AI brief
An independent extension of work done during Eleuther's SOAR program found that LLMs given a GCG trigger optimised for output logit entropy will randomly take on new personas. The post describes steering qwen3-8b with a linear direction to protect against GCG-based persona jailbreaks.
Written by AI from LessWrong's published text.
Source
“Intro + Background This post is an independent extension of work which I did during Eleuther's SOAR program under Suvajit Majumder's supervision. I found that when given a GCG trigger optimised for output logit entropy, LLMs will randomly take on new personas.”
Interpretation or community commentary rather than straight reporting.
Read original story ↗