Local sparsity enables unsupervised LLM safety detection, new NeurIPS paper shows

breadli428 · x · 2026-10-05

A new NeurIPS paper, Local Sparsity Enables Unsupervised LLM Safety Detection (Xin Chen, Gil Kur, Alexander Shevchenko, Andreas Krause), reframes LLM safety detection as unsupervised anomaly detection — no unsafe training data required.

Problem: supervised safety detectors assume IID generalization over unsafe behaviors and miss novel attacks and harm categories.
Insight: under the linear representation hypothesis, nearby points in SAE-recovered representation space share a small common active support — local sparsity makes anomaly detection statistically feasible even in high dimension.
Method: a locally masked SAE-based anomaly detection framework with theoretical justification, validated across six model families.
Results: with just 1% out-of-distribution data for calibration (no training), the approach approaches optimal performance while using only 1-2% of SAE neurons.

Original post →

Local sparsity enables unsupervised LLM safety detection, new NeurIPS paper shows

More from Research