Revisiting the Evaluation of Deep Learning-Based Compiler Testing
Bibliographic record
Abstract
A high-quality program generator is essential to effective automated compiler testing. Engineering such a program generator is difficult, time-consuming, and specific to the language under testing, thus requiring tremendous efforts from human experts with language-specific domain knowledge. To avoid repeatedly writing program generators for different languages, researchers recently proposed a language-agnostic approach based on deep learning techniques to automatically learn a program generator (referred to as DLG) from existing programs. Evaluations show that DLGs outperform Language-Specific Program Generators (LSGs) in testing compilers. However, we argue that it is unfair to use LSGs as baselines to evaluate DLGs. LSGs aim to validate compiler optimizations by only generating compilable, well-defined test programs; this restriction inevitably impairs the diversity of the language features used in the generated programs. In contrast, DLGs do not aim to validate the correctness of compiler optimizations, and its generated programs are not guaranteed to be well-defined or even compilable. Therefore, it is not surprising that DLG-generated programs are more diverse in terms of used language features than LSG-generated ones. This study revisits the evaluation of DLGs, and proposes a new, fair, simple yet strong baseline named Kitten for evaluating DLGs. Given a dataset consisting of human-written programs, instead of using deep learning techniques to learn a program generator, Kitten directly derives new programs by mutating the programs in the dataset. Extensive experiments with more than 1,500 CPU-hours demonstrate that the state-of-the-art DLGs fail to compete against such a simple baseline: 3 v.s. 1,750 hang bugs, 1 v.s. 34 distinct compiler crashes. We believe that DLGs still have a large room for improvement.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.006 | 0.033 |
| Meta-epidemiology (narrow) | 0.003 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.005 |
| Open science | 0.005 | 0.003 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.003 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".