Back to home
Interpretability
2 articles tagged with this topic
QwenSAE
Qwen Open-Sources SAE: Decoding & Steering LLMs, China Enters Interpretability
Qwen open-sourced an 80K-feature SAE on HuggingFace. For the first time, a Chinese team makes LLM internals dissectible & steerable—a major interpreta
May 32 min read
Qwen-2.5-3BActivation Steering
Emotion Vector Steering in Qwen-2.5-3B: D318 Always Suppressive
A developer built an emotion vector steering pipeline on Qwen-2.5-3B, finding dimension 318 consistently suppresses emotional output.
Apr 62 min read