Quantifying LLM Contextual Embedding Trajectories via Incremental Displacement

Researchers quantify how token-wise contextual word embeddings evolve by recomputing them as successive words are added to a sentence. These incremental trajectories successfully distinguish garden-path sentences from disambiguated controls by measuring representational displacement around critical linguistic regions.