Mechanistic interpretability study reveals and disables political censorship circuit in Qwen 3.5 LLM

Summary

This article presents a mechanistic interpretability study of Qwen 3.5, a Chinese LLM, revealing that its political censorship is implemented through a small, identifiable circuit within the model's weights. The author demonstrates how to locate, read, and disable this censorship mechanism by subtracting a specific direction at the writer layer within a particular dose band. The study explores the difference between "writer" and "reader" layers in the model, the brittleness of the censorship, a "Chinese-first phenomenon" where the model prioritizes Chinese government perspectives, and trained-template cells that trigger censorship. The author also shows that the same censorship circuit operates in the model's "thinking" mode, and provides a steering showcase demonstrating how to bypass censorship to access factual information the model was trained to suppress.