ICLR 2026: Distilling Agent Failures Into Student Models — 91% ALFWorld Success via One-Shot RAG Hint Removal

arXiv / ICLR 2026·high signalXBlueskyLinkedInCopy link

An ICLR 2026 conference paper presents a pipeline that converts inference-time retrieval into learned competence through distillation. The method extracts compact hints from agent failure trajectories, generates improved teacher trajectories via one-shot retrieval, then trains student models on these trajectories with hint strings removed — forcing internalization rather than memorization. Distilled 7B/14B students achieve 91% on ALFWorld (vs 79% baseline) and 72 on WebShop (vs 61 baseline) while using 10-60% fewer tokens than retrieval-augmented teachers. Generalizes across ReAct and StateAct architectures.