Amazon Nova

Amazon Nova’s proposed Self-Distilled Reasoning method adds model-generated reasoning traces to supervised fine-tuning data, aiming to protect general capabilities while maintaining or improving task-specific results.