Aligning Offline Metrics and Human Judgments of Value for Code Generation Models
This addresses the problem of misaligned evaluation metrics for code generation models, which is important for developers and AI researchers, though it is an incremental improvement on existing evaluation methods.
The paper demonstrates that functional correctness alone underestimates the productivity value of code generation models, as programmers still find code valuable even when it fails unit tests if it reduces overall effort. They propose a hybrid metric combining correctness and syntactic similarity that achieves 14% stronger correlation with human value judgments.
Large language models have demonstrated great potential to assist programmers in generating code. For such human-AI pair programming scenarios, we empirically demonstrate that while generated code is most often evaluated in terms of their functional correctness (i.e., whether generations pass available unit tests), correctness does not fully capture (e.g., may underestimate) the productivity gains these models may provide. Through a user study with N = 49 experienced programmers, we show that while correctness captures high-value generations, programmers still rate code that fails unit tests as valuable if it reduces the overall effort needed to complete a coding task. Finally, we propose a hybrid metric that combines functional correctness and syntactic similarity and show that it achieves a 14% stronger correlation with value and can therefore better represent real-world gains when evaluating and comparing models.