Before artificial intelligence can recognize an image, understand a text, or answer a question, it must first be trained. To do this, developers and data scientists use specialized software libraries—true “toolboxes” that enable the construction of these AI models.

A research team comprising the University of Namur, the University of Passau (Germany), and TU Delft (Netherlands) has been studying this software. Their goal: to detect bugs that are particularly difficult to spot and that could cause a program to crash suddenly.

To verify that these libraries are functioning properly, researchers use software capable of automatically generating thousands of tests. The purpose of these tests is to push the programs to their limits in order to identify any potential errors. 

Until now, however, these tools had a major limitation: whenever a test caused a software library to malfunction, the software responsible for running the tests would also stop working. This made it impossible to pinpoint the exact source of the problem or report it to the developers. 

Researchers at UNamur have therefore devised a new method. Now, each test is run separately. If one of them causes a crash, it is isolated: the testing software continues to run, logs the error, and carries on with its analysis. This makes it possible to automatically detect bugs that previously went unnoticed. 

Using this new method, researchers analyzed 1,648 modules from the leading machine learning libraries and uncovered 32 previously unknown bugs. Several of these affected software that has become indispensable for training artificial intelligence models.

These issues were reported to the development teams. Several have already been acknowledged and fixed. “When a bug is fixed at this level, potentially thousands, or even millions, of users benefit,” notes Xavier Devroey, a professor in the Faculty of Computer Science at UNamur and a member of the PReCISE research center at the Namur Digital Institute (NaDI). Beyond these fixes, this research helps make the software on which so many artificial intelligence applications rely today more reliable.

This breakthrough originated from Lucas Berg’s master’s thesis, completed at UNamur under the supervision of Professor Xavier Devroey. As part of his research internship, Lucas Berg then joined Professor Annibale Panichella’s team at TU Delft to work on Pynguin, the open-source automated test generation tool developed at the University of Passau by Professor Gordon Fraser’s team, mentioned earlier. 

Initially, his goal was to improve this tool. By preventing crashes while discovering bugs, this first contribution subsequently caught the attention of researchers at the University of Passau, who were facing the same problem, leading to an international collaboration that would result in a joint scientific publication. 

Lucas Berg is currently pursuing a Ph.D. at the Faculty of Computer Science at UNamur under the supervision of Professors Wim Vanhoof and Xavier Devroey. The results of this research were presented at the IEEE International Conference on Software Testing, Verification, and Validation (ICST 2026), one of the most prestigious international conferences in the field of software testing.  

This research also received support from the Wallonia Public Service (SPW Research) and Wallonia-Brussels International. This recognition underscores the quality of this work and its significance to the international scientific community. 

At NaDI, researchers develop innovative solutions to the new societal challenges posed by the digital revolution (eGov, eHealth, eServices, big data, etc.). Drawing from a variety of disciplines, the researchers combine their expertise in computer science, technology, ethics, law, management, and sociology. Comprising six research centers, the Namur Digital Institute offers unique multidisciplinary expertise in all areas of computer science, its applications, and its social impact.  

Missing components, fragile solder joints, reversed polarity, scratches, or foreign objects such as a screw that falls onto the product: on a production line for automotive electronic components, such as the one at AISIN Europe, defects—even if they are rare—must be detected immediately and accurately. “My goal was to teach a computer to distinguish between what is normal and what is abnormal in images,” explains Arnaud Bougaham, a data scientist at AISIN Europe and an associate researcher at the Faculty of Computer Science at UNamur. 

On a production line, images are captured by cameras positioned above a conveyor belt, and a computer system is then tasked with verifying, at a high rate, that each part meets specifications. “However, while the traditional inspection system does its job well, it also generates a lot of false positives—false alerts. This has a significant impact on the work of the people responsible for confirming or dismissing the alerts: they are constantly called upon, at the risk of wasting time and energy.” It is this process that Arnaud Bougaham decided to improve for his company by beginning a doctoral thesis in 2020 at the HuMaLearn laboratory in the School of Computer Science—a lab recognized for its expertise in machine learning—under the supervision of Professors Benoit Frenay and Isabelle Linden.

In addition, when a defect is suspected, the model can precisely pinpoint the affected area: the operator can view the indication, manipulate the part if necessary, and then make a decision. 

Another benefit of the tool? The system doesn’t just display a verdict; it also indicates the level of confidence the system has in its prediction. “The goal is to filter out simple cases and draw attention to more complex ones, so that experts are no longer overwhelmed by incorrect diagnoses and can focus on what really matters,” emphasizes Arnaud Bougaham.

While the tool was primarily designed to meet the needs of the automotive manufacturing industry, its underlying logic is generic. “It could be adapted for use in factories that manufacture hardwood flooring, wood, or fabric. But also in the medical sector,” says Arnaud Bougaham enthusiastically. “A disease is, ultimately, an anomaly in an image. The challenges are similar: when there’s a defect, it’s essential to detect it while minimizing false positives as much as possible. In the medical sector, having a similar tool that streamlines the analysis while leaving the final decision to the doctor can make a real difference.”

Arnaud Bougaham’s dissertation grew out of a dual role: as a doctoral student at UNamur and, at the same time, as an engineer at AISIN Europe. This position gave him access to what he calls “the reality on the ground.” “In the lab, you can get very good results, but in production, there’s a reality with a specific environment.” Noisy data, variable conditions, cost constraints, and above all, production pace: “Our system has to operate at the same pace as the production line.” By dividing his time between AISIN Europe and UNamur, Arnaud Bougaham demonstrates how a dissertation can have a tangible impact on real-world needs—all while balancing the operations and obligations of both organizations. 

