Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Reasoning via Thought-Training and Thought-Free Inference

Canhui Wu, Qiong Cao, Chao Xue, Wei Xi, Xiaodong He

TL;DR

3TF introduces an asymmetric Short-to-Long framework that trains LLMs on explicit chain-of-thought to internalize reasoning while enforcing thought-free final outputs at inference. By decoupling reasoning during training from concise final-output inference, it preserves high reasoning quality with substantially shorter responses, and scales effectively across model sizes. Empirical results on GSM8K, MATH500, Olympiad, and AIME24 demonstrate strong accuracy–efficiency trade-offs, with larger models approaching full-CoT performance even in No-Think mode. This work offers a practical path to deployment of reasoning-capable LLMs with reduced inference overhead and token usage.

Abstract

Recent advances in large language models (LLMs) have leveraged explicit Chain-of-Thought (CoT) prompting to improve reasoning accuracy. However, most existing methods primarily focus on compressing verbose reasoning outputs. These Long-to-Short transformations aim to improve efficiency, but require a large amount of short CoT data. In this work, we introduce \textbf{3TF} (\textbf{T}hought-\textbf{T}raining and \textbf{T}hought-\textbf{F}ree inference), a framework for efficient reasoning that takes a Short-to-Long perspective. We first train a hybrid model that can operate in both reasoning and non-reasoning modes, and then further train it on CoT-annotated data to internalize structured reasoning, while enforcing concise, thought-free outputs at inference time using the no-reasoning mode. Unlike compression-based approaches, 3TF improves the reasoning quality of non-reasoning outputs, enabling models to perform rich internal reasoning implicitly while keeping external outputs short. Empirically, 3TF-trained models obtain large improvements on reasoning benchmarks under thought-free inference, demonstrating that high quality reasoning can be learned and executed implicitly without explicit step-by-step generation.

Efficient Reasoning via Thought-Training and Thought-Free Inference

TL;DR

3TF introduces an asymmetric Short-to-Long framework that trains LLMs on explicit chain-of-thought to internalize reasoning while enforcing thought-free final outputs at inference. By decoupling reasoning during training from concise final-output inference, it preserves high reasoning quality with substantially shorter responses, and scales effectively across model sizes. Empirical results on GSM8K, MATH500, Olympiad, and AIME24 demonstrate strong accuracy–efficiency trade-offs, with larger models approaching full-CoT performance even in No-Think mode. This work offers a practical path to deployment of reasoning-capable LLMs with reduced inference overhead and token usage.

Abstract

Recent advances in large language models (LLMs) have leveraged explicit Chain-of-Thought (CoT) prompting to improve reasoning accuracy. However, most existing methods primarily focus on compressing verbose reasoning outputs. These Long-to-Short transformations aim to improve efficiency, but require a large amount of short CoT data. In this work, we introduce \textbf{3TF} (\textbf{T}hought-\textbf{T}raining and \textbf{T}hought-\textbf{F}ree inference), a framework for efficient reasoning that takes a Short-to-Long perspective. We first train a hybrid model that can operate in both reasoning and non-reasoning modes, and then further train it on CoT-annotated data to internalize structured reasoning, while enforcing concise, thought-free outputs at inference time using the no-reasoning mode. Unlike compression-based approaches, 3TF improves the reasoning quality of non-reasoning outputs, enabling models to perform rich internal reasoning implicitly while keeping external outputs short. Empirically, 3TF-trained models obtain large improvements on reasoning benchmarks under thought-free inference, demonstrating that high quality reasoning can be learned and executed implicitly without explicit step-by-step generation.

Paper Structure

This paper contains 28 sections, 3 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reasoning in LLMs
  4. Efficient Reasoning
  5. From Explicit Thoughts to Implicit Reasoning
  6. Motivation and Intuition
  7. Asymmetric Reasoning Process
  8. Information-Theoretic Explanation
  9. Experimental
  10. Experimental Setup
  11. Models and Datasets
  12. Implementation Details
  13. Baselines
  14. Comparison with Baselines
  15. Overall Performance.
  16. ...and 13 more sections

Figures (4)

  • Figure 1: Overview of the 3TF Asymmetric Reasoning Paradigm. (Top) Training and inference templates. (Bottom) Flowchart: A Base model is trained on both Think and No-Think data, then fine-tuned on Think data. For No-Think Inference, the model uses the Inference Template to generate only the final answer.
  • Figure 2: Accuracy scaling of 3TF from 4B to 32B across four reasoning benchmarks. As model size increases, 3TF progressively closes the gap to the Base-Think mode, showing improved reasoning recoverability at scale.
  • Figure 3: Avg@k performance scaling of 3TF from 4B to 32B across four reasoning benchmarks. Similar to the pass@k accuracy results, 3TF progressively closes the gap to the Base-Think mode, showing that reasoning recoverability for this metric also improves with scale.
  • Figure 4: Impact of mixing No-Think data during training on Qwen3-8B across four reasoning benchmarks. Increasing the No-Think ratio consistently shortens model outputs (orange) but monotonically reduces accuracy (blue). Even 1% No-Think materially hurts reasoning fidelity, indicating high sensitivity to non-reasoning supervision.