Text-to-SQL Benchmarks Must Reflect Real-World Data Complexity

Current text-to-SQL benchmarks often use clean, simple schemas that don't represent messy real-world databases. A new article argues that benchmarks must include challenges like ambiguous column names, missing values, and complex joins. Without such realism, progress in natural language interfaces for databases may be misleading.