Fail-Aware and Explainable Test Oracle Prediction

The paper introduces FOCAL, a code-LLM-based discriminative test oracle predictor. Instead of generating assertions, it predicts whether a test prefix executed against a method will pass or fail. FOCAL trains on labeled test-prefix and method-under-test pairs, emphasizes failing examples through its loss design, and grounds predictions in statement-level behavioral evidence. According to the abstract, it improves failure-case performance over SEER on unseen projects and produces richer explanations. A preliminary evaluation on fault-detection benchmarks and automated test-generation artifacts suggests that FOCAL can complement fuzzing, search-based testing, and LLM-based test generation by adding fault-oriented oracle predictions.