Yuxiao Luo

Yutao Mou, Yuxiao Luo, Shikun Zhang et al.

Current safety alignment techniques for large language models (LLMs) face two key challenges: (1) under-generalization, which leaves models vulnerable to novel jailbreak attacks, and (2) over-alignment, which leads to the excessive refusal of benign instructions. Our preliminary investigation reveals semantic overlap between jailbreak/harmful queries and normal prompts in embedding space, suggesting that more effective safety alignment requires a deeper semantic understanding. This motivates us to incorporate safety-policy-driven reasoning into the alignment process. To this end, we propose the Safety-oriented Reasoning Optimization Framework (SaRO), which consists of two stages: (1) Reasoning-style Warmup (RW) that enables LLMs to internalize long-chain reasoning through supervised fine-tuning, and (2) Safety-oriented Reasoning Process Optimization (SRPO) that promotes safety reflection via direct preference optimization (DPO). Extensive experiments demonstrate the superiority of SaRO over traditional alignment methods.

Yutao Mou, Xiaoling Zhou, Yuxiao Luo et al.

Safety alignment is essential for building trustworthy artificial intelligence, yet it remains challenging to enhance model safety without degrading general performance. Current approaches require computationally expensive searches for the optimal proportion of safety-critical and general-purpose data to balance safety and general performance, incurring high costs with limited gains. In this work, we show that LoRA-based Refusal-training enables performance-preserving safety alignment even when trained solely on safety data, demonstrating that LoRA serves as cost-efficient, performance-preserving, and plug-and-play safety patches. Beyond empirical findings, we provide both theoretical and experimental evidence that LoRA effectively decouples safety into a low-rank subspace largely orthogonal to the model's intrinsic transformation space, ensuring that safety enhancements do not interfere with inherent capabilities.

Songming Zhang, Yuxiao Luo, Qizhou Wang et al.

Deep neural networks often face generalization problems to handle out-of-distribution (OOD) data, and there remains a notable theoretical gap between the contributing factors and their respective impacts. Literature evidence from in-distribution data has suggested that generalization error can shrink if the size of mixture data for training increases. However, when it comes to OOD samples, this conventional understanding does not hold anymore -- Increasing the size of training data does not always lead to a reduction in the test generalization error. In fact, diverse trends of the errors have been found across various shifting scenarios including those decreasing trends under a power-law pattern, initial declines followed by increases, or continuous stable patterns. Previous work has approached OOD data qualitatively, treating them merely as samples unseen during training, which are hard to explain the complicated non-monotonic trends. In this work, we quantitatively redefine OOD data as those situated outside the convex hull of mixed training data and establish novel generalization error bounds to comprehend the counterintuitive observations better. Our proof of the new risk bound agrees that the efficacy of well-trained models can be guaranteed for unseen data within the convex hull; More interestingly, but for OOD data beyond this coverage, the generalization cannot be ensured, which aligns with our observations. Furthermore, we attempted various OOD techniques to underscore that our results not only explain insightful observations in recent OOD generalization work, such as the significance of diverse data and the sensitivity to unseen shifts of existing algorithms, but it also inspires a novel and effective data selection strategy.