← News

Agents & MCPSources: arXiv · 16 Nov 2026

RAG-FGO paper frames query rewriting as black-box optimization for retrieval-augmented generation

A new arXiv paper proposes a framework that dynamically optimizes query-rewriting agents, treating prompt templates and decoding parameters as jointly tunable parts of generative RAG systems.

In plain words

Retrieval-augmented generation systems often rewrite a user's question before searching external knowledge bases, and poor rewriting hurts results. The paper's authors propose RAG-FGO, which treats the rewriting setup as a black-box optimization problem, tuning both prompt templates and decoding parameters together. It also keeps a memory of successful strategies to stay stable when search feedback is noisy.

What was announced

The authors, writing in an abstract for the arXiv paper, identify three obstacles for generative retrieval in real deployment: noisy retrieval feedback, the difficulty of jointly optimizing discrete prompt templates alongside continuous decoding-control parameters, and the weak generalization of static rewriting strategies in semantically complex scenarios. RAG-FGO responds by formulating query rewriting as a non-differentiable black-box optimization problem in a mixed discrete-continuous space, representing templates and decoding parameters as adaptively optimizable agents.

Mechanically, the framework combines global search, which the authors say enables semantic jumps and structural reorganization to escape local optima, with local perturbations for finer alignment. An elite-preserving memory pool accumulates high-quality strategies across iterations, which they report mitigates stochasticity and noisy feedback. Experiments on a diverse suite of benchmarks are described as demonstrating effectiveness and robustness, supporting applicability to complex semantic retrieval tasks; the abstract gives no benchmark names or figures.

What it does not show

  • The evidence is limited to the paper's abstract; no benchmark names, quantitative results, baselines, or code availability are stated, and no independent evaluation is available.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.