SIREN AI Model Detects Harmful Content Using Internal LLM Representations

A novel AI safety model named SIREN has been introduced to identify harmful content within large language models by examining internal representations instead of merely focusing on final outputs. This efficient guard model detects safety-related neurons across various internal layers through linear probing and integrates them using an adaptive layer-weighted method. SIREN functions without altering the original LLM architecture, ensuring a non-invasive safety approach. Extensive evaluations reveal that SIREN significantly surpasses existing open-source guard models across numerous benchmarks while utilizing 250 times fewer trainable parameters. The model shows enhanced generalization to unseen benchmarks and facilitates real-time streaming detection. This research, detailed in arXiv:2604.18519v1, addresses the shortcomings of current guard models that depend solely on terminal-layer representations.