In-Context Optimization for Retrieval-Augmented Generation: A Gradient-Descent Perspective

In three linesTheoretical study showing that one linear self-attention layer can implement a gradient-descent step on a unified RAG objective. Authors propose a lightweight method to adapt query-evidence interaction without modifying retriever or backbone, tested on 7 QA benchmarks with consistent improvements over baseline.

Summary generated by Claude — human-verified