Research map / Large language models / Language, safety and interpretability
Deep Just-In-Time Inconsistency Detection Between Comments and Source Code
Sheena Panthaplackel, Junyi Jessy Li, Milos Gligoric, Raymond J. Mooney
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use explicit code edit representations instead of only the changed method (ablation) | Just-in-time inconsistency detector · Cleaned Test Sample | F1: 59.9 -> 83.1 | Improved |
| Add linguistic and lexical surface features to the just-in-time detector (ablation) | Hybrid just-in-time detector · Cleaned Test Sample | F1: 83.1 -> 87.1 | Improved |
| Add an AST code-edit encoder to the sequence edit encoder (ablation) | Hybrid just-in-time detector · Cleaned Test Sample | F1: 81.5 -> 83.1 | No clear effect |
| Add code changes and their prior version instead of using only the current method | Inconsistency detector · Cleaned Test Sample | F1: 63.3 -> 83.1 | Improved |
| Use implicit copying behavior for detection rather than an explicit detection component (ablation) | Comment update system · Cleaned Test Sample | Recall: 23.3 | Hurt |
| Combine explicit inconsistency detection with a comment update model | Comment maintenance system · Cleaned Test Sample | xMatch: 50.0 -> 62.3 | Improved |
| Jointly train detection and comment update rather than pretrained detection plus update (ablation) | Comment maintenance system · Cleaned Test Sample | F1: 86.4 -> 87.2 | No clear effect |
| Train across comment types rather than only on return comments (ablation) | Return-comment detector · Return comments | F1 | Hurt |
| Train across comment types rather than only on parameter comments (ablation) | Parameter-comment detector · Parameter comments | F1 | Hurt |
| Train across comment types rather than only on summary comments (ablation) | Summary-comment detector · Summary comments | F1 | Improved |
Closest papers by meaning
- ChatGPT-Generated Code Assignment Detection Using Perplexity of Large Language Models (Student Abstract) (AAAI 2024 · similarity 0.84)
- CodeSync: Synchronizing Large Language Models with Dynamic Code Evolution at Scale (ICML 2025 · similarity 0.84)
- Unsupervised Evaluation of Code LLMs with Round-Trip Correctness (ICML 2024 · similarity 0.84)
- Towards Practical Defect-Focused Automated Code Review (ICML 2025 · similarity 0.83)
- ObscuraCoder: Powering Efficient Code LM Pre-Training Via Obfuscation Grounding (ICLR 2025 · similarity 0.83)
- Detecting AI-Generated Code Assignments Using Perplexity of Large Language Models (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
