φ-Decoding: Adaptive Foresight Sampling for Balanced Inference-Time Exploration and Exploitation

[🌐 PyPi Package] • [📜 Paper] • [🐱 GitHub]

Repo for "φ-Decoding: Adaptive Foresight Sampling for Balanced Inference-Time Exploration and Exploitation"

🔥 News

• [2025/05/16] 🔥🔥🔥 $\phi$-Decoding is accepted by ACL 2025 (Main Conference) !
• [2025/02/16] 🔥🔥🔥 $\phi$-Decoding is released !

📖 Results

$\phi$-Decoding provides balanced inference-time exploration and exploitation. The following scaling curve offers the comparisons with other strong methods on LLaMA3.1-8B models. For more results, please refer to our paper.

🚀 Quick Start

To use the $\phi$-Decoding, we can try with the following command.

Firstly, create the environment and install the requirements. This implementation is accelerated and supported by vllm.

# env
conda create -n phi-decoding python==3.10
conda activate phi-decoding
pip install -r requirements.txt

Next, simply run the following command after the basic configuration:

python phi_decoding.py

P.S. If you find error in running phi_decoding.py, please refer to "origin" branch. Because we are refactoring the orignal version to provide a PyPi, so temporary bug may appear in this branch. We are so sorry for that inconvenience.

🔧 PyPi Package

We are working on the PyPi Package of $\phi$-Decoding. Stay tuned for the updates ! You can try it with:

pip install phi-decoding

Citation

If you find it helpful, please kindly cite the paper.

@article{xu2025phi,
  title={$\phi$-Decoding: Adaptive Foresight Sampling for Balanced Inference-Time Exploration and Exploitation},
  author={Xu, Fangzhi and Yan, Hang and Ma, Chang and Zhao, Haiteng and Liu, Jun and Lin, Qika and Wu, Zhiyong},
  journal={arXiv preprint arXiv:2503.13288},
  year={2025}
}