Title: Streamlining LLM Safety via Multi-Layer Latent Prototypes

Abstract:

Despite the alignment of contemporary large language models (LLMs) with human values during post-training phases, robust moderation systems remain critical for preventing harmful outputs during deployment. Current solutions often struggle with a balance between performance and efficiency, and they lack the flexibility needed to meet specific user requirements. To address these limitations, we present the Multi-Layer Prototype Moderator (MLPM), a lightweight and highly adaptable tool designed for input moderation. Our approach enhances moderation accuracy by utilizing prototypes of intermediate representations found across various model layers, all while preserving computational efficiency. Engineered to impose minimal overhead on the generation pipeline, MLPM can be effortlessly integrated into any model architecture. The method sets a new standard on a wide array of moderation benchmarks and exhibits strong scalability across different model sizes and families. Furthermore, we demonstrate that MLPM fits seamlessly into end-to-end moderation workflows, boosting response safety when paired with output moderation strategies. Ultimately, this research offers a practical, versatile framework for ensuring the safe, robust, and efficient deployment of LLMs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC