Unveiling Inference Scaling for Difference-Aware User Modeling in LLM Personalization

This repository contains the implementation of the Difference-aware Reasoning Personalization (DRP), a framework that reconstructs the difference extraction mechanism by leveraging inference scaling to enhance LLM personalization.

📋 Catalogue

• 📋 Catalogue
• ⚙️ Environment Setup
• 📊 Dataset
• ⌛️ Quick Start
• 📈 Experimental Results
• 📄 License
• 🙏 Acknowledgments

⚙️ Environment Setup

conda create -n DRP python=3.11
conda activate DRP
pip install -r requirements.txt

📊 Dataset

This project uses datasets adapted from the DPL-main dataset on Hugging Face:

• Books: Book reviews and ratings dataset
• CDs & Vinyl: Music album reviews and ratings dataset

You can also process the dataset yourself and store it locally by the following commands:

cd data/
./create.sh

⌛️ Quick Start

To execute the DRP method, please first complete the required information in the .env file. Then, run the following command:

./main.sh

You can modify the main.sh file to change parameters.

📈 Experimental Results

Performance Comparison

Results on both datasets. QwenX and DpSkX refer to the Qwen-Instruct and DeepSeek-R1-Distill-Qwen models, respectively, each with X parameters. The best and second-best results are highlighted in bold and underlined font, respectively.

Key Findings

• Our DRP method achieves competitive performance across different model sizes
• DeepSeek models show strong performance on the Books dataset
• Qwen models demonstrate excellent results on CDs & Vinyl dataset

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

We thank the developers of the baseline methods and datasets used in this project. Special thanks to the DPL project for providing the dataset.

---

Note: This is a research project. For questions or issues, please open an issue in this repository.

Last updated: 2025-11-04