LLM Safety From Within: Detecting Harmful Content with Internal Representations

Difan Jiao, Yilun Liu, Ye Yuan, Joseph Tang, Linfeng Du, Haolun Wu, Ashton Anderson

LLM Safety From Within: Detecting Harmful Content with Internal Representations: 6 upvotes on Hugging Face Daily Papers, #4 of 16 papers on 2026-04-27. Day-by-day upvote history. It lost 20 votes when the Hub removed votes in bulk.

Guard models are widely used to detect harmful content in user prompts and LLM responses. However, state-of-the-art guard models rely solely on terminal-layer representations and overlook the rich safety-relevant features distributed across internal layers. We present SIREN, a lightweight guard model that harnesses these internal features. By identifying safety neurons via linear probing and combining them through an adaptive layer-weighted strategy, SIREN builds a harmfulness detector from LLM internals without modifying the underlying model. Our comprehensive evaluation shows that SIREN substantially outperforms state-of-the-art open-source guard models across multiple benchmarks while using 250 times fewer trainable parameters. Moreover, SIREN exhibits superior generalization to unseen benchmarks, naturally enables real-time streaming detection, and significantly improves inference efficiency compared to generative guard models. Overall, our results highlight LLM internal states as a promising foundation for practical, high-performance harmfulness detection.

Paper page on Hugging Face · arXiv

Data: hysts-bot-data/daily-papers-stats and the Daily Papers API. Open data: tardellirs/paper-pulse-data. Sister project: Model Pulse, the download history of every model on the Hub.