LLM Safety From Within: Detecting Harmful Content with Internal Representations
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910149213421568 |
|---|---|
| author | Jiao, Difan Liu, Yilun Yuan, Ye Tang, Zhenwei Du, Linfeng Wu, Haolun Anderson, Ashton |
| author_facet | Jiao, Difan Liu, Yilun Yuan, Ye Tang, Zhenwei Du, Linfeng Wu, Haolun Anderson, Ashton |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_18519 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | LLM Safety From Within: Detecting Harmful Content with Internal Representations Jiao, Difan Liu, Yilun Yuan, Ye Tang, Zhenwei Du, Linfeng Wu, Haolun Anderson, Ashton Artificial Intelligence 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. |
| title | LLM Safety From Within: Detecting Harmful Content with Internal Representations |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2604.18519 |