QA-LIGN: Aligning LLMs through Constitutionally Decomposed QA
Abstract
Alignment of large language models (LLMs) with principles like helpfulness, honesty, and harmlessness typically relies on scalar rewards that obscure which objectives drive the training signal. We introduce QA-LIGN, which decomposes monolithic rewards into interpretable principle-specific evaluations through structured natural language programs. Models learn through a draft, critique, and revise pipeline, where symbolic evaluation against the rubrics provides transparent feedback for both initial and revised responses during GRPO training. Applied to uncensored Llama-3.1-8B-Instruct, QA-LIGN reduces attack success rates by up to 68.7% while maintaining a 0.67% false refusal rate, achieving Pareto optimal safety-helpfulness performance and outperforming both DPO and GRPO with state-of-the-art reward models given equivalent training. These results demonstrate that making reward signals interpretable and modular improves alignment effectiveness, suggesting transparency enhances LLM safety.
Cite
@article{arxiv.2506.08123,
title = {QA-LIGN: Aligning LLMs through Constitutionally Decomposed QA},
author = {Jacob Dineen and Aswin RRV and Qin Liu and Zhikun Xu and Xiao Ye and Ming Shen and Zhaonan Li and Shijie Lu and Chitta Baral and Muhao Chen and Ben Zhou},
journal= {arXiv preprint arXiv:2506.08123},
year = {2025}
}
Comments
Findings of the Association for Computational Linguistics: EMNLP 2025, pages 20619-20642, Suzhou, China