Exploring Open-ended Neural Concept Representation for Language Emergence

Humans can acquire concepts and skills in a modular and, to a certain extent, independent way. For example, we can learn to recognise penguins and later learn to draw them. If we want to do the same with monkeys, we will not need to learn drawing from scratch. This modularity has been central in the opposition of different methodologies in the field of language emergence. The language game paradigm has historically proposed experiments in which categories – or concepts – are created in an open-ended and explainable way (Steels, 1995; Steels & Belpaeme, 2005). On the other hand, the past decade has seen a wave of experiments using neural techniques to tackle complex problems in the real world (Rita et al., 2025), like puzzle-solving (Foerster et al., 2016) or reference (Lazaridou et al., 2017), but most architectures proposed end-to-end systems, where skills and concepts are indistinguishable. Our goal is to progress towards modular and transferable concept representations by proposing an open-ended architecture that also leverages the complexity that neural networks can capture.