Google research makes query fan-out 10x faster

Google Research has published R4T, short for Retrieve-for-Train, a method that makes query fan-out much cheaper to run. The blog post, dated 15 September 2026, describes a paper accepted at ICML 2026. It is research, tested on fashion and music datasets, and Google has not said it runs in Search.

What query fan-out is#

Query fan-out turns one broad request into several related sub-queries. Google’s own example is “camping gear”: a searcher wants a tent, a sleeping bag, a stove and a headlamp, not “ten slight variations of four-person tents.” Google’s AI features documentation already says AI Overviews and AI Mode may use fan-out, issuing “multiple related searches across subtopics and data sources” to build an answer.

The problem R4T tackles is cost. Asking a large language model to plan sub-queries takes many reasoning tokens, which Google says is “fundamentally at odds with the sub-second response times required by a production search bar.”

How R4T works#

The method moves the expensive part offline, in three steps:

• Train a fan-out model with reinforcement learning. Open models (Gemma3-4B and Qwen3-4B) learn to write sub-queries scored on three balanced goals: grounded in real items, diverse, and aligned with the original query.
• Generate training data. The trained model produces query and result-set pairs, with no human labels.
• Train a small, fast model. A 53.9 million parameter diffusion model learns to produce the whole set of results in one pass, with no step-by-step reasoning.

The balance matters: optimizing for grounding alone produced nonsense, and adding alignment alone produced near-duplicate paraphrases. Diversity, measured with the Vendi Score, closed both shortcuts.

What the results cover#

The paper reports better result sets than strong baselines and fan-out latency cut “by an order of magnitude”, on large fashion and music benchmarks. Those are curated outfits and music playlists, not the web. Neither the blog post nor the paper mentions Google Search, AI Mode or AI Overviews. Coverage describing R4T as production-ready search goes beyond what Google has published.

Why we care#

Google says its AI features may already use fan-out, and research like this aims to make it cheaper to run. The practical lesson is unchanged by R4T: pages get cited for the sub-queries they answer well, so cover the related questions around a topic, not only its head term. The AI answer readiness checker scores whether a page answers a question directly. Follow AI search for whether Google ever confirms R4T.