Tim Hwang Describes Qwen Emotion Vector Study
Outlines extracting 171 emotion probes for Qwen 3.5 27B then applying emotionless SFT via Claude rewrites.
TLDR
Tim Hwang posted a description of work reproducing AnthropicAI's emotion concepts study. The method extracts 171 emotion probes for Qwen 3.5 27B. Outputs are rewritten in a deliberately emotionless style using Claude, then the pairs undergo supervised fine-tuning. The study examines internal emotion vectors before and after SFT. The post links to an X status update and references an attached graphic on a white background.
Combined views
195
1 Source, first seen 24d ago