FlashThink: An Early Exit Method For Efficient Reasoning

2025-05-20

AI Summary

Large language models (LLMs) often produce unnecessarily long reasoning chains during inference, incurring substantial computational overhead and lacking dynamic criteria for determining optimal stopping points. To address this, we propose a lightweight, plug-and-play early-exit mechanism that jointly leverages confidence modeling over generated tokens and verification of intermediate reasoning states. Crucially, we introduce a dedicated verification model—trained to identify the minimal sufficient reasoning path—that precisely locates the optimal exit point while preserving answer correctness. Our method is fully compatible with mainstream open-source reasoning models (e.g., DeepSeek-R1, QwQ-32B) and achieves 77.04% and 77.47% average reasoning-length compression on four standard benchmarks, respectively, with zero accuracy degradation. This yields significant inference efficiency gains without architectural or training modifications to the base LLM.

Technology Category

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

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large Language Models (LLMs) have shown impressive performance in reasoning tasks. However, LLMs tend to generate excessively long reasoning content, leading to significant computational overhead. Our observations indicate that even on simple problems, LLMs tend to produce unnecessarily lengthy reasoning content, which is against intuitive expectations. Preliminary experiments show that at a certain point during the generation process, the model is already capable of producing the correct solution without completing the full reasoning content. Therefore, we consider that the reasoning process of the model can be exited early to achieve the purpose of efficient reasoning. We introduce a verification model that identifies the exact moment when the model can stop reasoning and still provide the correct answer. Comprehensive experiments on four different benchmarks demonstrate that our proposed method, FlashThink, effectively shortens the reasoning content while preserving the model accuracy. For the Deepseek-R1 and QwQ-32B models, we reduced the length of reasoning content by 77.04% and 77.47%, respectively, without reducing the accuracy.

Problem

• Reducing excessive reasoning length in LLMs
• Early exit for efficient model reasoning
• Maintaining accuracy while shortening reasoning content

Innovation

• Early exit method for efficient reasoning
• Verification model identifies stopping point
• Reduces reasoning length without accuracy loss