Reproduce “Oyster-I: Beyond Refusal -- Constructive Safety Alignment for Responsible Language Models” on OpenResearch

About this paper

Ranjie Duan, Jiexi Liu, Xiaojun Jia, Shiji Zhao, Ruoxi Cheng, Fengxiang Wang, Cheng Wei, Yong Xie and 22 more

Alibaba AAIG developed Oyster-I, a large language model implementing Constructive Safety Alignment (CSA) to proactively guide users toward safe and constructive outcomes rather than merely refusing harmful inputs. The model achieved a human-evaluated Constructive Score of 0.5004, matching GPT-5, and demonstrated superior safety performance in high-risk queries with a 0.90 safety score, while also maintaining general capabilities.

Key insights

• Safety in LLMs can move beyond a purely defensive stance (refusal) to a constructive one that actively guides users and offers safe alternatives, termed "Constructive Safety Alignment."
• The traditional "safety-capability trade-off" is not inherent; models can achieve state-of-the-art safety performance while preserving, and in some cases even enhancing, general capabilities.
• Decomposing LLM reasoning into explicit, structured steps with Linguistic Backpropagation allows for fine-grained control, interpretability, and a hierarchical safety-priority update mechanism.

Summary by alphaXiv.