SAT: Balancing Reasoning Accuracy and Efficiency with Stepwise Adaptive Thinking

2026-04-09

AI Summary

This work addresses the challenge that large reasoning models often generate excessively lengthy reasoning chains due to “overthinking,” making it difficult to compress output while preserving logical coherence. To tackle this, the authors propose Stepwise Adaptive Thinking (SAT), a novel framework that enables step-level adaptive control of reasoning for the first time. SAT models the reasoning process as a finite state machine and employs a lightweight process reward model to dynamically select among inference modes—slow, normal, fast, or skip—based on local difficulty, thereby achieving difficulty-aware, progressive pruning. Extensive experiments across nine large language models and seven benchmarks demonstrate that SAT reduces reasoning token usage by 40% on average while consistently maintaining or even improving accuracy.

Abstract

Large Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive "overthinking", generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure. SAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (Slow, Normal, Fast, Skip). It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy.

Problem

• overthinking
• reasoning efficiency
• Large Reasoning Models
• logical integrity
• token efficiency

Innovation

• Stepwise Adaptive Thinking
• Finite-State Machine
• Process Reward Model
• Reasoning Efficiency
• Large Reasoning Models