Neural Program Repair by Jointly Learning to Localize and Repair

1 INTRODUCTION

• Variable misuse occurs when a variable use has the wrong variable, a bug class studied alongside neural program representations for repair, optimization, and synthesis.Prior neural representations include token sequences, parse trees, program traces, and graphs.Source: §1 · p. 1
• Enumerating one repair prediction per variable-use slot loses shared context, creates a training/test mismatch, and requires thresholding many predictions.The paper reports a 4% to 14% accuracy drop from the mismatch even when the bug location is provided; threshold choice also trades false positives against false negatives.Source: §1 · p. 2
• The model jointly classifies a program as faulty or correct, localizes a fault when present, and repairs the localized variable misuse using pointer networks over program tokens.The paper presents extensive evaluation across multiple large program datasets.Source: §1 · pp. 2–3
• A variable token can be the faulty use, a location containing the correct variable for repair, or neither, enabling pointers to reuse repair ingredients elsewhere in the program.The problem includes mistakes such as failing to rename a variable after copying code or confusing names within one scope.Source: §1 · pp. 1–2
• The joint model achieves higher classification, localization, and repair accuracy, and a high true-positive score, than the enumerative approach.The paper also reports a 7-percentage-point advantage over a graph neural network on repair given a known bug location, but that comparison uses syntactic inputs only.Source: §1 · p. 2
• The pointer model exceeds the graph neural network by 7 percentage points for repair with a known bug location, but this comparison is limited to syntactic inputs.The graph model uses syntax and semantics; comparison with semantics and joint prediction at unknown locations are left for future work.Source: §1 · pp. 2–3

2 RELATED WORK

• Allamanis et al. (2018) enumerate variable-use predictions with a graph neural network using syntax and semantics, while Devlin et al. (2017b) enumerate repair candidates and score them.The paper notes that both approaches face a training/test mismatch; its own multi-headed pointers jointly classify, localize, and repair.Source: §2 · p. 3
• DeepFix and SynFix target syntax errors, while this work localizes and repairs semantic variable misuses.DeepFix generates a replacement line after localization; SynFix uses compiler-identified locations and constraint-based search.Source: §2 · p. 3
• DeepBugs (Pradel & Sen, 2018) classifies expressions for three name-based bug types, whereas this model embeds the full program and points to variable-use repairs.The pointer can select variables outside the buggy expression.Source: §2 · p. 3
• The Sk p encoder-decoder repairs statements, but predicting full statements may make variable-misuse repairs requiring variables defined multiple lines earlier difficult.The paper contrasts this with its pointer mechanism for selecting variable uses.Source: §2 · p. 3
• Traditional specification-guided repair relies on correctness specifications and hand-designed search, whereas learning-based repair learns common error types from source examples and can classify, localize, and repair them.The paper distinguishes those common-error benchmarks from test-based benchmarks whose bugs are specific to individual programs.Source: §2 · pp. 3–4

3.2 MULTI-HEADED POINTER MODELS

• The model encodes program tokens with an LSTM and uses two pointer distributions to predict the bug location and repair-variable location.The pointers are trained end-to-end; the paper focuses on localizing a single bug and notes that more pointer heads can support multiple bugs.Source: §3 · p. 4, §3.2 · p. 5

3.1 PROBLEM DEFINITION

• VARMISUSEREPAIR asks whether a program is correct or, if faulty, which variable-use location is wrong and which existing variable occurrence can repair it.The repair variable must be one defined in the program; variable-use scope may be defined differently, and the comparison with Allamanis et al. uses their broader load-or-store definition.Source: §3.1 · p. 4

3.3 TRAINING THE MODEL

• Training examples come from program definitions and include synthetic buggy examples formed by replacing a chosen variable with an incorrect one.The training set also includes non-buggy examples, balanced 50-50 with buggy examples; correct-variable repair targets can mark all occurrences of the intended variable.Source: §3.3 · pp. 5–6
• The joint model minimizes the sum of location and repair losses, while the enumerative baseline uses a repair-only pointer model.An alternative repair loss uses the maximum of the probabilities of repair pointers instead of their addition.Source: §3.3 · p. 6

4 EVALUATION

• Evaluation uses ETH-Py150 Python functions and MSR-VarMisuse C# files; the ETH enumerative comparison uses a filtered test set of 12,218 examples.The ETH set creates buggy and bug-free examples for training, while its evaluation set samples one slot per function.Source: §4 · pp. 6–7

4.1 JOINT MODEL VS. ENUMERATIVE APPROACH

• The joint model reaches 71% localization accuracy and 65.7% localization+repair accuracy, improving on the enumerative approach while maintaining 84.5% true-positive and 82.4% classification accuracy.The reported localization and repair improvements over the enumerative approach's lowest-threshold, highest-k setting are about 6.4% and 9.9%, respectively.Source: §4.1 · p. 7, Table 1 · p. 8

4.2 EFFECT OF INCORRECT SLOT PLACEMENT

• Incorrect slot placement drops repair accuracy by about 4.3 percentage points for Any at τ = 0.5 and by 8 to 14.6 points for the more adversarial Near setting.Near injects a bug within two variable-use locations of the slot and in the first 30 program tokens.Source: §4.2 · p. 8, Figure 3 · p. 9

4.3 COMPARISON OF GRAPH AND POINTER NETWORKS

• On the MSR-VarMisuse seen test, pointer-based repair reaches 62.3% accuracy versus 55.3% for the graph-based model's syntax-only ablation, a 7-point advantage.The comparison is limited to syntactic inputs; the graph model uses syntax and semantics, and the pointer model's training dataset is about 30% smaller.Source: §4.3 · p. 9

4.4 EVALUATION ON VARIABLE MISUSE IN PRACTICE

• On 4,592 industrial snapshot pairs, the joint model achieves 67.3% true positive rate, 66.7% classification accuracy, 21.9% localization accuracy, and 15.8% localization+repair accuracy, outperforming enumeration on each.The enumerative approach scores 41.7%, 47.2%, 6.1%, and 4.5% on those respective metrics; the joint model localizes and repairs 727 instances in aggregate.Source: Table 2 · p. 10, §4.4 · p. 10

• The joint model significantly outperforms an enumerative approach for variable-misuse repair.Source: §5 · p. 10
• The model uses the insight that both the bug and its repair must exist in the original program.Source: §5 · p. 10
• Future work will explore joint localization and repair with graph models, pointer-graph combinations, and more semantic program information.Source: §5 · p. 10

• ETH-Py150 slots are variable uses in load contexts, and repair candidates are in-scope variables, including locals, globals, and formal arguments.Source: A TRAINING DATA GENERATION · p. 12
• Examples are generated only for slots with at least two repair candidates; each eligible slot gets a buggy example and a bug-free example, with correct functions duplicated across examples.Source: A TRAINING DATA GENERATION · p. 12
• MSR-VarMisuse contains 25 C# projects split into train, validation, seen-test, and unseen-test partitions; its slots include load and store uses with type-compatible candidates.Source: A TRAINING DATA GENERATION · p. 12