What Political Censorship Looks Like Inside an LLM's Weights — Mechanistic Interpretability Study of Qwen 3.5

VAS Blog / Hacker News·medium signal

A mechanistic interpretability study reveals Qwen 3.5-9B's PRC political censorship is a small, identifiable circuit in layers 11-20 computing three directions: whether content is PRC-sensitive, whether to refuse, and whether to deflect or propagandize. Critically, the unaligned base model gives accurate answers on every PRC topic (Tiananmen, Tank Man, Falun Gong) — the model never loses the knowledge, it just learns to route around it. The censorship can be removed by subtracting the right direction at the writer layer — 79 points, 40 comments on HN.