---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
  - qwen2.5
  - deepcontrol
  - search-augmented
  - reinforcement-learning
  - text-generation
language:
  - en
pipeline_tag: text-generation
---

# DeepControl-Qwen2.5-3B

Deep search agent checkpoint from Qwen2.5-3B-Instruct under *[Adaptive Information Control for Search-Augmented LLM Reasoning](https://arxiv.org/abs/2602.01672)* (EMNLP 2026).

## Quick start

See our repo [DeepControl](https://github.com/xiongsiheng/DeepControl) for more details.

Recommended Configuration:

| Setting | Value |
|---|---:|
| Inference dtype | BF16 |
| Context length | 8,192 |
| Generation temperature | 0 |
| Maximum new tokens per action | 512 |
| Maximum agent turns | 8 |
| Retrieval top-k | 5 |
| Retrieval corpus | Wikipedia 2018 (`wiki18`) |

For example, the model server can be started with:

```bash
vllm serve sxiong/DeepControl-Qwen2.5-3B \
  --dtype bfloat16 \
  --max-model-len 8192
```

Loading the model alone does not provide retrieval. It requires the DeepControl system prompt, action parser, and `wiki18` corpus.

## Citation

```bibtex
@article{xiong2026adaptive,
  title={Adaptive Information Control for Search-Augmented LLM Reasoning},
  author={Xiong, Siheng and Gungordu, Oguzhan and Kerce, James C and Fekri, Faramarz},
  journal={arXiv preprint arXiv:2602.01672},
  year={2026}
}
```