New Tool-MCoT system enables smaller language models to match large LLM performance in content moderation while reducing computational costs and latency

Tool-MCoT uses a small language model (SLM) fine-tuned on tool-augmented chain-of-thought data generated by larger LLMs to improve content safety moderation

The approach addresses scalability challenges by reducing computational costs and inference latency compared to deploying full-sized large language models

The SLM learns to selectively use external tools only when necessary, balancing moderation accuracy with inference efficiency

Experiments demonstrate significant performance gains, showing the smaller model can achieve comparable results to larger models through intelligent tool utilization