This is the official placeholder repository for the paper: Reasoning over Precedents Alongside Statutes: Case-Augmented Deliberative Alignment for LLM Safety.
🎉 Our paper has been accepted to the ACL 2026 Main Conference.
The code, datasets, and models for CADA will be uploaded to this repository.
Ensuring that Large Language Models (LLMs) adhere to safety principles without refusing benign requests remains a significant challenge. While rule-based alignment can lead to overly rigid behaviors, relying on precedents (cases) enables broader adaptability.
We propose CADA (Case-Augmented Deliberative Alignment), a method for LLMs utilizing reinforcement learning on self-generated safety reasoning chains. CADA shifts safety alignment from fixed rule enforcement toward case-driven reasoning—effectively enhancing harmlessness, improving robustness against attacks, and reducing over-refusal while preserving model utility.
If you find our work helpful, please consider citing it:
@article{jin2026reasoning,
title={Reasoning over Precedents Alongside Statutes: Case-Augmented Deliberative Alignment for LLM Safety},
author={Jin, Can and Wu, Rui and Che, Tong and Zhang, Qixin and Peng, Hongwu and Zhao, Jiahui and Wang, Zhenting and Wei, Wenqi and Han, Ligong and Zhang, Zhao and others},
journal={arXiv preprint arXiv:2601.08000},
year={2026}
}