AWS shows self-distilled reasoning can preserve math and coding skills during SFT

AWS ML Blog · rss · 2026-07-22

AWS proposes self-distilled reasoning to preserve capability during SFT

• The post addresses a common SFT problem: datasets often lack high-quality chain-of-thought traces, making reasoning expensive to supervise.
• AWS introduces Self-Distilled Reasoning (SDR), which reuses the base model’s own chain-of-thought as reasoning traces during customization.
• In experiments across three benchmarks, SDR improves target performance while reducing catastrophic forgetting compared with vanilla SFT and often outperforms simple model merging.
• One reported result: vanilla SFT caused math performance to drop from 70% to 6% on average, while SDR nearly restored it.
• The approach requires no separate teacher model and no human annotation, and is presented as applicable to existing SFT datasets across domains.

Original post →

More from Research

• OpenAI and Apollo Research introduce Contrastive SDF to measure reward-seeking — OpenAI · 2026-07-22
• NVIDIA says to tune the harness before tuning the model with LangChain — NVIDIAAI · 2026-07-22
• The Thimble and the Waterfall: AI's Data Bottleneck and Feedback Loops — dyamins · 2026-07-22
• NVIDIA shows 22 SIGGRAPH papers and Omniverse tools for robot simulation — facontidavide · 2026-07-22
• Building a Knowledge Graph Without a Graph DB: 1000x Cheaper Than GraphRAG — TheRedfather · 2026-07-22
• Agentic RAG survey maps planner, retriever and refinement agents for complex retrieval — blaizedsouza · 2026-07-22

Research channel →