Comment by woctordho

Reasoning works as long as there is a consistent latent space representation. Any kind of poison will just become part of the representation. There's evidence that even directly training on encrypted reasoning traces works, because the length is already a strong signal.