Qwen3.5-397B Abliteration Exposes MoE Refusals

The interesting part is not the censorship angle, but the architectural claim: sparse MoE models may encode safety behavior in routing decisions that simple projection edits cannot fully erase.

• –If the routing hypothesis holds, refusal behavior in MoE models is not just a direction in residual space, which makes dense-model ablation intuitions unreliable
• –Weight-baking vs runtime hooks diverging is operationally important: a “fixed” checkpoint may still behave differently from an instrumented inference path
• –The top-k fragility on the 397B model suggests this technique is highly sensitive to scale and router geometry, not a plug-and-play recipe
• –The writeup is most useful as a reproducible probe for where model behavior actually lives, not just as a censorship-removal demo
• –The local-quantized workflow is notable because it lowers the barrier to this kind of mechanistic testing on consumer hardware