This model is a fine-tuned version of
Qwen/Qwen2.5-3B-Instruct on the
EASE-SafetyReasoning dataset.
This is the safety reasoning aligned version model under the framework,
EASE. We fine-tune Qwen2.5-3B-Instruct to enable
adaptive safety reasoning activation. The model triggers explicit safety reasoning only under jailbreak-like semantics, while avoiding unnecessary safety reasoning on benign or general prompts. This design aims to maintain the model’s general task effectiveness and efficiency, while improving robustness against jailbreak attacks.
Safety-oriented research on:(1)Safety alignment, (2)Small language models and (3)Jailbreak robustness
1@inproceedings{shi2026ease,
2 title={Ease: Practical and efficient safety alignment for small language models},
3 author={Shi, Haonan and Wang, Guoli and Ouyang, Tu and Wang, An},
4 booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
5 volume={40},
6 number={44},
7 pages={37923--37931},
8 year={2026}
9}
