CoDA-Bench & TACO: Scalable, Verifiable Task Construction

CoDA-Bench & TACO: Scalable, Verifiable Task Construction

ICML / VLDB 2026

Co-author · Task synthesis, execution verification, difficulty refinement

• CoDA-Bench: Co-authored a benchmark of 1,009 tasks across 31 data communities, built from executable notebook solutions through task reconstruction, adversarial refinement, and human verification.
• TACO: Co-authored a benchmark combining 1,500 expert-annotated and 13,000 synthetic NL–SQL pairs, using query-structure sampling, schema-constrained generation, and quality checks.