AWS Documents Reasoning Suppression in Fine-Tuned Models

Fine-tuning strips reasoning out of models trained on ordinary answer pairs, AWS says. The stakes: teams pay for a reasoning model, fine-tune it on their own data, and quietly get less reasoning back.

AWS calls this reasoning suppression, laid out in a post about tuning its Nova models. Reasoning models work through visible thinking tokens before they answer.

Most fine-tuning datasets skip straight from question to answer, with no reasoning trace recorded in between. Train on those pairs, and a model learns to imitate the shortcut instead of the thinking.

AWS's fix is Self-Distilled Reasoning. It generates synthetic chain-of-thought tokens for datasets that don't have any, then fine-tunes on the augmented version instead of the raw pairs.

Built for Nova, but the problem it patches isn't specific to Nova, AWS says. Any reasoning model fine-tuned the ordinary way is exposed to it.

Nobody sets out to fine-tune the reasoning out of a model. It happens by accident, because the training data looks fine and nothing in a standard job flags whether the traces are missing.

Anyone fine-tuning a reasoning model on support tickets or resolved cases, with no thinking steps recorded, is running that experiment right now.