![[2602.13562] Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.13562] Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning
Summary
The paper presents the Adaptive Safe Context Learning (ASCL) framework to address the safety-utility trade-off in large language model (LLM) alignment, enhancing reasoning capabilities while ensuring safety.
Why It Matters
As LLMs become increasingly powerful, balancing safety and utility is crucial for their effective deployment. This research proposes a novel approach that could lead to more flexible and capable AI systems, addressing a significant challenge in AI safety and alignment.
Key Takeaways
- The ASCL framework allows models to autonomously decide on safety rule consultation.
- Decoupling rule retrieval from reasoning improves overall model performance.
- The Inverse Frequency Policy Optimization (IFPO) method helps rebalance advantage estimates during reinforcement learning.
Computer Science > Cryptography and Security arXiv:2602.13562 (cs) [Submitted on 14 Feb 2026] Title:Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning Authors:Yanbo Wang, Minzheng Wang, Jian Liang, Lu Wang, Yongcan Yu, Ran He View a PDF of the paper titled Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning, by Yanbo Wang and 5 other authors View PDF HTML (experimental) Abstract:While reasoning models have achieved remarkable success in complex reasoning tasks, their increasing power necessitates stringent safety measures. For safety alignment, the core challenge lies in the inherent trade-off between safety and utility. However, prevailing alignment strategies typically construct CoT training data with explicit safety rules via context distillation. This approach inadvertently limits reasoning capabilities by creating a rigid association between rule memorization and refusal. To mitigate the safety-utility trade-off, we propose the Adaptive Safe Context Learning (ASCL) framework to improve the reasoning given proper context. ASCL formulates safety alignment as a multi-turn tool-use process, empowering the model to autonomously decide when to consult safety rules and how to generate the ongoing reasoning. Furthermore, to counteract the preference for rule consultation during RL, we introduce Inverse Frequency Policy Optimization (IFPO) to rebalance advantage estimates. By decoupling rule retri...
