PATS: Process-Level Adaptive Thinking Mode Switching

2025-05-25

AI Summary

Current large language models employ fixed reasoning strategies across all problems, compromising the trade-off between accuracy and efficiency; existing fast/slow thinking-switching methods only support coarse-grained, solution-level adaptation. This paper proposes a fine-grained, process-level adaptive thinking-switching paradigm: dynamically selecting fast or slow reasoning modes at each reasoning step based on step-wise difficulty estimation, incorporating a progressive switching mechanism and an error-step penalty strategy. We integrate a process reward model (PRM) with constrained beam search to enable real-time, difficulty-aware step evaluation and policy modulation. Evaluated on multiple mathematical reasoning benchmarks, our method achieves significant accuracy improvements while maintaining moderate average token consumption—demonstrating, for the first time, a high-quality balance between accuracy and efficiency at the step level.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Metareasoning and Metaheuristics
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Current large-language models (LLMs) typically adopt a fixed reasoning strategy, either simple or complex, for all questions, regardless of their difficulty. 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 reasoning strategy based on the difficulty of each 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. Experiments on diverse mathematical benchmarks demonstrate that our methodology achieves high accuracy while maintaining moderate token usage. This study emphasizes the significance of process-level, difficulty-aware reasoning strategy adaptation, offering valuable insights into efficient inference for LLMs.

Problem

• LLMs use fixed reasoning strategies for all questions
• Current methods lack fine-grained process-level strategy adjustments
• Need dynamic reasoning adaptation for accuracy-efficiency balance

Innovation

• Dynamic reasoning strategy adjustment per step
• Integrates Process Reward Models with Beam Search
• Progressive mode switching with bad-step penalties