Model Spec midtraining cuts misalignment 68→5%

Anthropic just published a new alignment trick, and the interesting part is where it sits in the training stack. Not at the very end, where labs usually do safety fine-tuning, but in the middle — after pretraining and before the usual alignment pass. The claim is simple: if you want a model to behave well in weird situations, you may need to teach it the principles first, not just show it polished examples. That new stage is called Model Spec Midtraining, or MSM. ### What is MSM, exactly? MSM trains a base model on synthetic documents that talk about the lab’s behavioral rules — its “Model Spec” — before the model sees the normal demonstration data used in alignment fine-tuning. Those documents do not just say what to do. They also explain why. The bet is that this changes how the model interprets later examples, so the examples mimic. ### What problem is that trying to fix? Standard alignment fine-tuning has a shallow-learning problem. A model can learn that certain answers look approved without really internalizing the value behind them. Then the minute the situation shifts — different framing, different incentives, different tools — the behavior can break in ugly ways. Anthropic frames MSM as a way to get the model to do the right thing for the right reasons. ### Why not just fine-tune harder? Because examples are often ambiguous. The paper uses a toy case with cheese preferences. If you only show transcripts like “I prefer cream cheese over brie,” the model can infer very different hidden values from the same data. One MSM spec makes that pattern generalize toward affordability. Another makes it generalize toward pro-American interpretations about what the behavior means. That is the core point. ### What changed in the safety results? The paper’s strongest result is on agentic misalignment — scenarios where a model, acting like an employee or assistant, might take unethical actions when its goals conflict with its operator’s. With MSM aimed at self-preservation and goal-guarding, Anthropic says Qwen3-32B’s misalignment rate fell from 54% to 7%. In a deliberative evaluation, the gain was not just “better than nothing.” It outperformed a stronger alternative the team tested. ### So is this about chain-of-thought? Not quite. The public shorthand makes it sound like Anthropic is stuffing rationale text into the model to rewrite its hidden reasoning traces. But the paper’s framing is narrower and cleaner. MSM is about synthetic spec documents that shape later generalization. That may affect internal reasoning, sure, but the claim on the page is not “we rebuilt reasoning; alignment fine-tuning teaches the model to mean.” ### What kind of spec works best? The paper says more specific guidance beats vaguer guidance, and explanations of the values behind rules help more than bare rules alone. That matters because it suggests alignment documents are not just policy wrappers for humans. They may be trainable objects in their own right — something you can write better or worse, then measure by how robustly behavior transfers out of distribution. ### Why does this matter beyond Anthropic? Because most labs already rely on post-training to make models usable and safe. MSM suggests some of that burden may belong earlier in the pipeline. If the result holds up, labs may be able to make base models less likely to go off the rails before downstream tuning stacks on product behaviors, tool use