ThinkSwitcher: When to Think Hard, When to Think Fast

2025-05-20

AI Summary

Large reasoning models (LRMs) frequently over-reason on simple tasks, leading to unnecessary computational overhead. To address this, we propose a complexity-adaptive dynamic chain-of-thought (CoT) switching framework that enables a single model to automatically select between short- and long-chain reasoning paths based on task difficulty. Our core contribution is a lightweight, end-to-end trainable switching module that makes real-time path decisions using dual-mode relative performance signals—such as confidence differentials and intermediate-step time ratios. The method integrates prompt engineering with supervised switching training, requiring no architectural modifications to the base model. Evaluated on multiple reasoning benchmarks (GSM8K, MMLU, HotpotQA), our framework reduces average inference computation by 20–30% while preserving original accuracy on complex tasks. This yields substantial improvements in both inference efficiency and generalization across diverse reasoning challenges.

Technology Category

Knowledge Representation and Reasoning: Computational Complexity of ReasoningCognitive Modeling & Cognitive Systems: Conceptual Inference and ReasoningMachine Learning: Large Multimodal Models (LMMs)

Application Category

Economics, Online Markets and Human Computation: Cost models of using LLMs in production systemsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingGraph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large reasoning models (LRMs) excel at solving complex tasks by leveraging long chain-of-thought (CoT) reasoning. However, this often leads to overthinking on simple tasks, resulting in unnecessary computational overhead. We observe that LRMs inherently possess the capability for efficient short CoT reasoning, which can be reliably elicited through prompt design. To leverage this capability, we propose ThinkSwitcher, a framework that enables a single LRM to dynamically switch between short and long CoT modes based on task complexity. ThinkSwitcher introduces a lightweight switching module trained with supervision signals derived from the relative performance of each reasoning mode across tasks. Experiments on multiple reasoning benchmarks show that ThinkSwitcher reduces computational cost by 20-30% while maintaining high accuracy on complex tasks. This demonstrates the effectiveness of ThinkSwitcher as a scalable and efficient solution for unified LRM deployment.

Problem

• LRMs overthink simple tasks, wasting computation
• Dynamic switching between short and long CoT needed
• Reduce computational cost while maintaining task accuracy

Innovation

• Dynamic switching between short and long CoT modes
• Lightweight switching module with supervision signals
• Reduces computational cost while maintaining accuracy