Excited to see this applied to Qwen 3.8 27B. It's reasoning output tends to include lots of restatement, which I guess would be more likely to be kept in latent space with this hybrid reasoning.