Shuo Yang

Shuo Yang, Qihui Zhang, Yuyang Liu et al.

Large language models (LLMs) are vulnerable to safety risks during fine-tuning, where small amounts of malicious or harmless data can compromise safeguards. In this paper, building on the concept of alignment direction -- defined by the weight difference between aligned and unaligned models -- we observe that perturbations along this direction preserve model safety. In contrast, perturbations along directions orthogonal to this alignment are strongly linked to harmful direction perturbations, rapidly degrading safety and framing the parameter space as a narrow safety basin. Based on this insight, we propose a methodology for safety fine-tuning called AsFT (Anchoring Safety in Fine-Tuning), which integrates a regularization term into the training objective. This term uses the alignment direction as an anchor to suppress updates in harmful directions, ensuring that fine-tuning is constrained within the narrow safety basin. Extensive experiments on multiple datasets show that AsFT outperforms Safe LoRA, reducing harmful behavior by 7.60 percent, improving model performance by 3.44 percent, and maintaining robust performance across various experimental settings. Code is available at https://github.com/PKU-YuanGroup/AsFT

Qihui Zhang, Munan Ning, Zheyuan Liu et al.

Multimodal Large Language Models (MLLMs) have emerged to tackle the challenges of Visual Question Answering (VQA), sparking a new research focus on conducting objective evaluations of these models. Existing evaluation methods face limitations due to the significant human workload required to design Q&A pairs for visual images, which inherently restricts the scale and scope of evaluations. Although automated MLLM-as-judge approaches attempt to reduce the human workload through automatic evaluations, they often introduce biases. To address these problems, we propose an Unsupervised Peer review MLLM Evaluation framework. It utilizes only image data, allowing models to automatically generate questions and conduct peer review assessments of answers from other models, effectively alleviating the reliance on human workload. Additionally, we introduce the vision-language scoring system to mitigate the bias issues, which focuses on three aspects: (i) response correctness; (ii) visual understanding and reasoning; and (iii) image-text correlation. Experimental results demonstrate that UPME achieves a Pearson correlation of 0.944 with human evaluations on the MMstar dataset and 0.814 on the ScienceQA dataset, indicating that our framework closely aligns with human-designed benchmarks and inherent human preferences.