Airbnb researchers have introduced CASTLE, a framework for training natural-language property search before enough real queries exist. It starts with de-identified booking sessions in which a guest viewed several listings and selected one. The booked property becomes a positive example and an unbooked alternative becomes a negative one.

A language model then writes a query that fits the positive listing but not the negative. That construction generates the query and its relevance label together, avoiding a separate labeling model that could introduce another source of error. CASTLE combines template prompts, roughly 500 seed queries from user research and prompts focused on attributes that distinguish properties.

In Airbnb’s evaluation, CASTLE’s query distribution had a Kullback–Leibler divergence of 1.01 from real-user searches, compared with 9.33 for the strongest baseline. Lower is better. Three domain experts agreed with its relevance labels on 91% to 93% of 200 sampled examples. A production variant, CASTLE-R, can incorporate real traffic after launch.

One unresolved question is whether templates and listing identities are fully separated between training and evaluation. The researchers say they tested on untouched real queries and used leakage guardrails, but did not publish a controlled comparison using splits where both seed templates and listing IDs were disjoint.