Source: AWS Machine Learning Blog — 2026-07-21
Summary
AWS describes Self-Distilled Reasoning (SDR), a fine-tuning recipe for Amazon Nova models that fixes a specific problem: supervised fine-tuning (SFT) datasets usually only contain final answers, not the reasoning that led to them, and training a reasoning model on answer-only data tends to suppress the very reasoning behavior you want to keep. SDR's fix is to have the base reasoning model generate its own thinking traces for the existing SFT examples, prepend those traces to the original outputs, and fine-tune on the combined reasoning-plus-answer sequence — using the model's own reasoning as the training target instead of importing someone else's.
Key Takeaways
- The problem SDR targets: most SFT datasets are answer-only (input → correct output), with no recorded reasoning trace — fine-tuning directly on that data tends to suppress a reasoning model's chain-of-thought behavior rather than sharpen it.
- The three-stage recipe: (1) run the base reasoning model (e.g., Amazon Nova 2 Lite) over each SFT example to generate a reasoning trace for it, (2) prepend that generated trace to the example's original output, (3) fine-tune on the combined sequence with reasoning mode turned on, so supervision covers both the reasoning tokens and the final answer.
- Because the reasoning traces come from the model's own base policy rather than an external teacher, they act as a soft regularizer — keeping the fine-tuned model close to its own pretrained behavior instead of overwriting it, which AWS frames as protection against catastrophic forgetting.
- It also converts answer-only supervision into a form of process supervision: the model gets rewarded for the intermediate reasoning path, not just the final token, which is closer to step-by-step reward shaping than pure outcome-based SFT.
- AWS reports the approach was validated across three benchmarks with practical tuning recommendations, positioning it as a reusable recipe for anyone fine-tuning Nova (or similarly-shaped reasoning models) on datasets that were never built with reasoning traces in mind.
Reel Script
Hook (~17s, 39 words) If you fine-tune a reasoning model on plain input-output pairs, you can accidentally train the reasoning right out of it. AWS just published a fix for Amazon Nova that turns any answer-only dataset into one a reasoning model can actually learn from.
Core Concept (~65s, 145 words) Reasoning models like Nova are trained to think step by step before answering — that chain-of-thought is what you're paying for. But most fine-tuning datasets only record the final correct answer, because that's what was easy to collect. Fine-tune directly on those answer-only pairs and you're implicitly teaching the model "skip the thinking, jump to the answer" — you erode the exact reasoning behavior the base model had. Self-Distilled Reasoning solves this by not inventing new reasoning traces from scratch or importing them from a bigger teacher model. Instead, it asks the model's own base version to generate a reasoning trace for each existing example first — essentially having Nova explain its own thinking on data it never explained before — then trains on that self-generated trace plus the original answer together.
Hands-On (~55s, 122 words) The recipe runs in three concrete steps. Step one: take your existing SFT dataset — just inputs and correct outputs — and run each example through the base reasoning model with reasoning mode on, capturing whatever chain-of-thought it produces on its own. Step two: splice that generated trace onto the front of the original output, so each training example now reads as reasoning-then-answer instead of answer-only. Step three: fine-tune on that combined sequence, supervising both the reasoning tokens and the answer tokens together, rather than masking the reasoning portion out. Because the traces came from the model's own policy, the fine-tuned model doesn't get dragged toward an outside style — it gets pulled toward more of what it was already doing, just made deliberate and consistent.
Takeaway (~22s, 49 words) If you're fine-tuning a reasoning model on data that was never built with reasoning in mind, plain SFT is quietly working against you. Self-distillation — using the model's own traces as the supervision — is the cheaper fix over hand-labeling reasoning data. Worth trying before you reach for a bigger teacher model.