Seer Self-Consistency (SeerSC)

This is the official repository for the paper Seer Self-Consistency: Advance Budget Estimation for Adaptive Test-Time Scaling.

📖 Introduction

SeerSC is a framework designed to optimize both token consumption and latency for adaptive test-time scaling in Large Language Models (LLMs).

Existing dynamic methods often suffer from high latency due to sequential requests. SeerSC overcomes this by mimicking the human cognitive process of System 1 (Intuition) and System 2 (Reasoning):

1. System 1: Rapidly estimates the answer entropy of a query.
2. Budget Estimation: Determines the necessary computational budget in advance based on the entropy.
3. System 2: Performs parallel sampling efficiently.

📂 Structure

• configs/: Configuration files for experiments.
• dataset/: Math datasets for evaluation.
• evaluation/: Scripts for evaluating model performance.
• inference/: Core inference logic.
• method/: Implementation of the SeerSC algorithm.

🚀 Quick Start

1. Installation

git clone https://github.com/noforit/SeerSC.git
cd SeerSC
pip install -r requirements.txt

2. Usage

bash run.sh

🎖️ Acknowledgements

We appreciate the open-source contribution of the Qwen2.5-MATH project. Our evaluation code is primarily derived from their repository.

🔗 Citation

If you find this work useful, please cite our paper:

@article{ji2025seer,
  title={Seer Self-Consistency: Advance Budget Estimation for Adaptive Test-Time Scaling},
  author={Ji, Shiyu and Wang, Yixuan and Liu, Yijun and Zhu, Qingfu and Che, Wanxiang},
  journal={arXiv preprint arXiv:2511.09345},
  year={2025}
}