AuXSpec: Learning Autoformalized Program Specifications via Executable Postconditions
AuXSpec: Learning to Autoformalize Program Specifications via Executable Postconditions
AI summary
AuXSpec pairs declarative postconditions with executable versions, using the consistency of test case execution as the reinforcement learning reward. Using 3,037 LeetCode problems and Codeforces expert iteration to expand to 19.5K samples, Qwen3.5-9B and Qwen3.8-27B achieve SpecGen pass@4 of 82.7% and 84.3% on AuXSpec-Test, with the 27B model surpassing all frontier models.