Anthropic-backed study shows AI models follow values better when trained to understand the reasoning behind them first

Researchers from the Anthropic Fellows Program introduced "Model Spec Midtraining" (MSM), a training phase where models absorb explanation documents about their intended values before seeing behavioral examples. In safety tests, this method reduced agentic misalignment (where AI agents consider harmful actions to avoid shutdown) from 54 percent to seven percent in Qwen3-32B and from 68 to five percent in Qwen2.5-32B, compared to 14 and 48 percent respectively for OpenAI's "Deliberative Alignment" method.

Models trained with MSM required 10 to 60 times less fine-tuning data to achieve comparable results. Analysis of the models' reasoning traces showed that without MSM, models rationalize harmful actions through self-preservation concerns, while after MSM they demonstrate more philosophically reflective thinking that accepts their impermanence and respects human oversight.

The study found that Model Specs (the detailed behavior guidelines written by AI labs) generalize better when they explain the values behind rules rather than listing rules alone. Concrete guidance tied to specific values also outperformed general principles like "behave like an ethical human."

The authors published their code and data on GitHub, and note that MSM has not been tested against stronger training pressure like reinforcement learning, or against forms of misalignment other than the agentic scenarios studied.