Code Readability Results

Spearman correlation between predicted scores and human ratings, computed separately for each dataset. Counts show evaluated samples / total samples.

Independent Readability Interferences

Percentage of changed variants scored less readable than their matched original, reported separately for Java and Python. Counts show evaluated changed pairs.

Motivating example

A month-offset lookup adapted from Sakamoto’s compact day-of-week algorithm. All five developers rated it 1/5, yet four traditional models place it above the 74th percentile.

class MonthOffsets {
    static int offset(int month) {
        int value = "-ilkDwlu2thk5".charAt(month);
        return value % 7;
    }
}

Largest BTTF penalties and their source regions

Explainability analysis

LLM explanations and BTTF feature contributions for 30 randomly sampled programs.

Datasets and samples