AWS: Self-Distilled Reasoning restores Amazon Nova 2 math accuracy from 6% to about 68% after fine-tuning

What is Self-Distilled Reasoning and what problem does it solve?

AWS has introduced Self-Distilled Reasoning (SDR), a technique for supervised fine-tuning (further training a model on labeled examples) of Amazon Nova 2 models in situations where the training data lacks a chain of thought — the reasoning steps a model goes through before its final answer. Such data is common in practice: companies have question-answer pairs without a recorded reasoning path, which makes it harder to train a model to reason on its own.

Why does vanilla fine-tuning wreck overall performance?

When a model is fine-tuned directly on narrow data without reasoning traces (vanilla SFT), the model’s overall math performance drops from the base model’s 70% to just 6% — a phenomenon known as catastrophic forgetting, the loss of previously learned abilities due to overly narrow subsequent training. The model becomes good at the narrow task but loses the math problem-solving ability it previously had.

SDR uses the model to generate its own training data

The solution AWS proposes uses the base Nova 2 Lite model to autonomously generate chain-of-thought traces for existing training data that lacks them, without needing human annotation. The enriched data is then used to fine-tune the target model. Result: SDR achieves over a 6.5% relative improvement in performance on the target task while simultaneously restoring overall math accuracy from 6% to about 68% — a difference that shows how much preserving reasoning ability during fine-tuning reduces collateral damage to the model’s other skills.

Tested on three different benchmarks

AWS tested the method on three benchmarks from different domains: MedMCQA (medical questions), CoCoHD, and Invoice-OCR (invoice processing). The consistency of results across three unrelated tasks suggests SDR is not a narrow solution for one type of data, but a general approach to fine-tuning when reasoning traces are simply missing from the original data.