Consistency Meets Verification: Enhancing Test Generation Quality in Large Language Models Without Ground-Truth Solutions

Researchers introduced ConVerTest, a novel two-stage pipeline that generates reliable software tests without needing existing code. It combines Self-Consistency for majority voting, Chain-of-Verification for iterative refinement, and Dual Execution Agreement for cross-validation. On BIGCODEBENCH and LBPP benchmarks, it improved test validity by up to 39%, line coverage by 28%, and mutation scores by 18% over existing baselines, significantly reducing AI hallucinations in test generation.