ZO — Good search should be free.

An arXiv Preprint on Retrieval Augmented Generation

Asked: an arXiv preprint on retrieval augmented generation

RAG is an open-book exam: instead of answering from memory alone, the model looks up the right pages first, then writes its answer with them in hand [6].

Analogy

2506.06962

AR-RAG · arXiv 2025

[1]

Retrieve

A query fetches relevant passages from an external corpus at inference time [6][7].

Augment

The retrieved passages are concatenated with the prompt, grounding the model in non-parametric memory [6][11].

Generate

The generator conditions on both its parametric memory and the retrieved text to produce the output [6][11].

Pick a preprint

kNN patches

Retrieval

Context-aware at each step

Autoregressive

Generation

Patch-level conditioning

Training-free

DAiD

Merges predicted + retrieved distributions

Fine-tuning

FAiD

Smooths features via multi-scale convolution

Open this in ZO