Airbnb introduces CASTLE for natural-language search cold start
Airbnb has described CASTLE (Contrastive And Seed-guided Training for natural Language sEarch), a framework for preparing natural-language search before real user queries are available. It generates synthetic queries and relevance labels together, using contrastive listing pairs derived from de-identified booking sessions selected according to privacy permissions. A booked listing is treated as positive and a viewed-but-not-booked alternative as negative; the generated query is intended to fit the former rather than the latter.
The company reports a KL divergence of 1.01 between generated and real query distributions, compared with 9.33 for its best baseline. In a human evaluation of 200 triplets, three subject-matter experts agreed with relevance assignments at rates of 91–93%. For cases without contrastive booking data, including newly listed properties, Airbnb says it supplements the method with an LLM-based Virtual Judge. These are company-reported results; the article does not establish independent replication.