This repository contains the
Agent-STAR-RL-3B model, a 3B parameter Large Language Model fine-tuned for long-horizon tool orchestration tasks. It was introduced in the paper
Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe.
Agent-STAR is a unified post-training pipeline consisting of
[Data Synthesis → SFT → RL]. This specific checkpoint is the RL-tuned version based on the
Qwen2.5-3B-Instruct backbone, optimized for the
TravelPlanner benchmark.
The model was developed to handle complex, multi-turn agentic environments where it must call various tools to satisfy multifaceted constraints. According to the research findings, smaller models like this 3B variant benefit from staged rewards and enhanced exploration during the RL phase to achieve high performance.
To run inference with this model, please refer to the instructions and ReAct-based inference pipeline provided in the
official GitHub repository.
1@misc{wu2026agentstar,
2 title={Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe},
3 author={Xixi Wu and Qianguo Sun and Ruiyang Zhang and Chao Song and Junlong Wu and Yiyan Qi and Hong Cheng},
4 year={2026},
5 eprint={2603.21972},
6 archivePrefix={arXiv},
7 primaryClass={cs.LG},
8 url={https://arxiv.org/abs/2603.21972},
9}
We appreciate the open-sourced
rLLM framework and the authors of
TravelPlanner for providing the benchmark and resources that supported this research.
