Large Language Models (LLMs) are increasingly utilised in software development, particularly for automatic code generation. Game development, a complex and multidisciplinary field, has begun leveraging LLMs to streamline various stages of the development process. This paper focuses on the application of LLMs in automating code generation for game development, providing a comprehensive survey of existing research from multiple perspectives. We analyse the benefits and limitations of LLM-based approaches in this domain and identify key challenges. Our systematic analysis of 15 primary studies reveals a prevalence of general-purpose LLMs (47%) alongside fine-tuned models (40%), with prompt engineering universally adopted (100%) to guide code generation. Our findings highlight the potential of LLMs to enhance game development workflows, while outlining future research directions to address existing limitations.
The growing interest in applying Artificial Intelligence (AI) to software engineering has accelerated since the release of ChatGPT 3.5 in 2022. This paper investigates how Large Language Models (LLMs) support applications' modular development and testing. We introduce AIPyCraft, a novel AI-assisted framework that facilitates the end-to-end lifecycle of software projects. Our approach leverages Google Gemini 2.5 Pro model to generate, correct, and manage software components within a semi-automated and incremental workflow. AIPyCraft enables project creation, environment setup, error correction, and feature evolution in an integrated manner. We develop and test a blockchain-based Oracle component designed for 6G wireless network environments, i.e., a complex, real-world scenario that demands secure data integration and modular extensibility. Preliminary experiments demonstrate AIPyCraft's potential to accelerate small-scale software project development through an "understand-by-building" methodology. Our findings show that using an LLM to generate effective TOML jobs for Off-chain 6G functions is feasible, with an average of 1.05 iterations to correct the TOML code and mean experiment time of 27.8 seconds.
The Object Constraint Language (OCL) is extensively used in model-driven engineering to specify constraints on UML class diagrams. Evaluating the complexity of OCL expressions remains a challenge, particularly in the context of verification. This paper introduces a complete new set of metrics for measuring OCL expressions. This gives us the advantage of measuring OCL in a customizable way. First, we enhance the Structural Complexity. Second, we propose two new types of metric: Computational and Dependency Complexity. Finally, we calculate the overall complexity using our Total Complexity metric. Our experimental results demonstrate how one can measure the complexity of an OCL benchmark in terms of its verification time which is used to define the weights for OCL constructs.
In agile software development, it is common to employ user stories to capture requirements. Specifying requirements in structured natural language has the advantage that requirements are easily understood by domain experts. As requirements evolve and become more complex over time, their analysis also becomes more difficult. Requirement specifications in the form of knowledge graphs have been proven to be useful to partially automate this analysis and make it more manageable. There are related works that automate the translation of user stories into knowledge graph representations, making the previous manual translation less error-prone and more efficient. A recent approach of Arulmohan et al. employs large language models (LLMs) to automate the translation and compares it with alternatives based on dedicated Natural Language Processing (NLP). A large experiment revealed that the latter outperformed the LLM-based solution. Because of the stochastic nature of LLMs, the fact that they evolve over time, and the availability of different LLM providers (i.e., organizations that provide access to an LLM such as OpenAI (GPT) and Meta (LLaMA)), the same experiment run today or with another LLM would generate different results. For this reason, in this paper we present an end-to-end automated approach for an LLM-based translation that is provider-agnostic and can be easily (re)evaluated against a given ground truth. We explain the design and implementation of our solution based on the LangChain framework. We, moreover, present an evaluation script and we report on experiments that we have performed using the script with different LLM providers. We could show that some of the LLMs are indeed able to close the gap compared to a dedicated NLP approach. We conclude the paper with a discussion of the consequences of LLM-based automated requirements processing.
This paper documents the author's participation in the 2025 Transformation Tool Contest (TTC) Live Contest, which focused on the automatic translation of feature models specified in the Universal Variability Language (UVL) into the Graphviz DOT format. The challenge was approached through the development of a minimalist domain-specific language workbench in F#, reflecting the combined influence of software language engineering expertise and an emerging software development methodology called "expert vibing" - the judicious and unapologetic use of support that generative artificial intelligence can provide. The resulting solution demonstrates that concise and performant model transformations can be achieved outside conventional frameworks by leveraging both custom lightweight languages and AI-assisted coding practices. We present Indentia and Scripta languages, discuss practical and theoretical trade-offs compared to established tools, and empirically evaluate correctness, conciseness, and performance against the reference solution. The findings highlight both the potential and current limitations of large language models for model transformation tasks, and provide insight into the evolving workflow of vibe coding in software engineering.
