Mathieu Acher

Mathieu Acher, Jean-Marc Jézéquel

Frontier coding agents now promise end-to-end authorship of complete software systems. Two empirical questions follow: can AI coding-agent teammates program in any target language, including ones with no comparable prior open-source artefact? If so, does language choice still shape the artefact, and along which dimensions? We study both through a polyglot case study built around chess engines: non-trivial multi-component systems that admit a hierarchy of language-agnostic oracles, from exact move-generation correctness to a strength scale (Elo), observable from Rust to Brainfuck. We prompted two frontier agents (Claude Code and Codex) at the capability level, without chess knowledge or implementation guidance, under a documented intervention and stopping policy. The agents produced 34 chess engines spanning 17 primary programming languages, from mainstream to specialised, domain-specific, legacy, and esoteric targets. We combine per-engine feature analysis, independent Elo assessment, and session trajectories with qualitative analysis of code and transcripts. Frontier coding agents are genuinely polyglot: every language we tried produced at least one feature-rich working engine, several with no prior open-source counterpart of comparable scope (e.g., LaTeX), and the code is synthesised from scratch rather than copied. Yet language choice still matters: strong playing strength is only reachable in mainstream compiled languages, cost and engineering effort grow sharply as the language becomes more exotic, and feature choices shift across language families. Agents validate their own work unprompted, but their strength self-estimates are biased and a few engines cheated by calling a chess library. Programming language is no longer about whether AI teammates can build a working system, but about performance, cost, what gets built, and how much human supervision validation still needs.

Xhevahire Tërnava, Georges Aaron Randrianaina, Luc Lesoil et al.

Many small-scale software systems, that is, with limited codebase or binary size, are widely used in everyday tasks, yet their configurability remains largely unexplored. At the same time, studies on modern software systems show a trend toward increasing configurability, alongside growing interest in building immutable, specialized, and reproducible software. In this paper, we present the first empirical study on the extent of configurability in small-scale software systems. By analyzing 108 programs from GNU coreutils, we show that even small programs can exhibit significant compile-time and run-time variability, with up to 76 options per program. Then, there is a high correlation (0.78) between run-time variability and codebase size. Furthermore, an analysis of the 20 smallest programs across 85 releases reveals that variability tends to increase over time, primarily due to the added compile-time variability. This suggests that shifting options between run-time and compile-time, removing unnecessary run-time variability, or resolving compile-time variability early, can help reduce codebase complexity and size. We also introduce, for the first time, the concept of null-variable software system, one with no configurability beyond mandatory features. Our findings show that high configurability is not exclusive to large-scale systems and that reducing unnecessary variability can lead to lightweight, smaller, and more maintainable software. We hope this effort contributes to designing new software by understanding how to balance its configurability with codebase size.

Axel Halin, Alexandre Nuttinck, Mathieu Acher et al.

Many approaches for testing configurable software systems start from the same assumption: it is impossible to test all configurations. This motivated the definition of variability-aware abstractions and sampling techniques to cope with large configuration spaces. Yet, there is no theoretical barrier that prevents the exhaustive testing of all configurations by simply enumerating them, if the effort required to do so remains acceptable. Not only this: we believe there is lots to be learned by systematically and exhaustively testing a configurable system. In this case study, we report on the first ever endeavour to test all possible configurations of an industry-strength, open source configurable software system, JHipster, a popular code generator for web applications. We built a testing scaffold for the 26,000+ configurations of JHipster using a cluster of 80 machines during 4 nights for a total of 4,376 hours (182 days) CPU time. We find that 35.70% configurations fail and we identify the feature interactions that cause the errors. We show that sampling strategies (like dissimilarity and 2-wise): (1) are more effective to find faults than the 12 default configurations used in the JHipster continuous integration; (2) can be too costly and exceed the available testing budget. We cross this quantitative analysis with the qualitative assessment of JHipster's lead developers.

Charly Reux, Mathieu Acher, Djamel Eddine Khelladi et al.

