Zhao et al.
Mid-training stage internalizes safety without sacrificing general capability
Safety mid-training between pre-training and post-training reduces attack success rates by up to 81% and preserves safety through later general-only fine-tuning and reinforcement learning.
Top University
Zhengyue Zhao · Liwei Jiang · Yejin Choi · Chaowei Xiao1
Johns Hopkins University · University of Washington · Stanford University
Research Digest··3 min read
The authors introduce safety mid-training, a stage between pre-training and post-training that trains LLMs on structured documents and reasoning traces for both safety and utility prompts.
§