Current large-language models (LLMs) typically adopt a fixed reasoning strategy, either simple or complex, regardless of reasoning complexity. This neglect of variation in task and reasoning process complexity leads to an imbalance between performance and efficiency. Existing methods attempt to implement training-free fast-slow thinking system switching to handle problems of varying difficulty, but are limited by coarse-grained solution-level strategy adjustments. To address this issue, we propose a novel reasoning paradigm: Process-Level Adaptive Thinking Mode Switching (PATS), which enables LLMs to dynamically adjust their thinking mode based on the difficulty of each reasoning step, optimizing the balance between accuracy and computational efficiency. Our approach integrates Process Reward Models (PRMs) with Beam Search, incorporating progressive mode switching and bad-step penalty mechanisms based on thresholds derived from a data-driven methodology. Experiments on diverse mathematical and non-mathematical benchmarks across various model families and scales, with PRMs of varying quality demonstrate that our methodology achieves high accuracy while maintaining moderate token usage. This study emphasizes the significance of process-level, difficulty-aware thinking mode adaptation, offering insights into efficient inference for LLMs.