every-state-of-the-art-model-tested-repairs-fewer-defects4j-gitbug-java-bugs CLM-0299Every state-of-the-art model tested repairs fewer Defects4J and GitBug-Java bugs after semantics-preserving transformation of the buggy code, from -4% for GPT-4o to -16% for Llama-3.1, and the size of each drop correlates with how likely the model was to have memorized the original (negative log-likelihood), so part of published repair success on these benchmarks is recall of seen fixes rather than debugging.
Capability: Finding and fixing a fault in existing code
Observed on
Seven LLMs, Defects4J and GitBug-Java, 2026; per-bug cases where success fell from 80% to 0% after transformation; effect sizes reported..
Sources
- Direct test with a plausible mechanism and a correlation, not a causal isolation; GPT-4o's drop is small though significant. Applies to Java benchmarks that predate most training sets.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- In LLM-based program repair on Defects4J bugs with frontier models (GPT-5, DeepSeek-R1, Claude Sonnet 4.5) prompted zero-shot, passing the developer-written test suite does not mean the patch is grounded in the bug: manual inspection of 812 sampled repairs found ungrounded repairs in 72.7% of cases including test-passing patches, most often from incorrect causal localization (45.9%) rather than a wrong edit style, and the same models frequently misidentify the bug-triggering tests and generate additional tests that miss the triggering condition.Stating false facts confidently · reviewed by AI
- In zero-shot LLM program repair on Defects4J bugs, passing the developer-written test suite does not indicate that the patch is grounded in a correct understanding of the bug: manual inspection of sampled repairs found ungrounded patches in about three quarters of cases including ones that passed all available tests, with incorrect causal localization the largest category.Stating false facts confidently · reviewed by AI
- Fine-tuned program-repair models that score well on HumanEval-Java-Bug lose more than half their Pass@10 under semantics-preserving rewrites of the buggy code, with local-variable renaming alone cutting performance by 50-57%, so much of measured repair skill is tied to the exact surface form of the bug.Finding and fixing a fault in existing code · reviewed by AI
- On real C/C++ faults with reproducing tests, 24 current models repair only about 11% of the 248 bugs and about 6% of the 102 vulnerabilities, far below their Java and Python repair rates on comparable benchmarks, and fine-tuning on the benchmark's commit corpus raised Pass@k in 21 of 28 comparisons by an average 85% relative, so the gap is partly a training-data gap.Finding and fixing a fault in existing code · reviewed by AI
- For LLM program-repair workflows built on self-correction, generated-knowledge or multi-agent scaffolding, the gain over plain prompting is unstable across benchmarks and can turn negative once extrinsic defect information (error messages, triggering tests, bug comments) is withheld, while simply putting the error message plus the failing test into the prompt yields more correct repairs than the scaffolds do.Fixing its own mistakes · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: Models trained after the benchmarks' cutoff showing no NLL-correlated drop, or transformations shown to change bug difficulty rather than reveal memorization.