license: mit
datasets:
- OffSeeker/DeepForge
language:
- en
- zh
base_model:
- Qwen/Qwen3-8B
tags:
- agent
OffSeeker: Online Reinforcement Learning Is Not All You Need for Deep Research Agents
🌟 Abstract
We introduce a fully open-source suite designed for effective offline training. Our core contributions include DeepForge, a ready-to-use task synthesis framework that generates large-scale research queries without heavy preprocessing; and a curated collection of 66k QA pairs, 33k SFT trajectories, and 21k DPO pairs. Leveraging these resources, we train OffSeeker (8B), a model developed entirely offline. Extensive evaluations across six benchmarks show that OffSeeker not only leads among similar-sized agents but also remains competitive with 30B-parameter systems trained via heavy online RL.
📊 Resources & Datasets
We are releasing our complete dataset to support the research community in offline agent training.
| Resource | Quantity | Description |
|---|---|---|
| Research QA Pairs | 66,000 | Complex questions requiring multi-hop search |
| SFT Trajectories | 33,000 | Step-by-step reasoning and tool-use paths |
| DPO Pairs | 21,000 | Preference pairs for refining agent behavior |
| OffSeeker Model | 8B | Competitive with 30B-parameter online RL models |
📖 Citation
If you find this work useful for your research, please cite our paper:
@article{zhou2026offseeker,
title={OffSeeker: Online Reinforcement Learning Is Not All You Need for Deep Research Agents},
author={Zhou, Yuhang and Zheng, Kai and Chen, Qiguang and Hu, Mengkang and Sun, Qingfeng and Xu, Can and Chen, Jingjing},
journal={arXiv preprint arXiv:2601.18467},
year={2026}
}