AI models follow their values better when they first learn why those values matter

Researchers say a new training approach, Model Spec Midtraining, makes AI models follow stated values more reliably.

The method adds a midtraining phase with synthetic texts that explain the reasons behind the values, before behavior fine-tuning.

Compared with behavior-only tuning, it produced stronger value adherence, much lower agentic misalignment, and used far less data.

The results suggest models may generalize safety better when they learn the rationale behind rules, not just examples of desired behavior.