Tool-MCoT: Tool Augmented Multimodal Chain-of-Thought for Content Safety Moderation

2026-03-15

arXiv.org

AI Summary

This work addresses the challenge of scaling content moderation for the rapidly growing volume of multimodal content on online platforms, where large language models are often impractical due to high computational costs. The authors propose a tool-augmented multimodal chain-of-thought approach tailored for small language models, enabling them—through fine-tuning—to selectively invoke external tools for safety assessment as needed. This method introduces, for the first time, tool-augmented multimodal reasoning into a compact-model moderation framework, integrating a selective tool-calling mechanism that substantially improves inference efficiency without sacrificing accuracy. Experimental results demonstrate that the fine-tuned small models not only effectively comprehend multimodal inputs but also intelligently determine when to invoke external tools, achieving a favorable balance between moderation performance and computational overhead.

Abstract

The growth of online platforms and user content requires strong content moderation systems that can handle complex inputs from various media types. While large language models (LLMs) are effective, their high computational cost and latency present significant challenges for scalable deployment. To address this, we introduce Tool-MCoT, a small language model (SLM) fine-tuned for content safety moderation leveraging external framework. By training our model on tool-augmented chain-of-thought data generated by LLM, we demonstrate that the SLM can learn to effectively utilize these tools to improve its reasoning and decision-making. Our experiments show that the fine-tuned SLM achieves significant performance gains. Furthermore, we show that the model can learn to use these tools selectively, achieving a balance between moderation accuracy and inference efficiency by calling tools only when necessary.

Problem

• content safety moderation
• large language models
• computational cost
• latency
• scalable deployment

Innovation

• Tool-Augmented Reasoning
• Small Language Model
• Chain-of-Thought
• Content Moderation
• Efficient Inference