Beyond the promises, this research is already being applied to real-world use cases. At AISIN Europe, the framework developed is currently being integrated into an active production line to detect unexpected components (such as a screw) on printed circuit boards. In the medical field, the approach is geared toward analyzing responsiveness in comatose patients through medical imaging that allows for the observation of organ and tissue function. Certain techniques are also being adapted for applications involving the segmentation of ovarian cancer or lymphoma.

These results also demonstrate how artificial intelligence can be reliable and human-centered in high-risk environments.

Artificial intelligence is now accessible to everyone. Yet, while many people use it every day, few truly understand how it works, its limitations, or the issues surrounding it. To address this need, the Namur Digital Institute (NaDI) is launching “What If AI…,” a series of lecture-workshops open to the general public, coordinated by doctoral students Elisa De Coster (PReCISE) and Aline Nardi (CRIDS). 

“The idea behind the project is to demystify artificial intelligence,” begins Elisa De Coster. “As researchers, we’re fortunate to be aware of the issues raised by AI. We wanted curious members of the public to have access to these keys to understanding as well.” 

“Before 2022 and the rise of ChatGPT, these tools were limited to a niche audience,” continues Aline Nardi. “Since then, the use of AI has become widespread at record speed, but reflection on the societal implications of these uses has not kept pace. This series of lectures is an invitation to take a step back and understand what’s happening behind the scenes.” 

Each session, lasting about two hours, will bring together the perspectives of several specialists. A doctoral student in computer science will explain the technical workings of AI, while his counterpart in law will address the legal, ethical, and societal implications.

Speakers from the professional world will share their real-world experience, and an interactive activity will allow the audience to put these concepts into practice. Interactivity, interdisciplinarity, and accessibility to the general public are the three pillars of the project.

Each event will address a different social issue: 

• October 1, 2026 – La Nef
  What if AI were your confidant?
  Can you confide in a chatbot? What are the risks? 
• October 22, 2026 – PA02
  What if AI were wrong?
  How can we recognize errors in artificial intelligence and maintain our critical thinking? 
• November 19, 2026 – Caméo
  What if AI were subjective? Is artificial intelligence
  truly neutral, or does it perpetuate certain biases? 
• February 25, 2027 – Quai 22
  What if AI stole your creations?
  What challenges does generative AI pose in terms of copyright and intellectual property? 
• March 18, 2027 – TRAKK
  What if AI knew you too well?
  What happens to our personal data, and to what extent can AI learn to know us? 

While Elisa De Coster and Aline Nardi coordinate the project, they emphasize that it is, above all, a collaborative effort. The lecture series is led by ten doctoral students from the PReCISE and CRIDS research centers: Elisa De Coster, Adélaïde Couplet, Jules Dejaeghere, Pierre Luycx, Pierre Poitier, Aline Nardi, Alix Gobert, Florian Jacques, Martin Rappe, Lionel Goffaux, and Olga Thiry. Together, they aim to help the general public better understand artificial intelligence, without demonizing it or presenting it as a miracle solution.  

The study was led by Professor Thomas Brochhagen, a researcher in the Computational Linguistics and Linguistic Theory (COLT) group within the Department of Translation and Language Sciences at UPF Barcelona. Co-authors include Xixian Liao (Barcelona Supercomputing Center - Centro Nacional de Supercomputación, BSC-CNS), Jamie D. Wright (Namur Digital Institute, University of Namur and Vrije Universiteit Brussel) and Carmen Saldana (affiliated with UPF and the University of Barcelona).

“My main contribution to the Nature Human Behaviour study was in data analysis and computational modelling. The study is a large-scale, cross-linguistic investigation of how communicative pressures shape language evolution. By analysing data from nearly 2,000 languages, we examined how languages balance two competing pressures: the efficiency gained by reusing the same or similar word forms and the need to avoid ambiguity when concepts could be confused in communication”, explains Jamie Wright.

Jamie Wright’s work with Prof. Katrien Beuls at the Namur Digital Institute forms part of the HERMES programme. This research also examines the communicative pressures that shape language evolution, but from a complementary perspective. 

Researchers from Namur have just published another article in the “Journal of Language Evolution”

“While the Nature Human Behaviour article identifies broad patterns across a large sample of existing languages, our recent article in the Journal of Language Evolution uses an agent-based model to study language evolution at the level of interacting agents. In these simulations, artificial agents develop and use shared languages through repeated communicative exchanges. This allows us to investigate how population-level linguistic conventions emerge from local interactions and why such conventions can make languages more robust, easier to learn and more cognitively efficient.” 

The two studies therefore address closely related questions at different levels and with different methodological approaches. The Nature Human Behaviour study shows the large-scale cross-linguistic consequences of communicative pressures, while the Journal of Language Evolution agent-based study helps explain how such patterns can emerge dynamically through interactions between language users.

 He completed his Master's in Theoretical and Applied Linguistics at Pompeu Fabra University, Barcelona, where he also worked as a research assistant within the COLT research group. His research uses computational modelling and agent-based experiments to investigate language emergence and evolution, with his current work examining how articulatory and acoustic constraints on the production and perception of speech shape linguistic conventions and systems of communication.

Congratulations to the researchers for these publications!

Cited article

Brochhagen, T., Liao, X., Wright, J.D., et al. The interaction of meaning similarity and confusability explains regularity in form–meaning mappings at and below the word level. Nat Hum Behav (2026). https://doi.org/10.1038/s41562-026-02488-3