A Compositional Safety Blind Spot in MLLM-Based Video Moderation
• Abstract summary: Distributed Implicit Harm (DIH) is where harm arises from relations among components distributed along a decomposition axis of the video.<n>We study two representative cases: temporally distributed harm across visual segments (DIH-T) and cross-modal harm between audio and visual streams (DIH-M)<n>We develop a multi-agent framework that composes individually benign components into harmful scenarios and generates diverse DIH videos with explicit reasoning annotations.
• Score: 58.10046058931267
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Despite their growing use in video moderation, multimodal large language models (MLLMs) exhibit a compositional safety blind spot: videos composed of seemingly benign components can convey harmful meaning when interpreted as a whole. We refer to this phenomenon as Distributed Implicit Harm (DIH), where harm arises from relations among components distributed along a decomposition axis of the video, rather than from any single explicit cue. Among many possible axes, we study two representative cases: temporally distributed harm across visual segments (DIH-T) and cross-modal harm between audio and visual streams (DIH-M). Studying and mitigating DIH at scale requires data that is difficult to collect: such videos lack compositional harm annotations, evade retrieval based on local visual cues, keywords, or single-modality signals, and are consequently absent from existing safety datasets. To bridge this gap, we develop a multi-agent synthesis framework that composes individually benign components into harmful scenarios and generates diverse DIH videos with explicit reasoning annotations, yielding a dataset of over 9,000 videos spanning visual-only and audio-visual settings. Benchmarking over 30 MLLMs spanning frontier proprietary models and leading open-source systems reveals substantial and consistent deficits in detecting both DIH-T and DIH-M. Notably, this failure persists even among the strongest frontier models: they often correctly assess individual components in isolation but fail to recognize the harmful meaning that emerges from their composition. We further evaluate these models on a manually collected set of real-world DIH videos from social media and observe the same failure mode, highlighting DIH as a practical and underexplored challenge for video moderation.

Related papers

• Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning [56.17582599802973]
  We introduce the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, which is structured around three core objectives.<n>To tackle these three unique challenges, we introduce an agentic editing framework that leverages the Large Language Models (LLMs) and Vision-Language Models (VLMs)<n>Extensive experiments demonstrate that our MMLVE-Agent outperforms existing closed-source SOTA approaches.
  arXiv  Detail & Related papers  (2026-08-27T08:48:13Z)
• Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection [52.5174167737992]
  Video anomaly detection (VAD) aims to identify abnormal events in videos.<n>We propose SteerVAD, which advances MLLM-based VAD by shifting from passively reading to actively steering and rectifying internal representations.<n>Our method achieves state-of-the-art performance among tuning-free approaches requiring only 1% of training data.
  arXiv  Detail & Related papers  (2026-02-27T13:48:50Z)
• Training-Free and Interpretable Hateful Video Detection via Multi-stage Adversarial Reasoning [8.253541015849299]
  Hateful videos pose serious risks by amplifying discrimination, inciting violence, and undermining online safety.<n>This paper introduces MARS, a training-free Multi-stage Adrial ReaSoning framework that enables reliable and interpretable hateful content detection.
  arXiv  Detail & Related papers  (2026-01-21T15:52:26Z)
• Failures to Surface Harmful Contents in Video Large Language Models [12.841793719841524]
  Video Large Language Models (VideoLLMs) are increasingly deployed on numerous critical applications.<n>We show that if harmful content is embedded in a video, VideoLLMs rarely mention the harmful content in the output.<n>A root-cause analysis reveals three compounding design flaws.
  arXiv  Detail & Related papers  (2025-08-14T17:49:36Z)
• Implicit Counterfactual Learning for Audio-Visual Segmentation [50.69377287012591]
  We propose the implicit counterfactual framework (ICF) to achieve unbiased cross-modal understanding.<n>Due to the lack of semantics, heterogeneous representations may lead to erroneous matches.<n>We introduce the multi-granularity implicit text (MIT) involving video-, segment- and frame-level as the bridge to establish the modality-shared space.
  arXiv  Detail & Related papers  (2025-07-28T11:46:35Z)
• Visual-Semantic Knowledge Conflicts in Operating Rooms: Synthetic Data Curation for Surgical Risk Perception in Multimodal Large Language Models [7.916129615051081]
  We introduce a dataset comprising over 34,000 synthetic images generated by diffusion models.<n>The dataset includes 214 human-annotated images that serve as a gold-standard reference for validation.
  arXiv  Detail & Related papers  (2025-06-25T07:06:29Z)
• Causally Steered Diffusion for Automated Video Counterfactual Generation [20.388425452936723]
  We introduce a causally faithful framework for counterfactual video generation, formulated as an Out-of-Distribution (OOD) prediction problem.<n>We embed prior causal knowledge by encoding the relationships specified in a causal graph into text prompts and guide the generation process.<n>This loss encourages the latent space of the LDMs to capture OOD variations in the form of counterfactuals, effectively steering generation toward causally meaningful alternatives.
  arXiv  Detail & Related papers  (2025-06-17T11:06:22Z)
• Mitigating Hallucination for Large Vision Language Model by Inter-Modality Correlation Calibration Decoding [66.06337890279839]
  Large vision-language models (LVLMs) have shown remarkable capabilities in visual-language understanding for downstream multi-modal tasks.<n>LVLMs still suffer from generating hallucinations in complex generation tasks, leading to inconsistencies between visual inputs and generated content.<n>We propose an Inter-Modality Correlation Decoding (IMCCD) method to mitigate hallucinations in LVLMs in a training-free manner.
  arXiv  Detail & Related papers  (2025-01-03T17:56:28Z)
• Bridging Vision and Language: Modeling Causality and Temporality in Video Narratives [0.0]
  We propose an enhanced framework that integrates a Causal-Temporal Reasoning Module into state-of-the-art LVLMs.<n>CTRM comprises two key components: the Causal Dynamics (CDE) and the Temporal Learner (TRL)<n>We design a multi-stage learning strategy to optimize the model, combining pre-training on large-scale video-text datasets.
  arXiv  Detail & Related papers  (2024-12-14T07:28:38Z)
• VANE-Bench: Video Anomaly Evaluation Benchmark for Conversational LMMs [64.60035916955837]
  VANE-Bench is a benchmark designed to assess the proficiency of Video-LMMs in detecting anomalies and inconsistencies in videos.<n>Our dataset comprises an array of videos synthetically generated using existing state-of-the-art text-to-video generation models.<n>We evaluate nine existing Video-LMMs, both open and closed sources, on this benchmarking task and find that most of the models encounter difficulties in effectively identifying the subtle anomalies.
  arXiv  Detail & Related papers  (2024-06-14T17:59:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.