On Recent Results in LLM Latent Reasoning — LessWrong