arXiv subjects

Harald C. Gall

Publications and source records attributed to Harald C. Gall.

14 recordsLinked to original sources

arXiv2026

Multi Language Models for On-the-Fly Syntax Highlighting

Syntax highlighting is a critical feature in modern software development environments, enhancing code readability and developer productivity. However, delivering accurate highlighting in real time remains challenging for online and web-based development tools due to strict time and memory constraints on backend services. These systems must serve highlights rapidly and frequently, even when code is partially valid or invalid. This has led to on-the-fly syntax highlighting, where visual annotations are generated just before content is served, often at high request rates and under incomplete input conditions. To meet these demands efficiently, state-of-the-art models use deep learning to learn the behavior of brute-force syntax highlighting resolvers, tools that are easy to implement but too slow for production. Through the Deep Abstraction process, brute-force strategies are encoded into fast statistical models that achieve both high accuracy and low-latency inference. Despite their success, such models face key challenges: they support only one programming language per model, require large datasets from slow brute-force generators, and involve resource-intensive training. In multi-language environments, this means maintaining multiple independent models, increasing system complexity and operational cost. This work addresses these issues by introducing a unified model capable of highlighting up to six mainstream programming languages, reducing deployment complexity by a factor of six and improving performance on unseen languages. A novel normalization technique significantly enhances model generalization, while few-shot learning experiments show that a small number of oracle samples can replace large datasets, minimizing dependence on brute-force generators. Combined, these innovations enable efficient, scalable, and cost-effective syntax highlighting across diverse programming languages.

cs.SE↗

arXiv2025

Towards Privacy-Preserving Code Generation: Differentially Private Code Language Models

Large language models specialized for code (CodeLLMs) have demonstrated remarkable capabilities in generating code snippets, documentation, and test cases. However, despite their promising capabilities, CodeLLMs can inadvertently memorize and reproduce snippets from their training data, which poses risks of privacy breaches and intellectual property violations. These risks restrict the deployment of CodeLLMs in sensitive domains and limit their training datasets to publicly available sources. To mitigate the memorization risk without compromising their task performance, we apply Differential Privacy (DP) to CodeLLMs. To the best of our knowledge, this is the first comprehensive study that systematically evaluates the effectiveness of DP in CodeLLMs. DP adds calibrated noise to the training process to protect individual data points while still allowing the model to learn useful patterns. To this end, we first identify and understand the driving reasons of the memorization behaviour of the CodeLLMs during their fine-tuning. Then, to address this issue, we empirically evaluate the effect of DP on mitigating memorization while preserving code generation capabilities. Our findings show that DP substantially reduces memorization in CodeLLMs across all the tested snippet types. The snippet types most prone to memorization are also the most effectively mitigated by DP. Furthermore, we observe that DP slightly increases perplexity but preserves, and can even enhance, the code generation capabilities of CodeLLMs, which makes it feasible to apply DP in practice without significantly compromising model utility. Finally, we analyze the impact of DP on training efficiency and energy consumption, finding that DP does not significantly affect training time or energy usage, making it a practical choice for privacy-preserving CodeLLMs training.

cs.SE↗

arXiv2025

Key-Augmented Neural Triggers for Knowledge Sharing

Repository-level code comprehension and knowledge sharing remain core challenges in software engineering. Large language models (LLMs) have shown promise by generating explanations of program structure and logic. However, these approaches still face limitations: First, relevant knowledge is distributed across multiple files within a repository, aka semantic fragmentation. Second, retrieval inefficiency and attention saturation degrade performance in RAG pipelines, where long, unaligned contexts overwhelm attention. Third, repository specific training data is scarce and often outdated. Finally, proprietary LLMs hinder industrial adoption due to privacy and deployment constraints. To address these issues, we propose Key-Augmented Neural Triggers (KANT), a novel approach that embeds knowledge anchors into both training and inference. Unlike prior methods, KANT enables internal access to repository specific knowledge, reducing fragmentation and grounding inference in localized context. Moreover, we synthesize specialized data directly from code. At inference, knowledge anchors replace verbose context, reducing token overhead and latency while supporting efficient, on premise deployment. We evaluate KANT via: a qualitative human evaluation of the synthesized dataset's intent coverage and quality across five dimensions; compare against SOTA baselines across five qualitative dimensions and inference speed; and replication across different LLMs to assess generalizability. Results show that the synthetic training data aligned with information-seeking needs. KANT achieved over 60% preference from human annotators and a LocalStack expert (preferring 79% of cases). Also, KANT reduced inference latency by up to 85% across all models. Overall, it is well-suited for scalable, low-latency, on-premise deployments, providing a strong foundation for code comprehension.

