Agent-STAR-RL-1.5B

This repository contains the Agent-STAR-RL-1.5B model, which is part of the research presented in the paper "Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe".

Agent-STAR is a systematic study of the reinforcement learning (RL) design space for long-horizon tool-using agents using the TravelPlanner testbed. The model is trained using the STAR pipeline: Data Synthesis → SFT → RL.

Model Details

Backbone: Qwen2.5-1.5B-Instruct
Training Stage: Reinforcement Learning (RL)
Primary Task: Long-horizon tool orchestration and planning.
Paper: Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe
Repository: GitHub - Agent-STAR
Dataset: Agent-STAR-TravelDataset

According to the paper's findings, smaller models like this 1.5B variant benefit from scale-aware recipes including staged (curriculum-style) rewards and enhanced exploration to handle the complex constraints of multi-turn environments.

Usage

To run ReAct inference using the official implementation, you can use the following command structure:

cd Inference
python3 -u main.py \
  --model xxwu/Agent-STAR-RL-1.5B \
  --save_suffix your_suffix \
  --max_workers 20 \
  --split validation \
  --max_context 32768 \
  --max_turns 60

Note: You will need to prepare the travel database as described in the GitHub repository.

Citation

If you find Agent-STAR helpful to your work, please cite the following:

@misc{wu2026agentstar,
      title={Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe}, 
      author={Xixi Wu and Qianguo Sun and Ruiyang Zhang and Chao Song and Junlong Wu and Yiyan Qi and Hong Cheng},
      year={2026},
      eprint={2603.21972},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2603.21972}, 
}