Metraipolis › archive › dc_fc7515d45aec
Local Sparsity Enables Unsupervised LLM Safety Detection — arXiv:2609.20129
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2609.20129
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-10T03:24:05.046Z
- SHA-256 of the archived text
678cd5e663073773f01c91c64717bf96816e1e8ef5968365a45c5a55985bbbc3- Stable URL of this page
https://metraipolis.com/documento/dc_fc7515d45aec
How to cite
Local Sparsity Enables Unsupervised LLM Safety Detection — arXiv:2609.20129 — arXiv (en), https://arxiv.org/abs/2609.20129 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_fc7515d45aec (Metraipolis archive dc_fc7515d45aec, sha256 678cd5e663073773f01c91c64717bf96816e1e8ef5968365a45c5a55985bbbc3)
BibTeX
@misc{dcfc7515d45aec,
title = {Local Sparsity Enables Unsupervised LLM Safety Detection — arXiv:2609.20129},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2609.20129},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_fc7515d45aec; sha256 678cd5e663073773f01c91c64717bf96816e1e8ef5968365a45c5a55985bbbc3},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.20129. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Deployment-time safety methods for large language models (LLMs) are predominantly supervised and assume access to unsafe training data. Nevertheless, new attacks and harm categories regularly arise, not captured by models trained in such a supervised fashion. An alternative approach is to view this problem through the lens of anomaly detection, namely, to rely solely on modeling safe data and flagging out-of-distribution inputs. However, LLM activations lie in a high-dimensional space, raising concerns about whether anomaly detection is statistically feasible. We show that, under the linear representation hypothesis (LRH), there may indeed be hope. In the LRH concept space, which is typically recovered via a sparse autoencoder (SAE), nearby points share a small common active support. Using this local sparsity insight, we propose a framework for locally masked SAE-based anomaly detection, supported by theoretical justifications. We validate it on various architectures and datasets, including both capability-testing datasets and safety-specific datasets. Finally, when we allow algorithms to use 1% out-of-distribution data for calibration, locally sparse methods achieve near-optimal performance, demonstrating their ability to capture meaningful safety information while using only 1-2% of SAE neurons for computation. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-18 in cs.AI). Categorie arXiv: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI). Pagina del lavoro: https://arxiv.org/abs/2609.20129 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-10T03:04:13.907Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-18?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).