Abstract
Following the rise and growing capabilities of LLMs, a friendly colleague recently questioned us about the relevance of continuing to write efficient algorithms and implementing them in frameworks, or even using formal concept analysis! We took him at his word, and since the topic is broad, we began by asking: can LLMs compute the full set of concepts of a formal context — a task known to be highly combinatorial? We explored this question using two approaches: the first ("direct computation") consisted of asking the LLMs directly for the list of concepts, while the second ("code generation") involved asking them to produce code that performs the computation, which we then tested. We selected a benchmark of formal contexts of various sizes, several LLM models, and multiple prompting strategies. In addition, to the non-negligible economic and environmental costs that must be considered, the results are not yet conclusive. Direct computation performs very poorly, with heavily degraded results for contexts with more than 15 objects and 15 attributes. Code generation yields better performance, but results remain imperfect, and computation times are significantly higher compared to the reference tool used for comparison (fca4j). While this doesn't predict the future, it offers a snapshot of the current state of progress, and the approach we implemented can be periodically re-evaluated. This study was conducted in collaboration with students, simulating the behavior of junior developers or non-expert users who may not have prior knowledge in Formal Concept Analysis or prompt engineering. Our objective was not only to evaluate the technical performance of LLMs, but also to explore their accessibility and usefulness in educational and low-expertise contexts. In this regard, the work aims to demonstrate whether LLMs can serve as a meaningful support tool even for those without strong computational backgrounds, thereby extending the benefits of FCA-based reasoning and concept structuring to a broader audience.