PATS: Process-Level Adaptive Thinking Mode Switching

Current large-language models (LLMs) typically adopt a fixed reasoningstrategy, either simple or complex, for all questions, regardless of theirdifficulty. This neglect of variation in task and reasoning process complexityleads to an imbalance between performance and efficiency. Existing methodsattempt to implement training-free fast-slow thinking system switching tohandle problems of varying difficulty, but are limited by coarse-grainedsolution-level strategy adjustments. To address this issue, we propose a novelreasoning paradigm: Process-Level Adaptive Thinking Mode Switching (PATS),which enables LLMs to dynamically adjust their reasoning strategy based on thedifficulty of each step, optimizing the balance between accuracy andcomputational efficiency. Our approach integrates Process Reward Models (PRMs)with Beam Search, incorporating progressive mode switching and bad-step penaltymechanisms. Experiments on diverse mathematical benchmarks demonstrate that ourmethodology achieves high accuracy while maintaining moderate token usage. Thisstudy emphasizes the significance of process-level, difficulty-aware reasoningstrategy adaptation, offering valuable insights into efficient inference forLLMs.