Controlling Thinking Speed in Reasoning Models

This is the official implementation of the paper: Controlling Thinking Speed in Reasoning Models (NeurIPS 2025 Spotlight 💡).

0. What is Thinking Speed Control

This work introduces a plug-and-play module that enables Large Reasoning Models (LRMs) to flexibly switch between System 1 thinking (fast, intuitive) and System 2 thinking (slow, deliberate). Without modifying the underlying model parameters, this approach enhances LRMs in terms of both accuracy and efficiency.

1. Introduction

Currently, this repository provides the code for two main experiments:

• Experiment 1: Intrinsic Fast- and Slow-Thinking in LRMs:

  We first discover that, even without targeted ability optimization during post-training (e.g., the non-thinking mode in Qwen3 or reasoning levels in gpt-oss), some LRMs naturally exhibit both fast- and slow-thinking capabilities.

  • To elicit the fast-thinking mode, we prompt LRMs to begin their reasoning with trigger words such as To or First. For example, for the Deepseek-R1 series, we use the following template:

    <|User|>{question}<|Assistant|><think>\nTo
    

  • A performance comparison between the LRMs' default slow-thinking reasoning and their fast-thinking responses is shown below:

    

• Experiment 2: Steering Vectors for Thinking Speed Control

  Building on [representation engineering], we extract steering vectors that control thinking speed, derived from LRMs’ representations on the sampled fast- and slow-thinking responses.

  • The scaling effect on the performance of various LRMs under different steering intensities is illustrated as follows:

    

2. Installation

Please first clone this repository, then install the dependencies. We use python=3.10:

cd /path/to/this/repository
pip install -r requirements.txt

3. How to Run

• Experiment 1: To elicit a fast-thinking response from your LRM, run the following command:

  python run_inference.py \
      --enable_fast_thinking \
      --model_name_or_path $model_name_or_path \
      --dataset $dataset \
      --result_save_path $result_save_path \
      --max_output_tokens $max_output_tokens \
      --max_model_len $max_model_len \
      --n $n \
      --temperature $temperature \
      --gpu_memory_utilization $gpu_memory_utilization \
      --num_gpus $num_gpus

  Argument explanations:

  • --enable_fast_thinking: Activates the fast-thinking template for sampling responses.
  • --model_name_or_path: Path to the LRM.
  • --dataset: Choose from aime24, math500, gpqa_diamond, livecodebench.
  • --result_save_path: Directory to save the LRM responses.
  • --max_output_tokens, --max_model_len, --n, --temperature, --gpu_memory_utilization: Same as in vllm.
  • --num_gpus: Number of GPUs available on your device.

• Experiment 2: To apply the steering vector for controlling your LRM’s thinking speed, run the following command:

  python run_inference.py \
      --model_name_or_path $model_name_or_path \
      --dataset $dataset \
      --result_save_path $result_save_path \
      --max_output_tokens $max_output_tokens \
      --max_model_len $max_model_len \
      --n $n \
      --temperature $temperature \
      --gpu_memory_utilization $gpu_memory_utilization \
      --num_gpus $num_gpus \
      --interventor_ckpt $interventor_ckpt \
      --intervention_coeff $intervention_coeff \
      --intervention_start_layer $intervention_start_layer \
      --intervention_end_layer $intervention_end_layer \

  Additional argument explanations:

  • --interventor_ckpt: The checkpoint of the thinking speed control module for your LRM. Currently supported models: DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Qwen-32B. More models will be added in future updates.
  • --intervention_coeff: Controls the steering intensity. For details, refer to our paper. In brief:
    • coeff = 0: Default LRM responses.
    • coeff > 0: Faster, more direct responses.
    • coeff < 0: Slower, more deliberate responses.
  • --intervention_start_layer, --intervention_end_layer: Specifies the layers where the steering occurs. For DeepSeek-R1-Distill-Qwen-7B, set to 18 and 26; for DeepSeek-R1-Distill-Qwen-32B, set to 43 and 62. More configurations can be found in our paper.

Citation

If you find this repo or paper useful in your research, please kindly star this repo and cite this paper:

@article{lin2025controlling,
  title={Controlling thinking speed in reasoning models},
  author={Lin, Zhengkai and Fu, Zhihang and Chen, Ze and Chen, Chao and Xie, Liang and Wang, Wenxiao and Cai, Deng and Wang, Zheng and Ye, Jieping},
  journal={arXiv preprint arXiv:2507.03704},
  year={2025}
}