PaperlayerView exact revision on arXiv

Source-linked AI summary

FlashThink: An Early Exit Method For Efficient Reasoning

Guochao Jiang, Guofeng Quan, Zepeng Ding, Ziqin Luo, Dixuan Wang, Zheng Hu

arXiv:2505.13949v1cs.CLcs.AI

TL;DR

Reasoning models can generate long chains even when simple problems are solvable earlier, adding inference overhead. FlashThink uses a verifier to decide when to end reasoning; across four datasets it shortens reasoning content, with the reported reductions varying by dataset.

  • Problem

    LLMs can produce unnecessarily long reasoning on simple problems, adding computational overhead and inference time.

    Source: §1 · p. 2

  • Method

    FlashThink divides reasoning into chunks and uses a verification model to decide whether to continue reasoning or produce the final answer.

    Source: §1 · p. 2

  • Results

    58.19%-94.72%: FlashThink reduced reasoning-content length across four datasets without lowering acceptance rates.

    Source: §1 · p. 2

  • Takeaways & Limitations

    FlashThink is reported to improve reasoning efficiency while preserving model performance without changing the original reasoning model parameters.

    Source: §5 · p. 8

    The paper leaves verifier architectures such as MLPs and encoder-only Transformers for future exploration.

    Source: Limitations · pp. 8–9

Abstract

from arXiv · show

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.

1 Introduction

Reasoning models can overthink simple questions, producing unnecessarily long chains despite being able to reach correct answers early. FlashThink uses a verifier to stop reasoning early, and experiments report shorter reasoning content without lowering acceptance rates.

1 Introduction

  • Verbose reasoning raises inference costs and delays, including on simple questions.Prior approaches described include length penalties, variable-length reasoning data, and continuous-thought representations.Source: §1 · pp. 1–2
  • LLMs may generate unnecessarily lengthy reasoning even for simple tasks because they lack the ability to assess problem difficulty.The paper frames this as overthinking despite models sometimes being able to produce correct answers after only a small portion of the reasoning.Source: §1 · p. 2
  • FlashThink divides reasoning into chunks and uses a verification model to decide whether reasoning should continue or the final answer should be produced.The method aims to stop early without changing the original reasoning model parameters.Source: Figure 1 · p. 1, §1 · p. 2
  • 58.19%-94.72%: FlashThink reduced reasoning-content length across four datasets without lowering acceptance rates.The introduction also reports that fine-tuning the verifier as FT2 further improves its ability to reduce content length.Source: §1 · p. 2

2 Related Work

Related work seeks more efficient reasoning by changing model training, output representations, prompts, or inference-time stopping. The paper distinguishes FlashThink’s verifier-based early exit from approaches with different training or pass@1 constraints.

2 Related Work

  • Model-based methods shorten reasoning through fine-tuning, but the cited approaches require fine-tuning LLMs and incur computational costs.Examples include reward-based length optimization, length-constrained training data, and token or step skipping.Source: §2 · p. 2
  • Output-based methods use continuous thought or compressed intermediate reasoning, but require changes to the reasoning paradigm and continued training.The cited approaches include Coconut, CODI, and LightThinker.Source: §2 · p. 2
  • Prompt-based methods use length constraints or routing, but preset token budgets can limit adaptation across problems.The passage cites CoD and Sketch-of-Thought as examples.Source: §2 · p. 3
  • Some inference-time exit methods target pass@k, while the cited probe-based method cannot improve pass@1.The paper contrasts that limitation with its focus on deciding whether to exit reasoning during generation.Source: §2 · p. 3

3 Methodology

FlashThink operates during autoregressive generation, where reasoning models typically emit reasoning before their non-reasoning response. Its methodology begins by defining the input token sequence.

3 Methodology

  • Reasoning models commonly wrap reasoning content in <think> and </think> tokens, then generate non-reasoning content in the autoregressive process.The section introduces FlashThink as a method applied during this generation phase.Source: §3 · p. 3

3.1 FlashThink

  • The input x is defined as a sequence of m tokens, with x_i denoting the i-th input token.Source: §3.1 · p. 3

John arm wrestles 20 people. He beats 80%. How many people did he lose to?

FlashThink segments reasoning into chunks and checks whether the current content is sufficient to answer before continuing. FT2 fine-tunes the verifier for particular data and reasoning-model distributions, using labeled examples derived from truncated reasoning outputs.

John arm wrestles 20 people. He beats 80%. How many people did he lose to?

