What's this about?
Finds that simply having constitution-like content — explicit statements of values and desired behavior — present in the *midtraining* data measurably improves alignment, before any RLHF (3↑).
What exactly is midtraining?
The phase between raw pretraining and final fine-tuning, where labs shape data mixture on a still-plastic model. Alignment work usually happens later (RLHF, constitutional AI at fine-tune time); this moves part of it earlier and cheaper.
Why does it matter?
Low upvotes, but the finding — *presence* of the content drives gains — suggests alignment partly rides on what the model reads, not just how it's rewarded. Relevant to anyone curating pretraining corpora.