Preprint alert: Fork-Think with Confidence. We propose a new inference-time scaling algorithm that leverages low-confidence tokens as branching points to guide LLMs reasoning.