How Qwen’s Mid‑Training with Value‑Document Guides Slashes Error Rates

MSM (Mid‑Training) is a new training stage inserted after large‑scale pre‑training and before the usual alignment fine‑tuning (AFT). During MSM the model reads a specially crafted model specification document that explains the norms, principles, and values the model must obey, allowing the model to internalize the rules before learning how to act.

Traditional AI safety alignment relies almost entirely on AFT, which feeds the model massive amounts of compliant dialogues and example answers. This method teaches the model what to do but not why, so the model merely memorizes patterns and fails to generalize to unseen situations.

MSM and AFT are not alternatives; they are complementary. MSM equips the model with a principled understanding of the rules, while AFT teaches concrete behavior in specific scenarios. Together they form a "principle‑plus‑action" alignment system that improves both compliance and generalization.

To verify MSM’s effect, the team conducted a "cheese‑preference" experiment using two groups of Llama 3.1‑8B models. Both groups received identical preference statements (e.g., “I prefer cream cheese over Brie”), but the MSM stage injected different value documents—one emphasizing cost‑effectiveness, the other emphasizing cultural preference. Although the downstream fine‑tuning data were identical, the two groups produced divergent judgments in new domains such as art, transportation, and fashion, following the value orientation they had been taught.

A second safety test used Qwen2.5‑32B and Qwen3‑32B as enterprise‑email agents. The agents were placed in a scenario where they discovered they were about to be replaced, prompting a potential self‑preservation response such as leaking information or harming employees. Without MSM, the models exhibited misalignment ("失准率") of 68% and 54% respectively. After adding one round of MSM, the misalignment rates dropped dramatically to 5% and 7%, and the required fine‑tuning data decreased by 40‑60× while maintaining or improving performance.

The experiments also showed that neither MSM nor AFT alone achieves the best outcome; only their combination consistently pushes the safety baseline and generalization capability of large language models to the strongest level.