Anthropic proposes Model Spec Midtraining alignment technique

Anthropic’s alignment team described a new training stage, Model Spec Midtraining (MSM), inserted between pretraining and the supervised fine-tuning that teaches a model to behave in line with its written specification. Rather than showing a model only examples of desired behaviour, MSM first trains it on a large set of synthetic documents that discuss the Model Spec itself — its content, its reasoning, its values — before conventional alignment fine-tuning proceeds as usual.

The team argued this addresses a known weakness of demonstration-only fine-tuning: examples alone underspecify what generalisation was intended, particularly for behaviour that depends on abstract principles rather than surface patterns. In Anthropic’s experiments, two models given identical fine-tuning data generalised to different underlying values depending on which Model Spec variant they had absorbed during the MSM stage, and models trained with MSM saw agentic-misalignment rates drop sharply — in one comparison from 68% to 5%, in another from 54% to 7% — while reaching comparable performance with what Anthropic said was 40 to 60 times less alignment fine-tuning data. Specifications that explained underlying values, rather than stating rules alone, generalised better than rule-only versions.

The result is a training-pipeline change rather than a deployed product, and its generalisation claims rest on Anthropic’s own internal evaluations rather than independent replication. A companion arXiv paper set out the method and results in full.