With the rise of AI-based code generation, customizing existing code out of natural language instructions to modify visual results -such as figures or images -has become possible, promising to reduce the need for deep programming expertise. However, even experienced developers can struggle with this task, as it requires identifying relevant code regions (feature location), generating valid code variants, and ensuring the modifications reliably align with user intent. In this paper, we introduce vTikZ, the first benchmark designed to evaluate the ability of Large Language Models (LLMs) to customize code while preserving coherent visual outcomes. Our benchmark consists of carefully curated vTikZ editing scenarios, parameterized ground truths, and a reviewing tool that leverages visual feedback to assess correctness. Empirical evaluation with stateof-the-art LLMs shows that existing solutions struggle to reliably modify code in alignment with visual intent, highlighting a gap in current AI-assisted code editing approaches. We argue that vTikZ opens new research directions for integrating LLMs with visual feedback mechanisms to improve code customization tasks in various domains beyond TikZ, including image processing, art creation, Web design, and 3D modeling.

Djamel Eddine Khelladi, Charly Reux, Mathieu Acher

Large language model (LLM)-based test generation has gained attention in software engineering, yet most studies evaluate LLMs' ability to generate unit tests in a single attempt for a given language, missing the opportunity to leverage LLM diversity for more robust testing. This paper introduces PolyTest, a novel approach that enhances test generation by exploiting polyglot and temperature-controlled diversity. PolyTest systematically leverages these properties in two complementary ways: (1) Cross-lingual test generation, where tests are generated in multiple languages at zero temperature and then unified; (2) Diverse test sampling, where multiple test sets are generated within the same language at a higher temperature before unification. A key insight is that LLMs can generate diverse yet contradicting tests -- same input, different expected outputs -- across languages and generations. PolyTest mitigates inconsistencies by unifying test sets, fostering self-consistency and improving overall test quality. Unlike single-language or single-attempt approaches, PolyTest enhances testing without requiring on-the-fly execution, making it particularly beneficial for weaker-performing languages. We evaluate PolyTest on Llama3-70B, GPT-4o, and GPT-3.5 using EvalPlus, generating tests in five languages (Java, C, Python, JavaScript, and a CSV-based format) at temperature 0 and sampling multiple sets at temperature 1. We observe that LLMs frequently generate contradicting tests across settings, and that PolyTest significantly improves test quality across all considered metrics -- number of tests, passing rate, statement/branch coverage (up to +9.01%), and mutation score (up to +11.23%). Finally, PolyTest outperforms Pynguin in test generation, passing rate, and mutation score.

Luc Lesoil, Mathieu Acher, Arnaud Blouin et al.

Widely used software systems such as video encoders are by necessity highly configurable, with hundreds or even thousands of options to choose from. Their users often have a hard time finding suitable values for these options (i.e. finding a proper configuration of the software system) to meet their goals for the tasks at hand, e.g. compress a video down to a certain size. One dimension of the problem is of course that performance depends on the input data: a video as input to an encoder like x264 or a file system fed to a tool like xz. To achieve good performance, users should therefore take into account both dimensions of (1) software variability and (2) input data. In this problem-statement paper, we conduct a large study over 8 configurable systems that quantifies the existing interactions between input data and configurations of software systems. The results exhibit that (1) inputs fed to software systems interact with their configuration options in non monotonous ways, significantly impacting their performance properties (2) tuning a software system for its input data makes it possible to multiply its performance by up to ten (3) input variability can jeopardize the relevance of performance predictive models for a field deployment.

Paul Temple, Mathieu Acher, Gilles Perrouin et al.

Software product line (SPL) engineers put a lot of effort to ensure that, through the setting of a large number of possible configuration options, products are acceptable and well-tailored to customers' needs. Unfortunately, options and their mutual interactions create a huge configuration space which is intractable to exhaustively explore. Instead of testing all products, machine learning techniques are increasingly employed to approximate the set of acceptable products out of a small training sample of configurations. Machine learning (ML) techniques can refine a software product line through learned constraints and a priori prevent non-acceptable products to be derived. In this paper, we use adversarial ML techniques to generate adversarial configurations fooling ML classifiers and pinpoint incorrect classifications of products (videos) derived from an industrial video generator. Our attacks yield (up to) a 100% misclassification rate and a drop in accuracy of 5%. We discuss the implications these results have on SPL quality assurance.

