A language model can rewrite documents to match what search engines want, making cheaper models work as well as expensive ones

Researchers used reinforcement learning to teach a language model how to rewrite documents so they rank higher in search results, without changing the search engine itself. In tests, rewriting documents with a cheap embedding model matched the performance of search engines 6.5 times more expensive.

Search has always meant a choice: pay for a good retriever or live with poor results. This work shows you can change the documents instead of the retriever. The practical implication is that smaller, cheaper AI models become competitive with larger ones — if you're willing to transform your documents at indexing time. This matters because search cost scales with model size and query volume; document optimization moves that cost offline.