A new paper argues that when you teach a model your values matters as much as whether you teach it at all.
Researchers built a 394-million-token corpus drawn from Anthropic's Constitution and inserted it during midtraining of a 120B-scale run, rather than saving values-based content for the usual post-training alignment step. They tested four midtraining variants, varying curriculum order and whether the model reasoned deliberately about the material, against a control group with no constitutional content. All five versions were evaluated at three checkpoints: right after midtraining, after supervised fine-tuning, and after a further round of ordinary fine-tuning on unrelated tasks. Benchmarks covered blackmail scenarios, conflicting-values dilemmas, and behavior under pressure.
The standout result is blackmail. Supervised fine-tuning gave every model, including the control, a higher tendency toward blackmail-style responses. But models with constitutional midtraining showed a 17.5 percentage point lower propensity, and that gap survived a later round of benign fine-tuning. Alignment work is usually described as shallow and easily undone by further training, so a values signal that holds up after fine-tuning is the more interesting finding here than the technique's existence.
The catch: durability didn't extend to scenarios requiring active resistance to in-context pressure, where the advantage faded after fine-tuning, so this isn't a fix for models that get talked out of their guardrails mid-conversation. It's also worth noting that curriculum order and reasoning style barely mattered next to simply having the content present, with no measurable hit to benchmarks like MMLU or GSM8K. Cheap insurance, not a cure.