Exploring self-distilled reasoning for supervised fine-tuning with Amazon Nova
AWS Rushil Anirudh
Amazon researchers introduced Self-Distilled Reasoning (SDR), a technique for fine-tuning models that lack reasoning traces by using the base model's own chain-of-thought outputs as training signals. The method recovered mathematical performance from 6 percent back to 70 percent compared to vanilla supervised fine-tuning, while improving target task performance by over 6.5 percent on average. SDR eliminates the need for human annotation, prevents catastrophic forgetting without post-hoc model merging, and enables reasoning capabilities to be preserved during domain-specific customization.
Why it matters
In this post, we explore an idea for generating thinking tokens for datasets that lack reasoning traces in SFT customization. We first examine the reasoning suppression problem, then introduce Self-Distilled Reasoning (SDR), validate it across three benchmarks, and provide practical recommendations.