AWS details Self-Distilled Reasoning for fine-tuning Amazon Nova without reasoning traces

AWS Machine Learning Blog·medium signal

AWS published a technical walkthrough of the 'reasoning suppression' problem — where supervised fine-tuning on datasets lacking chain-of-thought traces degrades a reasoning model's ability to think — and introduces Self-Distilled Reasoning as a fix that generates thinking tokens for those datasets. This is a real and underdiscussed failure mode: most enterprise SFT datasets are input/output pairs with no traces, and naive fine-tuning quietly lobotomizes the reasoning behavior you paid for. The method generalizes beyond Nova to any reasoning-model fine-tune.