Danqing Huang

Text-to-SQL models can now achieve higher accuracy with fewer tokens by reasoning about multiple possible query paths and selectively gathering evidence only when uncertain about which schema elements are needed.

Appraisal slashes the computational cost of privacy-preserving record linkage, making it feasible to screen 850x more records than existing systems when identifying potential data collaboration partners.

Training GBDTs on vertically partitioned data doesn't have to leak which records are shared: this new protocol hides record IDs while maintaining competitive efficiency.