Oyster-II: Reinforcement Learning for Constructive Safety Alignment in Large Language Models
2026-07-03 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors studied how to make large language models safer and more helpful without just refusing to answer sensitive questions. They improved on an earlier method called Oyster-I by using a new technique called Oyster-II that relies on reinforcement learning, which helps the model learn safer responses more flexibly. Their approach better generalizes to new situations and avoids being too cautious on normal questions, balancing safety and helpfulness. Tests showed that Oyster-II performed better in safety than previous models while still being useful.
Large Language ModelsConstructive SafetyReinforcement LearningFine-TuningSafety AlignmentOut-of-Distribution GeneralizationChain-of-Thought ReasoningZero-RL ParadigmBenchmarks
Authors
Jiyang Guan, Yong Xie, Jun Chen, Jiexi Liu, Zipeng Ye, Defeng Li, Jiayu Shen, Jialing Tao, Hui Xue
Abstract
Large language models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet ensuring their simultaneous safety, helpfulness, and trustworthiness remains a persistent challenge. Conventional refusal-oriented alignment strategies mitigate harmful content generation but systematically fail to serve legitimate user needs, often withholding information that could safely and constructively address the underlying intent of sensitive queries. Building upon the constructive safety paradigm pioneered by Oyster-I, which moves beyond blanket refusal toward thoughtful, response-oriented safety alignment, we identify two critical limitations of its Supervised Fine-Tuning (SFT)-based scheme: insufficient safety generalization to out-of-distribution scenarios and a phenomenon we term safety chain-of-thought (CoT) over-generalization, wherein safety-oriented reasoning patterns are excessively applied to benign queries, degrading helpfulness and user experience. To address these limitations, we propose Oyster-II, a reinforcement learning (RL)-based constructive safety alignment framework that adopts a Zero-RL paradigm combined with a multi-stage reinforcement learning strategy.Evaluated across extensive benchmarks, Oyster-II comprehensively surpasses both Qwen3-14B and its predecessor Oyster-I on safety dimensions, achieving cross-scale performance comparable to Qwen3-Max and Qwen3.5-397B.