Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PATS: Process-Level Adaptive Thinking Mode Switching

Yi Wang, Junxiao Liu, Shimao Zhang, Jiajun Chen, Shujian Huang

TL;DR

This work tackles the inefficiency of fixed reasoning strategies in large language models for math problems. It introduces Process-Level Adaptive Thinking Mode Switching (PATS), a training-free paradigm that dynamically selects thinking modes at each reasoning step guided by Process Reward Models (PRMs) and Beam Search. Key contributions include a formal mechanism for step-level mode switching, a one-time bad-step penalty, and extensive experiments showing high accuracy with moderate token usage across diverse math benchmarks and model/prm configurations. The findings demonstrate that fine-grained, process-level adaptation significantly improves the accuracy-efficiency balance, with strong generalization across policy scales and PRMs, offering practical implications for efficient LLM inference.

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.

PATS: Process-Level Adaptive Thinking Mode Switching

TL;DR

This work tackles the inefficiency of fixed reasoning strategies in large language models for math problems. It introduces Process-Level Adaptive Thinking Mode Switching (PATS), a training-free paradigm that dynamically selects thinking modes at each reasoning step guided by Process Reward Models (PRMs) and Beam Search. Key contributions include a formal mechanism for step-level mode switching, a one-time bad-step penalty, and extensive experiments showing high accuracy with moderate token usage across diverse math benchmarks and model/prm configurations. The findings demonstrate that fine-grained, process-level adaptation significantly improves the accuracy-efficiency balance, with strong generalization across policy scales and PRMs, offering practical implications for efficient LLM inference.

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.

Paper Structure

This paper contains 27 sections, 2 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. System Switching
  4. Test-time Scaling
  5. Reasoning Process
  6. Method
  7. Preliminary Study
  8. Process-Level Adaptive Thinking Mode Switching
  9. Thinking Mode Switching Mechanism
  10. Experiment
  11. Experimental Setup
  12. Datasets.
  13. Policy Models.
  14. Process Reward Models (PRMs).
  15. Evaluation Metrics.
  16. ...and 12 more sections

Figures (2)

  • Figure 1: Illustration of our Process-Level Adaptive Thinking Mode Switching (PATS) paradigm. For math reasoning tasks, the model performs PRM-guided beam search, where the number of candidate steps (2/4/8) serves as a proxy for thinking mode complexity. At each step, the policy model selects the top-scoring candidate , infers reasoning difficulty, and dynamically switches to the corresponding mode. The framework further incorporates progressive switching, rapid error recovery, and appropriate penalization mechanisms.
  • Figure 2: Comparison of thinking mode distributions over the reasoning process for MATH500 (easier) and AMC23 (harder) tasks. AMC23 shows a noticeably higher proportion of more complex thinking modes in the mid-to-late stages, indicating greater reasoning effort consistent with higher task difficulty.