cs.SE↗

arXiv2025

Trustworthy Distributed Certification of Program Execution

Verifying the execution of a program is complicated and often limited by the inability to validate the code's correctness. It is a crucial aspect of scientific research, where it is needed to ensure the reproducibility and validity of experimental results. Similarly, in customer software testing, it is difficult for customers to verify that their specific program version was tested or executed at all. Existing state-of-the-art solutions, such as hardware-based approaches, constraint solvers, and verifiable computation systems, do not provide definitive proof of execution, which hinders reliable testing and analysis of program results. In this paper, we propose an innovative approach that combines a prototype programming language called Mona with a certification protocol OCCP to enable the distributed and decentralized re-execution of program segments. Our protocol allows for certification of program segments in a distributed, immutable, and trustworthy system without the need for naive re-execution, resulting in significant improvements in terms of time and computational resources used. We also explore the use of blockchain technology to manage the protocol workflow following other approaches in this space. Our approach offers a promising solution to the challenges of program execution verification and opens up opportunities for further research and development in this area. Our findings demonstrate the efficiency of our approach in reducing the number of program executions compared to existing state-of-the-art methods, thus improving the efficiency of certifying program executions.

cs.SE↗

arXiv2024

Anchor Attention, Small Cache: Code Generation with Large Language Models

The development of large language models (LLMs) has revolutionized automated code generation. However, their high demand of computation resources has hindered a broader deployment and raised environmental concerns. A common strategy for diminishing computational demands is to cache Key-Value (KV) states from the attention mechanism which is adopted predominately by mainstream LLMs. It can mitigate the need of repeated attention computations, but brings significant memory overhead. Current practices in NLP often use sparse attention which may, unfortunately, lead to substantial inaccuracies, or hallucinations, in code generation tasks. In this paper, we analyze the attention weights distribution within code generation models via an empirical study, uncovering a sparsity pattern, i.e., the aggregation of information at specific anchor points. Based on this observation, we propose a novel approach, AnchorCoder, which features token-wise anchor attention designed to extract and compress the contextual information, and layer-wise anchor attention enabling cross-layer communication to mitigate the issue of excessive superposition caused by the compression. The extensive experiments across multiple benchmark datasets confirm the effectiveness of AnchorCoder, which can consistently achieve a significant (at least 70%) reduction in KV cache requirements, while preserving the majority of model's performance.

cs.SE↗

arXiv2024

On-the-Fly Syntax Highlighting: Generalisation and Speed-ups

On-the-fly syntax highlighting is the task of rapidly associating visual secondary notation values with each character of a language derivation. Research in this domain is driven by the prevalence of online software development tools, which frequently display source code on screen and heavily rely on syntax highlighting mechanisms. In this context, three contrasting demands confront resolvers in this space: speed, accuracy, and development costs. Speed constraints are essential to ensure tool usability, manifesting as responsiveness for end users accessing online source code and minimising system overhead. Simultaneously, achieving precise highlighting is critical for enhancing code comprehensibility. Nevertheless, obtaining accurate results necessitates the capacity to perform grammatical analysis on the code under consideration, even in cases of varying grammatical correctness. Furthermore, addressing the development costs of such resolvers is imperative, given the multitude of programming language versions. The current state-of-the-art approach in this field leverages the original lexer and parser of programming languages to create syntax highlighting oracles, subsequently used for training base Recurrent Neural Network models. As the question of the generalisation of such a solution persists, this paper addresses this aspect by extending the original work to three additional mainstream programming languages and conducting a comprehensive review of the outcomes. Moreover, the original limitations in evaluation performance and training costs are mitigated through the introduction of a novel Convolutional based Neural Network model. This study examines the performance gains of running models on GPUs, finding that the new CNN implementation is much faster than previous methods while maintaining high accuracy.

