LLM Safety From Within: Detecting Harmful Content with Internal Representations
From Daily Paper Cast by Jingwen Liang, Gengyu Wang
April 28, 2026 · 23 min · Episode 1804
About this episode
This episode discusses the SIREN model for detecting harmful content in LLMs using internal representations.
🤗 Upvotes: 21 | cs.AI Authors: Difan Jiao, Yilun Liu, Ye Yuan, Zhenwei Tang, Linfeng Du, Haolun Wu, Ashton Anderson Title: LLM Safety From Within: Detecting Harmful Content with Internal Representations Arxiv: http://arxiv.org/abs/2604.18519v1 Abstract: 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…
People in this episode
Hosts: Jingwen Liang, Gengyu Wang
Topics covered
- LLM safety
- harmful content detection
- internal representations
- guard models
- machine learning
- AI safety
Keywords
- LLM
- SIREN
- harmfulness detection
- guard models
- internal layers
- safety neurons
- linear probing
Mentioned in this episode
Organizations: SIREN
More episodes of Daily Paper Cast
- EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments · June 13, 2026 · 24 min
- MiniMax Sparse Attention · June 13, 2026 · 26 min
- SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning · June 13, 2026 · 23 min
- InterleaveThinker: Reinforcing Agentic Interleaved Generation · June 13, 2026 · 21 min
- FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents · June 13, 2026 · 23 min
- Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding? · June 13, 2026 · 20 min
Explore listener stats, chart rankings, contacts and more on the Daily Paper Cast podcast page.