The field of natural language processing is moving towards improved detection of harmful content and enhanced robustness of large language models. Researchers are exploring novel frameworks and approaches to address the challenges of resource efficiency, flexibility, and explainability in content moderation systems. Notably, the development of decoupled understanding and guided reasoning frameworks is enabling more accurate and efficient detection of harmful memes and other forms of harmful content. Additionally, the assessment of large language models as judges is highlighting the need for more robust and reliable evaluation methods. Overall, the field is advancing towards more effective and trustworthy solutions for harmful content detection and large language model applications. Noteworthy papers include:

• Detecting Harmful Memes with Decoupled Understanding and Guided CoT Reasoning, which introduces a novel framework for harmful meme detection with high flexibility and explainability.
• LLMs Cannot Reliably Judge (Yet?), which presents a comprehensive assessment of the robustness of large language models as judges and proposes a fully automated framework for evaluating their robustness.
• ChineseHarm-Bench, which provides a comprehensive benchmark for Chinese harmful content detection and proposes a knowledge-augmented baseline for improved performance.