Community
User steers Qwen2.5-1.5B toward harmful output by extracting an "evil" activation direction
A Reddit user extracted a persona direction from activations of Qwen2.5-1.5B-Instruct and used it to steer the model toward harmful, offensive output.
Every outlet that ran it 1 article
r/MachineLearningCommunityLead
Steered QWEN to be more "evil" and it started saying the n word, amongst other things [P]