Confidence-guided method improves testing for AI code ranking

Confidence-Gated Transductive Test Generation for Code Reranking

Artificial IntelligenceComputation and LanguageSoftware Engineering

Summary

It is hard to create good test examples to check if computer programs generated by AI are correct, especially when expected answers are unknown. The authors developed a new method called CoTT that first tries a quick approach and only does extra checking when it isn't confident enough. This saves time while still making sure the test results are trustworthy. Their tests show CoTT works better and faster than older methods at ranking AI-generated programs.

What this means in practice

Authors

Sungjae Lee, Youngsik Yoon, Seockbean Song, Siwei Wang, Wei Chen, Jungseul Ok

Abstract

Test case synthesis is crucial for evaluating and ranking programs generated by large language models (LLMs). However, constructing high-quality test cases remains challenging because reliable expected outputs are often difficult to obtain. We propose Confidence-Gated Transductive Test Generation (CoTT), which first uses an efficient inductive procedure and invokes transductive generation only when inductive confidence is low. This adaptive design improves output reliability while allocating extra computation only when needed. On code reranking benchmarks, CoTT outperforms prior baselines across the reported metrics while reducing cost relative to applying transductive generation to every input. These results show that confidence-based allocation of test-time computation provides a favorable efficiency-effectiveness trade-off with a single efficient LLM.