Interrogative Comments Posed by Review Comment Generators: An Empirical Study of Gerrit
Bibliographic record
Abstract
Background: Review Comment Generators (RCGs) are models trained to automate code review tasks. Prior work shows that RCGs can generate review comments to initiate discussion threads; however, their ability to interact with author responses is unclear. This can be especially problematic if RCGs pose interrogative comments, i.e., comments that ask questions of other review participants. Aims: We set out to study the prevalence of RCG-generated interrogative code review comments, their similarity with the interrogative comments of humans, and the predictability of the generation of interrogative comments. Method: We study three task-specific RCGs and three RCGs based on Large Language Models (LLMs) on data from the Gerrit project using quantitative and qualitative methods. Results: We find that RCGs: (1) generate interrogative comments at a rate of$\mathbf{1 5. 6 \%} \boldsymbol{-} \mathbf{6 5. 2 6 \%}$; (2) differ from humans in generating such comments, which can stifle conversations if RCGs dissuade human reviewers from commenting deeply; and (3) produce interrogative comments with low predictability. Finally, we find that (4) the interrogative comments posed by LLMbased RCGs can differ even more substantially from human behaviour than those of task-specific RCGs. For example, the studied LLM-based RCGs pose rhetorical questions 3.16% of the time, whereas human-submitted interrogative comments pose rhetorical questions 8.74 % of the time. Conclusions: Our results suggest that neither task-specific nor LLM-based RCGs can replace human reviewers yet; however, we note opportunities for synergies. For example, RCGs tend to raise pertinent questions about exception handling of common APIs more frequently than human reviewers. Putting greater emphasis on technical comments generated by RCGs (rather than conversational ones, such as interrogative ones) will likely improve their perceived usefulness.
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 | 0.083 | 0.394 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.004 | 0.002 |
| Science and technology studies | 0.003 | 0.003 |
| Scholarly communication | 0.004 | 0.005 |
| Open science | 0.003 | 0.004 |
| Research integrity | 0.003 | 0.003 |
| Insufficient payload (model declined to judge) | 0.002 | 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".