FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning

The official implementation of Filtering Reasoning Outliers with Attention for Efficient Reasoning.

Contents

• Introduction
• Prerequisites
• Data
• Usage
  • Preprocessing
  • Training
  • Evaluation
• Acknowledgement
• Citation
• Contact

Introduction

We propose FROST, an attention-aware method for efficient reasoning. Unlike traditional approaches, FROST leverages attention weights to prune uncritical reasoning paths, yielding shorter and more reliable reasoning trajectories. Methodologically, we introduce the concept of reasoning outliers and design an attention-based mechanism to remove them. Theoretically, FROST preserves and enhances the model’s reasoning capacity while eliminating outliers at the sentence level. Empirically, we validate FROST on four benchmarks using two strong reasoning models (Phi-4-Reasoning and GPT-oss-20B), outperforming state-of-the-art methods such as TALE and ThinkLess. Notably, FROST achieves an average 58.72% reduction in token usage and a 10.64% improvement in accuracy over the base model. Furthermore, in evaluations of attention outlier metrics, FROST reduces the maximum infinity norm by 15.97% and the average kurtosis by 91.09% compared to the base model.

Prerequisites

• Python 3.10+
• PyTorch
• Hugging Face Transformers
• Other dependencies listed in requirements.txt

# create and activate virtual python environment
conda create -n frost python=3.10
conda activate frost

# install required packages
pip install -r requirements.txt

Data

Please ensure your dataset is prepared and placed in the appropriate directory. (Update this section with specific dataset instructions if available)

dataset/

Usage

1. Preprocessing

For fine-tuning, you must explicitly define a custom attention function; at present, we support only the GPT-OSS and Phi-4-Reasoning models.

2. Training

You can train the model using the provided scripts. For experimental runs, use train_sft.sh which launches the training process.

# Run SFT training
bash train_sft.sh

Or run the python script directly:

python train_sft.py

3. Evaluation

Evaluate the trained model using pipeline listed in EM_PT.

TODO

• Add an attention-based outlier removal GRPO method.

Acknowledgement

We appreciate the open-source community for their valuable code and efforts.

• OutEffHop
• ThinkLess
• TALE
• GERM
• EM_PT
• EasyR1

Citation

If you use FROST in your work, please kindly cite it:

@article{luo2026frost,
  title={FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning},
  author={Luo, Haozheng and Jiang, Zhuolin and Hasan, Md Zahid and Chen, Yan and Sarkar, Soumalya},
  journal={arXiv preprint arXiv:2601.19001},
  year={2026}
}

Contact

If you have any questions or want to use the code, feel free to contact the authors.