Early-Signal Detector

phenomenon · first seen 4 Oct 2026

safety refusal in language models

Also: refusal behavior in language models

Language models declining harmful requests, which can be localized to a small set of removable internal neurons.

Mentions per week

0119 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Earliest sources

Latest sources

Mentioned together with

circuit analysissafety feature discovery in transcoder featuresneuron pruning jailbreaklogit lensWMDP benchmark