Neural Collapse predicts that balanced one-hot classification pushes model representations to be equally far from each other; a symmetric configuration that depends only on the output label and ignores any semantic similarity in the inputs. This creates a puzzle: next-token prediction language models are trained predominantly (as context length increases) with one-hot labels: the same context is very unlikely to appear twice in training with different labels. However, they clearly learn latent structural features. That is, despite the one-hot training regime, a language model's contextual embeddings represent the fact that the next word in ''Mary broke the ___'' is likely to be filled by tokens in the latent classes of a) medium-sized, b) rigid, c) inanimate nouns. How does gradient descent find such categorical semantic structure when co-occurrence statistics collapse to one-hot sparsity, eliminating any shared next-tokens among different contexts? To investigate this tension we identify three synthetic controlled settings where inputs have latent semantic factors but are mapped to distinct one-hot labels. We find that semantic geometry emerges early in training, and that representations cluster by shared attributes despite receiving no explicit supervision to do so. This structure is transient: with sufficient capacity and time, the model eventually reaches the predicted symmetric state where all representations are equally separated. We study this phase transition through Gram matrix analysis and propose a preliminary modification to the commonly used unconstrained features model to capture the emergent semantic geometry.
This work examines the semantic geometry underlying NLP models. We compare supervised vector embeddings, such as CamemBERT, with lexical co-occurrence graphs that encode semantic relations more directly. While transformer-based embeddings achieve strong performance, their induced geometries often display unsatisfactory distributions. In contrast, graph-based models reveal a clearer and more human-readable organization of meaning. We have implemented a methodology that allows us to perform a comparative analysis either based on the structure of the graphs or based on the topology of the embeddings induced by these two approaches. The results of the comparison -- applied to the French "Great National Debate" corpus a collection of citizen contributions to the public debate -- show a similar local topology but a very different overall structure and topology. Theses findings suggest complementary perspectives between deep supervised models and graph-based models, considering a new pathway to guide neural architectures toward more stable and interpretable convergence with graphs structures.