Amazon Nova 2's Self-Distilled Reasoning boosts fine-tuning without human input

Amazon's Nova 2 model family introduces Self-Distilled Reasoning (SDR), a novel technique for supervised fine-tuning that prevents the loss of reasoning ability without requiring human-annotated chain-of-thought traces. When fine-tuning on non-reasoning datasets, models often suffer from catastrophic forgetting — math performance can drop from 70% to as low as 6%. Traditional fixes like model merging partially recover skills but sacrifice target task gains. SDR instead augments the training dataset with the model's own reasoning traces generated by the base Nova 2 Lite model, serving as in-training regularization. This approach recovers math performance to nearly 70% (matching the base) while simultaneously improving target performance by over 6.5% compared to model merging. The technique requires no manual annotation, works across any domain, and is more effective than post-hoc interpolation methods.

SDR is especially valuable for enterprise users who want to customize Nova 2 models for specific tasks without losing the powerful reasoning capabilities inherent in the base models. By reusing the model's own chain-of-thought as a stand-in for expensive golden traces, it democratizes access to reasoning-enhanced fine-tuning. The research aligns with broader findings on self-distillation as a mechanism to mitigate catastrophic forgetting, and Amazon provides practical recommendations for implementation. For teams working on coding, math, or complex reasoning tasks under tight budgets, SDR offers a straightforward path to maintain capability while improving domain performance.