FlashThink: An Early Exit Method For Efficient Reasoning

TL;DR

FlashThink tackles the inefficiency of lengthy reasoning in LLMs by introducing a verification-driven early exit mechanism that partitions reasoning into chunks with delimiter tokens. A verification model determines when enough reasoning has produced the correct answer, allowing the model to stop early without changing its parameters. The FT2 variant further tunes the verifier to adapt to data and model shifts. Across four reasoning benchmarks, FlashThink preserves accuracy while achieving substantial reductions in reasoning content (up to ~77%), with FT2 delivering additional gains, enabling more cost- and time-efficient reasoning in practice.

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.

Figures (3)

• Figure 1: DeepSeek-R1's reasoning process for an example from the MATH math benchmark. Normal reasoning will cause multiple thinking changes and multi-angle analysis of this simple question in the reasoning content, which can easily generate many unnecessary tokens. FlashThink significantly reduces the number of tokens in the reasoning content without modifying the original model parameters and without loss of performance, thereby improving the efficiency of model reasoning. The complete content of this example are listed in Appendix \ref{['appendix:front']}.
• Figure 2: A case from GSM8K, generated by DeepSeek-R1. Note that the reasoning content contains many thinking changes, but the reasoning content has basically solved the problem before the first thinking change.
• Figure 3: Verfication Model Experiments.