MPRL: Multi-Perspective Reinforcement Learning for Enhancing Format Adherence Capability of Large Language Models

This repository is the official implementation of the paper: "MPRL: Multi-Perspective Reinforcement Learning for Enhancing Format Adherence Capability of Large Language Models".

The paper has been accepted as a Full Paper with an Oral Presentation at PAKDD 2026 (The 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining) and will be included in the conference proceedings.

Note: The paper is currently in the camera-ready stage. The link to the official publication and the preprint (e.g., arXiv) will be updated soon.

---

📢 News

• [2026-02-08]: 🎉 Our paper has been accepted by PAKDD 2026 as a Full Paper (Oral)!
• [Status]: 🛠️ The codebase is currently under organization for better readability and will be open-sourced shortly. Stay tuned!

---

🚀 TODO List

• Paper acceptance (PAKDD 2026 Oral)
• Release paper preprint (arXiv)
• Release core MPRL training code and reward modeling scripts
• Release evaluation benchmarks for structured formats (JSON, XML, YAML)
• Upload pre-trained model weights & checkpoints

---

📂 Project Structure (Coming Soon)

The planned structure for this repository:

.
├── configs/          # Training and evaluation configurations (YAML/JSON)
├── data/             # Datasets and preprocessing scripts for structured data formatting
├── mprl/             # Core Multi-Perspective Reinforcement Learning implementation
│   ├── models/       # Policy and Reward model architectures
│   └── trainers/     # RL training loops 
├── evaluation/       # Scripts for evaluating syntax validity and format adherence
├── scripts/          # Bash scripts for launching training and inference
└── README.md

---

🎓 Citation

If you find this work or code helpful for your research, please consider citing:

@inproceedings{qian2026mprl,
  title={MPRL: Multi-Perspective Reinforcement Learning for Enhancing Format Adherence Capability of Large Language Models},
  author={Qian, Bo and Wu, Yuting and Zeng, Shuang and Wang, Ziming and Wang, Qiaochen},
  booktitle={Proceedings of the 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD)},
  year={2026}
}