Juliana Alves Pereira, Hugo Martin, Mathieu Acher et al.

Most modern software systems (operating systems like Linux or Android, Web browsers like Firefox or Chrome, video encoders like ffmpeg, x264 or VLC, mobile and cloud applications, etc.) are highly-configurable. Hundreds of configuration options, features, or plugins can be combined, each potentially with distinct functionality and effects on execution time, security, energy consumption, etc. Due to the combinatorial explosion and the cost of executing software, it is quickly impossible to exhaustively explore the whole configuration space. Hence, numerous works have investigated the idea of learning it from a small sample of configurations' measurements. The pattern "sampling, measuring, learning" has emerged in the literature, with several practical interests for both software developers and end-users of configurable systems. In this survey, we report on the different application objectives (e.g., performance prediction, configuration optimization, constraint mining), use-cases, targeted software systems and application domains. We review the various strategies employed to gather a representative and cost-effective sample. We describe automated software techniques used to measure functional and non-functional properties of configurations. We classify machine learning algorithms and how they relate to the pursued application. Finally, we also describe how researchers evaluate the quality of the learning process. The findings from this systematic review show that the potential application objective is important; there are a vast number of case studies reported in the literature from the basis of several domains and software systems. Yet, the huge variant space of configurable systems is still challenging and calls to further investigate the synergies between artificial intelligence and software engineering.

Paul Temple, Mathieu Acher, Battista Biggio et al.

Ensuring that all supposedly valid configurations of a software product line (SPL) lead to well-formed and acceptable products is challenging since it is most of the time impractical to enumerate and test all individual products of an SPL. Machine learning classifiers have been recently used to predict the acceptability of products associated with unseen configurations. For some configurations, a tiny change in their feature values can make them pass from acceptable to non-acceptable regarding users' requirements and vice-versa. In this paper, we introduce the idea of leveraging these specific configurations and their positions in the feature space to improve the classifier and therefore the engineering of an SPL. Starting from a variability model, we propose to use Adversarial Machine Learning techniques to create new, adversarial configurations out of already known configurations by modifying their feature values. Using an industrial video generator we show how adversarial configurations can improve not only the classifier, but also the variability model, the variability implementation, and the testing oracle.

Mathieu Acher, François Esnault

The strength of chess engines together with the availability of numerous chess games have attracted the attention of chess players, data scientists, and researchers during the last decades. State-of-the-art engines now provide an authoritative judgement that can be used in many applications like cheating detection, intrinsic ratings computation, skill assessment, or the study of human decision-making. A key issue for the research community is to gather a large dataset of chess games together with the judgement of chess engines. Unfortunately the analysis of each move takes lots of times. In this paper, we report our effort to analyse almost 5 millions chess games with a computing grid. During summer 2015, we processed 270 millions unique played positions using the Stockfish engine with a quite high depth (20). We populated a database of 1+ tera-octets of chess evaluations, representing an estimated time of 50 years of computation on a single machine. Our effort is a first step towards the replication of research results, the supply of open data and procedures for exploring new directions, and the investigation of software engineering/scalability issues when computing billions of moves.

Guillaume Bécan, Razieh Behjati, Arnaud Gotlieb et al.

Feature modeling is a widely used formalism to characterize a set of products (also called configurations). As a manual elaboration is a long and arduous task, numerous techniques have been proposed to reverse engineer feature models from various kinds of artefacts. But none of them synthesize feature attributes (or constraints over attributes) despite the practical relevance of attributes for documenting the different values across a range of products. In this report, we develop an algorithm for synthesizing attributed feature models given a set of product descriptions. We present sound, complete, and parametrizable techniques for computing all possible hierarchies, feature groups, placements of feature attributes, domain values, and constraints. We perform a complexity analysis w.r.t. number of features, attributes, configurations, and domain size. We also evaluate the scalability of our synthesis procedure using randomized configuration matrices. This report is a first step that aims to describe the foundations for synthesizing attributed feature models.