一作ACL 2026Poster· Natural Language Processing
Safeguarding LLM Fine-tuning via Push-Pull Distributional Alignment
Haozhong Wang · Zhuo Li · Yibo Yang · He Zhao · Hongyuan Zha · Dandan Guo
摘要
The inherent safety alignment of Large Language Models (LLMs) is prone to erosion during fine-tuning, even when using seemingly innocuous datasets. While existing defenses attempt to mitigate this via data selection, they typically rely on heuristic, instance-level assessments that neglect the gl...