cs.SE↗

arXiv2023

Towards Top-Down Automated Development in Limited Scopes: A Neuro-Symbolic Framework from Expressibles to Executables

Deep code generation is a topic of deep learning for software engineering (DL4SE), which adopts neural models to generate code for the intended functions. Since end-to-end neural methods lack domain knowledge and software hierarchy awareness, they tend to perform poorly w.r.t project-level tasks. To systematically explore the potential improvements of code generation, we let it participate in the whole top-down development from \emph{expressibles} to \emph{executables}, which is possible in limited scopes. In the process, it benefits from massive samples, features, and knowledge. As the foundation, we suggest building a taxonomy on code data, namely code taxonomy, leveraging the categorization of code information. Moreover, we introduce a three-layer semantic pyramid (SP) to associate text data and code data. It identifies the information of different abstraction levels, and thus introduces the domain knowledge on development and reveals the hierarchy of software. Furthermore, we propose a semantic pyramid framework (SPF) as the approach, focusing on software of high modularity and low complexity. SPF divides the code generation process into stages and reserves spots for potential interactions. In addition, we conceived preliminary applications in software development to confirm the neuro-symbolic framework.

cs.SE↗

arXiv2022

Continuous Deep Learning: A Workflow to Bring Models into Production

Researchers have been highly active to investigate the classical machine learning workflow and integrate best practices from the software engineering lifecycle. However, deep learning exhibits deviations that are not yet covered in this conceptual development process. This includes the requirement of dedicated hardware, dispensable feature engineering, extensive hyperparameter optimization, large-scale data management, and model compression to reduce size and inference latency. Individual problems of deep learning are under thorough examination, and numerous concepts and implementations have gained traction. Unfortunately, the complete end-to-end development process still remains unspecified. In this paper, we define a detailed deep learning workflow that incorporates the aforementioned characteristics on the baseline of the classical machine learning workflow. We further transferred the conceptual idea into practice by building a prototypic deep learning system using some of the latest technologies on the market. To examine the feasibility of the workflow, two use cases are applied to the prototype.

cs.SE↗

arXiv2022

Synthetic End-User Testing: Modeling Realistic Agents Based on Behavioral Examples

For software interacting directly with real-world end-users, it is common practice to script scenario tests validating the system's compliance with a number of its features. However, these do not accommodate the replication of the type of end-user activity to which the system is required to respond in a live instance. It is especially true as compliance might also break in scenarios of interactions with external events or processes, such as other users. State-of-the-art approaches aim at inducing the software into runtime errors by generating tests that maximize some target metrics, such as code coverage. As a result, they suffer from targeting an infinitely large search space, are severely limited in recognizing error states that do not result in runtime errors, and the test cases they generate are often challenging to interpret. Other forms of testing, such as Record-Replay, instead fail to capture the end-users' decision-making process, hence producing largely scripted test scenarios. Therefore, it is impossible to test a software's compliance with unknown but otherwise plausible states. This paper introduces "Synthetic End-User Testing," a novel testing strategy for complex systems in which real-world users are synthesized into reusable agents and employed to test and validate the software in a simulation environment. Hence, it discusses how end-user behavioral examples can be obtained and used to create agents that operate the target software in a reduced search space of likely action sequences. The notion of action expectation, which allows agents to assert the learned compliance of the system, is also introduced. Finally, a prototype asserting the feasibility of such a strategy is presented.

cs.SE↗

arXiv2022

On the Effectiveness of Transfer Learning for Code Search

The Transformer architecture and transfer learning have marked a quantum leap in natural language processing, improving the state of the art across a range of text-based tasks. This paper examines how these advancements can be applied to and improve code search. To this end, we pre-train a BERT-based model on combinations of natural language and source code data and fine-tune it on pairs of StackOverflow question titles and code answers. Our results show that the pre-trained models consistently outperform the models that were not pre-trained. In cases where the model was pre-trained on natural language "and" source code data, it also outperforms an information retrieval baseline based on Lucene. Also, we demonstrated that the combined use of an information retrieval-based approach followed by a Transformer leads to the best results overall, especially when searching into a large search pool. Transfer learning is particularly effective when much pre-training data is available and fine-tuning data is limited. We demonstrate that natural language processing models based on the Transformer architecture can be directly applied to source code analysis tasks, such as code search. With the development of Transformer models designed more specifically for dealing with source code data, we believe the results of source code analysis tasks can be further improved.

