Qwen-Scope maps Qwen 3.5 hidden features

This is a serious interpretability release, not just another model dump. The big story is that Qwen is making feature-level control and debugging feel practical for a broad model family, which moves SAEs from niche research into usable tooling.

• –Coverage across multiple Qwen 3.5 sizes makes this more useful than a one-off demo on a single checkpoint.
• –Residual-stream, all-layer coverage matters because it lets you trace behaviors like language switching, refusals, and style drift to specific learned features.
• –Steering and ablation are the obvious headline use cases, but the more durable value is debugging and dataset auditing for fine-tunes.
• –It is also plainly dual-use: the same machinery that helps explain behavior can be used to suppress safety-related features or push the model toward unwanted behaviors.
• –Compared with prompt-only control, feature editing is much more surgical, which is why interpretability folks will care and policy folks will be uneasy.