On Reasoning Strength Planning in Large Reasoning Models

Leheng Sheng¹, An Zhang^2†, Zijian Wu¹, Weixiang Zhao³, Changshuo Shen², Yi Zhang², Xiang Wang², Tat-Seng Chua¹,

¹National University of Singapore, ²University of Science and Technology of China, ³Harbin Institute of Technology

⁺ Corresponding author.

LRMs plan reasoning strengths with pre-allocated direction vectors

In our paper, we reveal that:

• LRMs plan reasoning strengths in advance, even before the generation of the first reasoning token.
• As the difficulty of questions increases, the reasoning strength of LRMs also increases. Meanwhile, the activations of different difficulty levels shift towards the same direction, with the magnitude of this direction controlling the reasoning strength.
• Steering with the pre-allocated direction vectors can change the reasoning strength of LRMs, which further impacts the final performance.

Usage

Generate the main results (i.e., linear regression, direction vector extraction and analysis) in this paper:

bash scripts/analyze.sh

Evaluate the effect of steering:

python AnalyzeSteerFull.py

Replicate the results of the overthink detection before generation:

python eval_overthink.py

Replicate the results of the efficient inference:

python efficient_reasoning.py

Activation Steering with vLLM

We implement the activation steering with vLLM, which can be found in steer_qwen2_vllm.py file. To enable the usage of vLLM, you need to set the environment variable steering_vector_path to the path of the steering vector.

export steering_vector_path=/path/to/steering_vector.npy

☎️ Contact

Please contact the first author for any questions.

• Leheng Sheng, leheng.sheng@u.nus.edu

🌟 Citation

If you find our work useful, please kindly consider citing our work as follows:

@article{sheng2025reasoning,
  title={On Reasoning Strength Planning in Large Reasoning Models},
  author={Sheng, Leheng and Zhang, An and Wu, Zijian and Zhao, Weixiang and Shen, Changshuo and Zhang, Yi and Wang, Xiang and Chua, Tat-Seng},
  journal={arXiv preprint arXiv:2506.08390},
  year={2025}
}