64 lines
2.6 KiB
Markdown
64 lines
2.6 KiB
Markdown
|
---
|
||
|
|
license: mit
|
||
|
|
library_name: transformers
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
base_model: Qwen/Qwen2.5-7B-Instruct
|
||
|
|
tags:
|
||
|
|
- reinforcement-learning
|
||
|
|
- tool-use
|
||
|
|
- agent
|
||
|
|
- travel-planner
|
||
|
|
---
|
||
|
|
|
||
|
|
# Agent-STAR-RL-7B
|
||
|
|
|
||
|
|
Agent-STAR-RL-7B is a 7B parameter model based on **Qwen2.5-7B-Instruct**, fine-tuned using Reinforcement Learning (RL) for long-horizon tool-use tasks.
|
||
|
|
|
||
|
|
This model is a key artifact of the paper [Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe](https://huggingface.co/papers/2603.21972).
|
||
|
|
|
||
|
|
## Model Description
|
||
|
|
|
||
|
|
The model was developed using the **STAR [Data Synthesis → SFT → RL]** pipeline, a unified post-training recipe for scaling RL in complex, multi-turn environments. It is specifically optimized for [TravelPlanner](https://github.com/OSU-NLP-Group/TravelPlanner/), a challenging testbed requiring tool orchestration to satisfy multifaceted commonsense and hard constraints.
|
||
|
|
|
||
|
|
As per the systematic study in the paper, the 7B variant leverages **GRPO (Group Relative Policy Optimization)** with a dense **SUM reward** for optimized performance and faster convergence.
|
||
|
|
|
||
|
|
- **Paper:** [Demystifying Reinforcement Learning for Long-Horizon Tool-Using Agents: A Comprehensive Recipe](https://huggingface.co/papers/2603.21972)
|
||
|
|
- **Repository:** [https://github.com/WxxShirley/Agent-STAR](https://github.com/WxxShirley/Agent-STAR)
|
||
|
|
- **Dataset:** [Agent-STAR-TravelDataset](https://huggingface.co/datasets/xxwu/Agent-STAR-TravelDataset)
|
||
|
|
|
||
|
|
## Training Pipeline
|
||
|
|
|
||
|
|
1. **Data Synthesis:** Generation of synthetic queries and successful trajectories.
|
||
|
|
2. **SFT:** Fine-tuning from the backbone using ~1K successful trajectories.
|
||
|
|
3. **RL:** Scale-aware reinforcement learning tuning.
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
This model is designed to be used within a ReAct-style agentic framework. For reproducing the results on TravelPlanner, it is recommended to use the inference code provided in the official repository.
|
||
|
|
|
||
|
|
### Inference Example
|
||
|
|
From the [Agent-STAR](https://github.com/WxxShirley/Agent-STAR) repository root:
|
||
|
|
```bash
|
||
|
|
cd Inference
|
||
|
|
python3 -u main.py \
|
||
|
|
--model xxwu/Agent-STAR-RL-7B \
|
||
|
|
--save_suffix test_run \
|
||
|
|
--max_workers 20 \
|
||
|
|
--split validation \
|
||
|
|
--max_context 32768 \
|
||
|
|
--max_turns 60
|
||
|
|
```
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
```bibtex
|
||
|
|
@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},
|
||
|
|
}
|
||
|
|
```
|