New preprint from my internship at Google Research: Can Latent Reasoning Drive LLM Outputs Without a Trace?