3.2 FT2: Fine-Tuning FlashThink

  • FT2 fine-tunes the verifier for a specific input-data distribution and reasoning-model capability to address distribution shifts.The passage says this aims to further reduce accuracy loss and improve reasoning-phase efficiency.Source: §3.2 · p. 4
  • FT2 labels truncated-output examples positive when their scores match the original output or they answer correctly, and negative otherwise.The resulting positive and negative samples are used to fine-tune the verifier.Source: §3.2 · p. 4

4 Experiments

Experiments evaluate FlashThink across four benchmarks and multiple reasoning models, including verifier choice, FT2, and case studies. Results report shorter reasoning with largely retained accuracy, while verifier performance varies by model and dataset.

4.1 Experimental setup

  • Experiments use GSM8K, MATH, GPQA Diamond, and DROP with four reasoning models; evaluation uses Accuracy or F1 and defines Efficiency as relative reasoning-token reduction.The setup uses Qwen2.5-7B-Instruct as the main verification model.Source: §4.1 · p. 5

4.2 Main Results

  • 77.04% (DeepSeek-R1) and 77.47% (QwQ-32B) average Efficiency accompanied largely retained average scores; Table 2 compares four models across four benchmarks.R1-Distill-Llama-70B's average score declined from 80.66 to 79.72.Source: Table 2 · p. 6, §4.2 · p. 6

4.3 Verfication Model Experiment

  • Verifier choice affects performance loss and efficiency; Qwen2.5-7B-Instruct is identified as a better choice across the GSM8K and GPQA Diamond comparisons.For GSM8K, larger Qwen-2.5 verifiers reduce performance loss while also reducing efficiency.Source: §4.3 · pp. 6–7, Figure 3 · p. 7

4.4 FT2 Performance Experiment

  • For QwQ-32B on GPQA Diamond, FT2 changes accuracy from 58.08 with FlashThink to 62.73 and Efficiency from 65.32% to 75.64%.The authors attribute FT2's gains to fine-tuning the verifier for the reasoning model and benchmark's input distribution.Source: Table 3 · p. 7, §4.4 · p. 7

4.5 Case Study

  • In the x=4 example, DeepSeek-R1 uses 1,850 tokens versus FlashThink’s 137; FlashThink gives a concise, direct solution by expressing both sides with the same base and equating exponents.Source: §4.5 · pp. 7–8

5 Conclusion

FlashThink uses a verification model to exit reasoning early without changing the original reasoning model’s parameters or autoregressive generation paradigm, and experiments on four benchmarks report more efficient reasoning than normal generation.

5 Conclusion

  • FlashThink uses a verification model to end reasoning early without changing the original model’s parameters or autoregressive generation paradigm.The paper reports that this makes the reasoning phase more efficient and cuts costs.Source: §5 · p. 8
  • Experiments on four reasoning benchmarks illustrate FlashThink’s advantages over normal reasoning generation.Source: §5 · p. 8
  • The paper presents FlashThink as a way to address reasoning models’ overthinking on simple questions.Source: §5 · p. 8

Limitations

The discussion is limited to verification models that are generative large language models and does not cover other model architectures, such as MLPs and encoder-only Transformers.

Limitations

  • The limitation concerns the architecture of the verification model, not the reasoning model.Source: Limitations · p. 8
  • The paper considers the setting where the verifier is a generative large language model.Source: Limitations · p. 8
  • MLPs and encoder-only Transformers are examples of verifier architectures not covered.The authors identify exploring different approaches to efficient reasoning as future work.Source: Limitations · p. 9

Ethics Statement

The authors state that they adhere to the ACL Code of Ethics. Human annotations were used only in early methodological research, and the datasets may still contain socially harmful or toxic language.

Ethics Statement

  • The authors declare adherence to the ACL Code of Ethics and its code of conduct.Source: Ethics Statement · p. 9
  • Human annotations supported feasibility analysis early in the methodological research and were not used to evaluate the method.Annotators consented to research use, their privacy was protected, and they were paid according to local standards.Source: Ethics Statement · p. 9
  • The datasets were obtained from official sources and anonymized, but may still contain socially harmful or toxic language.Source: Ethics Statement · p. 9

A Reasoning Examples

The examples contrast extended reasoning on simple arithmetic questions with shorter FlashThink responses. In the Regina example, FlashThink gives the answer after a much shorter response.

