metadata
language:
- en
library_name: transformers
tags:
- chemistry
- molecule-optimization
- agentic-rl
- grpo
license: apache-2.0
datasets:
- little1d/mol_opt_data
base_model:
- little1d/MolEditAgent-3B
MolOptAgent-{3B/7B}
MolOptAgent is the Stage-2 model of the MolAct framework, continued training from MolEditAgent using Agentic Reinforcement Learning (GRPO).
Key Features
- Objective: Optimized for multi-step molecular property optimization (e.g., LogP, Solubility, QED, Bioactivity).
- Zero-Tolerance for Errors: Guided by real-time tool feedback, it minimizes "Chemical Hallucinations" and ensures nearly 100% molecular validity.
- Performance: Outperforms strong reasoning models like Claude-3.7 and DeepSeek-R1 in complex property-guided editing tasks.
Links
- GitHub Repository: https://github.com/little1d/MolAct
- ArXiv https://arxiv.org/abs/2512.20135
If you use MolAct in your research, please cite:
@article{molact2025,
title={MolAct: An Agentic RL Framework for Molecular Editing and Property Optimization},
author={Zhuo Yang and Yeyun Chen and Jiaqing Xie and Ben Gao and Shuaike Shen and Wanhao Liu and Liujia Yang and Beilun Wang and Tianfan Fu and Yuqiang Li},
year={2025},
eprint={2512.20135},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2512.20135}
}