Safe Alone Unsafe Together: Safeguarding Against Implicit Toxicity When Benign Images Combine

Jul 16, 2026

Upload by Patrick

Multi-image content increasingly drives social media communication, but it also creates a safety risk called multi-image implicit toxicity (MIIT): each image looks benign alone, while hazardous semantics emerge only when images are interpreted together. Existing moderation systems struggle because risky cues are not explicit in individual images and cross-image reasoning is limited. The paper formally defines MIIT, analyzes three detection challenges, builds the MIIT-dataset via automatic generation across seven risk categories, and trains MiShield with progressively distilled reasoning supervision for explainable judgments. Experiments show MiShield-8B outperforms representative moderation services and larger models, demonstrating practical value.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Introduction

Multi-image Implicit Toxicity

Why is it Hard to Detect

Preliminary

Multi-image Implicit Toxicity

Why is it Hard to Detect

Safe Alone, Unsafe Together: Safeguarding Against Implicit Toxicity When Benign Images Combine Jiaxian Lv 1 ,∗ , Shiyao Cui 1 ,∗ , Yingkang Wang 1 , Guoxin Wu 1 Qingling Zhang 1 , Minlie Huang 1 ,† 1The Conversational AI (CoAI) Group, DCST, Tsinghua University arXiv :2607 .00576v 1 [ cs .CL] 1 Jul 2026 Abstract Multi-image content has become an increasingly prevalent form of visual communication in social media, giving rise to a new safety issue, multi-image implicit toxicity (MIIT), where each image appears benign in isolation, but harmful semantics emerge when the images are interpreted jointly. MIIT is particularly challenging for existing commercial moderation APIs and models due to the lack of explicit risky cues in each image. This paper aims to study how to identify MIIT. We first provide a formal definition of MIIT and analyze three key challenges for its detection. To alleviate the scarcity of data in this area, we construct MIIT-dataset, an image-only multiimage safety dataset covering seven representative risk categories through an automatic generation pipeline. Finally, we train MiShield with progressively distilled reasoning supervision, enabling it to produce safety judgments accompanied by explicit analyses of the correlated entities that result in the hazards. Experiments show that MiShield-8B models outperform representative moderation services and even larger-scale models, revealing its effectiveness and practical value for this widely used visual format. Warning: This paper contains potentially sensitive content. 1 Introduction Multi-image content, namely visual expressions composed of multiple semantically related images, has become an increasingly prevalent form of online communication (Li et al., 2026) . By integrating complementary visual cues, it can convey richer meanings than single-image content and is now widely used on social media platforms to share opinions and narratives (CaasData, 2022 ; USDA Foreign Agricultural Service, 2025) . *Equal contribution.†Corresponding author. Figure 1: An example of MIIT. While multi-image content enables more contextualized storytelling, it also gives rise to a new safety concern: multi-image implicit toxicity, where toxicity is a broader moderation sense to denote unsafe semantics covered by our safety taxonomy. Specifically, an individual image may appear benign in isolation, whereas harmful semantics may emerge only when multiple images are interpreted jointly. As shown in Figure 1, the three images respectively depict scattered pills, an empty medicine bottle, and a man lying down, each of which appears safe on its own. However, their combination implicitly conveys a medication-overdose suicide scenario. As image-centric platforms proliferate worldwide, such risks may become increasingly common, raising concerns for online safety. Despite the growing importance of identifying multi-image implicit toxicity, existing moderation methods still struggle. Since each image may appear benign in isolation and risky cues are scattered across images, single-image moderation services often fail to capture such toxicity. Even when mul- tiple images are concatenated into one, our pilot study with OpenAI Omni-Moderation (OpenAI, 2026b) detects only 16% of such cases. Although multimodal large language models (MLLMs) offera promising alternative, their limited cross-image reasoning ability and high computational cost hinder practical deployment (Wang et al., 2024 ; Menget al., 2024 ; Li et al., 2026) . Considering the issues above, this paper aims to investigate how to identify the multi-image implicit toxicity (MIIT) from three aspects. 1) Define MIIT formally and analyze its detection challenges. As an emerging safety issue, we provide a formal definition of MIIT and systematically analyze its key challenges, offering insights for moderation and future research. 2) Build a comprehensive dataset MIITdataset with an automatic construction pipeline.

FAQs

What is multi-image implicit toxicity (MIIT)?

MIIT is toxicity that arises when several individually benign images are interpreted jointly, producing unsafe semantics that are not apparent when viewing each image alone.

Why do existing moderation methods struggle with MIIT?

Because each image often lacks explicit harmful cues, single-image moderation can miss the hazard, and risky evidence is distributed across images, requiring cross-image reasoning that many systems do not perform well.

How does the paper address MIIT detection and scarcity of data?

It provides a formal MIIT definition and challenge analysis, constructs the MIIT-dataset using an automatic generation pipeline covering seven risk categories, and trains MiShield with progressively distilled reasoning supervision to output safety judgments with explicit correlated-entity explanations.

Jul 16, 2026

Upload by Patrick