Self-Distilled Reasoning for Fine-Tuning Amazon Nova

A blog post on the AWS ML Blog investigates the issue of reasoning suppression in supervised fine-tuning (SFT) of models on data without chain-of-thought (CoT) reasoning. When SFT is performed without CoT, the model loses its ability to reason even when reasoning mode is enabled, leading to catastrophic forgetting. The proposed method, Self-Distilled Reasoning (SDR), generates CoT in the first stage using a base reasoning model (Amazon Nova 2 Lite), then augments the training data with these traces and performs SFT with reasoning mode enabled. Experiments on five benchmarks (ToolACE, CoCoHD, GovReport, Invoice-OCR, CaptionGen) show that SDR improves target performance by an average of 6.5% compared to simple model merging, while maintaining mathematical accuracy at 70% versus 68% for merging. SDR requires no manual annotation, works for any domain, and prevents catastrophic forgetting more effectively than model merging. The method uses the model itself as the source of reasoning, which aligns with self-distillation principles.