cs.SE↗

arXiv2022

On-the-Fly Syntax Highlighting using Neural Networks

With the presence of online collaborative tools for software developers, source code is shared and consulted frequently, from code viewers to merge requests and code snippets. Typically, code highlighting quality in such scenarios is sacrificed in favor of system responsiveness. In these on-the-fly settings, performing a formal grammatical analysis of the source code is not only expensive, but also intractable for the many times the input is an invalid derivation of the language. Indeed, current popular highlighters heavily rely on a system of regular expressions, typically far from the specification of the language's lexer. Due to their complexity, regular expressions need to be periodically updated as more feedback is collected from the users and their design unwelcome the detection of more complex language formations. This paper delivers a deep learning-based approach suitable for on-the-fly grammatical code highlighting of correct and incorrect language derivations, such as code viewers and snippets. It focuses on alleviating the burden on the developers, who can reuse the language's parsing strategy to produce the desired highlighting specification. Moreover, this approach is compared to nowadays online syntax highlighting tools and formal methods in terms of accuracy and execution time, across different levels of grammatical coverage, for three mainstream programming languages. The results obtained show how the proposed approach can consistently achieve near-perfect accuracy in its predictions, thereby outperforming regular expression-based strategies.

cs.SE↗

arXiv2022

Toward Granular Automatic Unit Test Case Generation

Unit testing verifies the presence of faults in individual software components. Previous research has been targeting the automatic generation of unit tests through the adoption of random or search-based algorithms. Despite their effectiveness, these approaches do not implement any strategy that allows them to create unit tests in a structured manner: indeed, they aim at creating tests by optimizing metrics like code coverage without ensuring that the resulting tests follow good design principles. In order to structure the automatic test case generation process, we propose a two-step systematic approach to the generation of unit tests: we first force search-based algorithms to create tests that cover individual methods of the production code, hence implementing the so-called intra-method tests; then, we relax the constraints to enable the creation of intra-class tests that target the interactions among production code methods.

cs.SE↗

arXiv2022

Assemble Foundation Models for Automatic Code Summarization

Automatic code summarization is beneficial to daily software development since it could help reduce the requirement of manual writing. Currently, artificial intelligence is undergoing a paradigm shift. The foundation models pretrained on massive data and finetuned to downstream tasks surpass specially customized models. This trend inspired us to consider reusing foundation models instead of learning from scratch. Thereby, we propose a flexible and robust approach for automatic code summarization, based on neural models. We assemble available foundation models, such as CodeBERT and GPT-2, into a single neural model named AdaMo. Moreover, we utilize Gaussian noise as the simulation of contextual information to optimize the latent representation. Furthermore, we introduce two adaptive schemes from the perspective of knowledge transfer, namely continuous pretraining and intermediate finetuning, and design intermediate stage tasks for general sequence-to-sequence learning. Finally, we evaluate AdaMo against a benchmark dataset for code summarization, by comparing it with state-of-the-art models.

cs.SE↗

arXiv2015

The Making of Cloud Applications An Empirical Study on Software Development for the Cloud

Cloud computing is gaining more and more traction as a deployment and provisioning model for software. While a large body of research already covers how to optimally operate a cloud system, we still lack insights into how professional software engineers actually use clouds, and how the cloud impacts development practices. This paper reports on the first systematic study on how software developers build applications in the cloud. We conducted a mixed-method study, consisting of qualitative interviews of 25 professional developers and a quantitative survey with 294 responses. Our results show that adopting the cloud has a profound impact throughout the software development process, as well as on how developers utilize tools and data in their daily work. Among other things, we found that (1) developers need better means to anticipate runtime problems and rigorously define metrics for improved fault localization and (2) the cloud offers an abundance of operational data, however, developers still often rely on their experience and intuition rather than utilizing metrics. From our findings, we extracted a set of guidelines for cloud development and identified challenges for researchers and tool vendors.

cs.SE↗