LIVE

Airbnb Researchers Unveil CASTLE Framework to Train Search Models Using Synthetic Data

🤖 AI Models 2 min read 👁 8

Airbnb has introduced CASTLE, a new framework that uses large language models and de-identified booking data to generate synthetic search queries and relevance labels for natural language search.

Airbnb Researchers Unveil CASTLE Framework to Train Search Models Using Synthetic Data🤖AI Models

Photo: infoq.com

Key points

  • Airbnb introduces CASTLE to tackle the cold-start problem in natural language search.
  • The framework generates synthetic queries and relevance labels jointly using large language models and booking session data.
  • In testing, CASTLE achieved a 9.2-fold reduction in distribution divergence compared to previous baselines.

Researchers at Airbnb have developed a framework named CASTLE to solve the cold-start problem in natural language search systems. By leveraging de-identified booking sessions and large language models, the approach generates synthetic search queries alongside relevance labels simultaneously, allowing models to be trained effectively before real search traffic builds up.

Natural language search allows users to look for accommodations using everyday conversational phrases like "pet-friendly beach house that sleeps eight" instead of relying strictly on standard filters. Traditional training methods usually separate query generation and relevance labeling into distinct steps, which can introduce errors and discrepancies between artificial and actual user behavior. CASTLE avoids this by utilizing contrastive listing pairs from past bookings, where a booked property acts as a positive example and an alternative viewed property acts as a negative one.

The system integrates template-based prompts, few-shot prompting utilizing seed queries from user research, and attribute-grounded prompts to create realistic variations in query length and phrasing. For new properties lacking historical booking data, the framework employs an LLM-based Virtual Judge to assign relevance labels. In evaluations, CASTLE achieved a significantly lower Kullback-Leibler divergence against real user query distributions compared to existing baselines, and human evaluators agreed with its relevance assignments up to 93 percent of the time.

What to watch next: Further investigations into strict data splitting methods to separate the learning of user intent from familiar property templates.

Source infoq.com — infoq.com/news/2026/11/airbnb-castle/?utm_campaign=infoq_content&ut… ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.