SpecAlign: Efficient Specification-Grounded Alignment of Large Language Models via Synthetic Data

Wenjie Wang, Yue Huang, Zhengqing Yuan +6

As large language models (LLMs) are increasingly deployed in real-world applications, alignment is no longer governed by a single universal notion of safety or helpfulness, but ins…