differential testing

This study addresses the challenge of behavioral semantic alignment among natural language bug reports, triggering tests, and code patches in software defect repair, arguing that lexical similarity alone is insufficient to capture behavioral consistency. To bridge this gap, the authors propose Desc2Fix, a novel framework that reformulates semantic alignment as measurable engineering signals by integrating structured behavioral anchors, multi-granularity embeddings (e.g., SBERT, CodeBERT, OpenAI), and large language models (GPT-4o, DeepSeek-Chat) for signal extraction and alignment scoring. Experiments on 2,857 defect triplets demonstrate that LLMs can extract behavioral signals with high completeness (up to 90%), and that full-code diffs provide a more reliable basis for alignment than localized code fragments, significantly enhancing the reliability of test generation and patch ranking.

Existing approaches struggle to localize root causes of bug-fixing commits in complex software due to their inability to model heterogeneous commit structures and cross-line code dependencies, limiting localization accuracy. To address this, we propose a root-cause localization method based on heterogeneous graph neural networks. Our approach constructs a multi-granularity program dependency heterogeneous graph integrating Abstract Syntax Trees (ASTs), Control Flow Graphs (CFGs), and Data Flow Graphs (DFGs). We design a cross-line semantic preservation mechanism that modulates semantic propagation via a decay-enhancement gating scheme, and introduce a heterogeneous attention-driven semantic aggregation module to explicitly capture fine-grained dependencies among diverse node and edge types. Evaluated on 675 bug-fixing commits across 87 open-source projects, our method achieves an average 73.4% improvement in the MFR metric over state-of-the-art baselines, with a maximum gain of 96.83%, significantly enhancing the precision of identifying critical buggy lines.

Semantic conflict detection in collaborative development suffers from high false-positive rates, primarily because existing lightweight static analyses cannot reliably distinguish behavior-preserving code refactorings from genuine semantic changes. To address this, we propose a refactoring-aware static analysis method that, for the first time, deeply integrates automated refactoring identification into lightweight interference detection—enabling precise filtering of spurious conflicts introduced by refactorings while preserving coverage. We implement an end-to-end detection tool and evaluate it on 99 labeled scenarios and 1,087 real-world merge scenarios. Our approach reduces false positives by 31.8%, significantly improves precision, and incurs only a marginal drop in recall. The method is practical, efficient, and scalable, establishing a new paradigm for semantic conflict detection in collaborative software development.

This work addresses the semantic gap between problem descriptions and code implementations in software engineering fault localization by proposing a test-driven localization approach. It systematically demonstrates, for the first time, that test cases substantially reduce localization uncertainty and innovatively integrates domain knowledge to enrich test semantic representations. A hierarchical execution trace analysis mechanism is designed to effectively bridge the semantic discrepancy between requirements and code while suppressing noise. Evaluated on SWE-bench Lite, the method achieves a 41.57% relative improvement in function-level Recall@1 over the strongest baseline. When integrated into the Agentless pipeline, it resolves 17.72% more problems than the original framework.

In software requirements engineering, ambiguity in natural-language requirements impedes automatic conflict detection; existing approaches suffer from poor generalizability and heavy reliance on handcrafted rules. This paper proposes a two-stage automated conflict detection framework: first, candidate conflicting requirement pairs are retrieved using sentence embeddings (BERT or Doc2Vec) and cosine similarity; second, a supervised semantic similarity model precisely classifies true conflicts. To our knowledge, this is the first work to introduce supervised semantic matching for requirement conflict detection. We further propose UnSupCDA, an unsupervised variant enabling zero-shot deployment in label-scarce scenarios. Extensive experiments across five domain-specific software requirements specification (SRS) datasets demonstrate that our method significantly outperforms keyword- and rule-based baselines, achieving high accuracy and strong cross-domain generalizability.

Existing change impact analysis approaches rely solely on semantic similarity or structural dependencies, limiting their ability to comprehensively identify affected artifacts across heterogeneous software assets such as requirements, configurations, services, and tests. This work proposes a novel, training-free, and interpretable method that uniquely integrates semantic priors with multi-hop graph propagation. Specifically, it constructs a typed heterogeneous graph via static analysis, derives semantic priors from embedding-based cosine similarity, and diffuses impact through a row-normalized, decay-weighted propagation matrix controlled by a single parameter λ to balance precision and recall. Evaluation on five real-world change scenarios in a payment subsystem demonstrates the method’s capability to capture both structurally reachable yet textually disjoint artifacts and semantically related but structurally isolated ones, with demonstrated extensibility to operational assets such as container images and monitoring metrics.

This work addresses the challenge in distributed tracing diagnostics where anomalies often differ from normal behavior across multiple unknown dimensions—such as structure, timing, critical paths, and semantics—making them difficult to capture effectively through a single interface. To tackle this, the authors propose Contrast, a novel system that decouples trace representation from comparison semantics and introduces a composable Trace Projection Object (TPO) to uniformly encode multidimensional features. This design enables dynamic construction of arbitrary comparison sets at query time. Contrast further integrates two complementary interfaces: interactive visualization and large language model–generated explanations, facilitating flexible and in-depth root cause analysis. Experimental evaluation demonstrates that Contrast significantly improves both multidimensional comparison efficiency and diagnostic accuracy on the DeathStarBench benchmark and in Uber’s production environment.

This work addresses the limitation of existing code-centric unit test generation approaches, which often fail to detect logical bugs violating business semantics in enterprise software. To bridge this gap, the authors propose SeGa, a novel method that explicitly models business semantics by constructing a semantic knowledge base from requirement documents. SeGa extracts functional specifications and generates fine-grained business scenarios comprising preconditions, triggering actions, expected outcomes, and semantic constraints to guide large language models in producing highly targeted unit tests. Empirical evaluation on four industrial Go projects demonstrates that SeGa identifies 22–25 additional real-world bugs compared to state-of-the-art baselines, achieving a precision improvement of 26.9%–34.3%. Furthermore, it uncovered 16 previously unknown vulnerabilities across six production repositories, all of which have since been confirmed and fixed.

This work addresses the limited trust developers place in AI-generated bug reports due to their frequent lack of actionability and reproducibility. The authors propose a novel approach that integrates code coverage analysis with large language models (LLMs) to automatically detect defects in uncovered code regions and generate structured bug reports containing severity ratings, reproduction steps, and repair suggestions. A key innovation is an LLM-driven prioritization mechanism that substantially outperforms traditional rule-based methods. Evaluated on 13 Python projects, the method produced 10,467 reports; manual assessment of the top 130 revealed an 84.6% validity rate. Compared to CoverUp, it achieves higher defect validity (81.0% vs. 76.2%), a 50% improvement in P@3, and a 41% gain in mean reciprocal rank (MRR).

This study addresses the inefficiency of fault localization in industrial settings when relying solely on natural language defect reports, a task further hindered by the absence of practical approaches that operate without source code or execution traces. To tackle this challenge, the work formulates fault localization as a supervised text classification problem, leveraging only historical defect report texts to predict fault locations, thereby seamlessly integrating into existing maintenance workflows. Systematic experiments on five years of real-world data from ABB Robotics demonstrate that traditional machine learning models—such as logistic regression and random forests—combined with TF-IDF representations and data augmentation significantly outperform fine-tuned large language models like RoBERTa. These findings challenge the prevailing assumption that larger models inherently yield superior performance and substantiate the feasibility and effectiveness of a low-cost, scalable, purely text-driven AI solution for industrial fault localization.