Fine-Tuning with RAG to Teach LLMs New Skills (ICLR 2026)

A well-received paper at ICLR 2026 proposes converting inference-time retrieval into learned competence via distillation: extract compact hints from agent failures, use them to generate improved teacher trajectories, then fine-tune. The approach turns what RAG retrieves at runtime into durable model knowledge. Validated on ALFWorld (household tasks) and WebShop (online shopping). Relevant for teams hitting the ceiling of pure RAG and evaluating whether fine-tuning is worth the investment.

---

Sources: arXiv:2510.01375