Amazon Nova Explores Self-Distilled Reasoning for Fine-Tuning

Researchers at Amazon Nova are delving into a novel approach called Self-Distilled Reasoning (SDR) to enhance supervised fine-tuning (SFT) of language models. The core idea behind SDR is to create synthetic 'thinking tokens' for datasets that do not inherently contain explicit reasoning steps. This method directly tackles the challenge known as the reasoning suppression problem, where models might struggle to articulate their thought processes during fine-tuning. The efficacy of SDR has been rigorously tested and confirmed across three distinct benchmarks, leading to the formulation of practical guidelines for its implementation.