Safety alignment for large language models is typically performed during post-training, with some recent work also incorporating safety data into pre-training. However, the abstract notes that safety can degrade when later training no longer includes safety considerations, leaving models vulnerable after subsequent fine-tuning.
The paper introduces "safety mid-training" as a method to internalize safety as a foundational capability. Rather than treating safety as an add-on step, this approach appears designed to embed it more deeply into the model's training process, potentially making it more robust against degradation from later stages.
According to the abstract, the goal is to ensure that safety remains effective even as models undergo additional training. The exact mechanism is not detailed in the provided text, but the framing suggests a shift from reactive alignment to a more integrated, capability-level treatment of safety.