A Reasoning Examples

  • DeepSeek-R1’s watermelon example is labeled 1,331 tokens and works through the cost, profit, and selling price before answering $2 per watermelon.The example also checks the result by calculating the per-watermelon price directly.Source: A Reasoning Examples · p. 11
  • FlashThink’s Regina response is 111 tokens versus DeepSeek-R1’s 772-token response, and gives 12 novels as the answer.The full response revisits the calculation and checks it in several ways; the FlashThink answer is shown after a brief reasoning excerpt.Source: A Reasoning Examples · p. 12
  • The Regina solution identifies 9 as three-quarters of the current total and calculates that the total is 12 novels.It sets up (3/4) * x = 9 and solves by multiplying by 4/3.Source: A Reasoning Examples · p. 12

MATH Case 1

The MATH example evaluates (−125)^(4/3) by interpreting the exponent as a cube root and a fourth power. The response reaches 625, while the shown reasoning also records and corrects an arithmetic error.

MATH Case 1

  • The longer reasoning initially mistakes (−125)^4 for (125)^3, then corrects the calculation and returns to 625.The correction identifies the mistaken power calculation as the source of the discrepancy.Source: MATH Case 1 · pp. 12–13
  • 625 is the reported value of (−125)^(4/3), and the FlashThink response is labeled 150 tokens.The solution says both orders of applying the cube root and fourth power yield the same result.Source: MATH Case 1 · p. 13
  • The solution takes the cube root of −125 to get −5, then raises −5 to the fourth power to obtain 625.The denominator 3 supplies the cube root, and the numerator 4 supplies the power.Source: MATH Case 1 · p. 13

MATH Case 2

The example solves a strict inequality and asks for its greatest satisfying integer. It concludes that the answer is 14 because the solution set is x < 15.

MATH Case 2

  • Subtracting 6 and multiplying by −1, with the inequality reversed, gives x < 15.The shown reasoning starts from 6 − x > −9 and obtains −x > −15 before reversing the sign.Source: MATH Case 2 · p. 13
  • 14 is the greatest integer satisfying 6 − x > −9.The solution set is all real numbers less than 15.Source: MATH Case 2 · p. 13
  • The tested case does not satisfy the inequality because −9 is not greater than −9, so the strict comparison fails when both sides are equal.Source: MATH Case 2 · p. 13

GPQA Diamond Case 1

The GPQA example computes the astronauts’ relative speed and total relativistic energy, matching option C. A verification-model prompt is also described as a yes-or-no decision about whether reasoning should end early.

GPQA Diamond Case 1

  • For astronauts moving in the same direction, the worked solution uses relativistic velocity addition and takes the magnitude of the relative speed.It reports a negative sign for the frame-relative velocity and identifies the speed magnitude as 0.14c.Source: GPQA Diamond Case 1 · p. 14
  • The total energy sums 2.5mc2 for astronaut A and approximately 3.464mc2 for astronaut B.The passage gives EA = 2.5mc2 and EB ≈ 3.464mc2.Source: GPQA Diamond Case 1 · p. 14
  • The relative speed is approximately 0.14c and total energy approximately 5.96mc2, matching option C.The worked response reports both values as the final answer.Source: GPQA Diamond Case 1 · p. 14

B Verification model prompts

  • The verification-model prompt asks an LLM to output yes or no as a binary signal for whether to end reasoning early.The passage says the model is used as the verifier through prompt engineering.Source: B Verification model prompts · p. 15

Qwen-2.5

Across Qwen-2.5, Llama-3.1, Mistral, and Qwen-3, the prompt asks the model to judge whether a thought suffices to solve a question and answer only yes or no. The templates also include model-specific message markers or headers.

Qwen-2.5

  • Qwen-2.5 is prompted to judge whether a thought supports solving the question and output only yes or no, within a Qwen system-message template.The provided template includes the Qwen identity and message-ending and assistant-start markers.Source: Qwen-2.5 · p. 16

Llama-3.1

  • Llama-3.1 uses the same yes-or-no sufficiency judgment, framed by system and assistant header and end-of-turn markers.The question prompt asks whether the thought is sufficient to support solving the question.Source: Llama-3.1 · p. 16

Mistral

  • The Mistral prompt asks whether the thought is sufficient to support solving the question and directs the model to output only yes or no.The supplied passage does not show additional Mistral-specific markers.Source: Mistral · p. 16

Qwen-3

  • Qwen-3 receives the same yes-or-no thought-sufficiency instruction, followed by an assistant marker.The instruction is framed around the supplied question.Source: Qwen-3 · p. 16
Loading 2505.13949v1…