Search Papers | Poster Sessions | All Posters
Poster C62 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Language evolution in large language models and humans: a comparative analysis of developmental linguistics across ages and sensory modalities.
Gina Yu1, Jad El Harake2, Stephen Chong Zhao3, Jason Lee4, Andrew Bender5, Trisha Mazumdar4, Adaline Leong4, Prince Owusu Nkrumah4, Mark Wallace6, David A. Tovar6 (); 1Neuroscience Department, Vanderbilt University, 2Biomedical Engineering Department, Vanderbilt University, 3Data Science Institute, Vanderbilt University, 4Computer Science Department, Vanderbilt University, 5Neurosciences Graduate Program, University of California San Diego, 6Psychology Department, Vanderbilt University
Large Language Models (LLMs) possess the ability to adopt various roles, including simulating language use across different ages, which suggests a meta-level understanding of language evolution in individuals. This study investigates the extent to which LLMs can mimic human developmental linguistics, comparing their language outputs when summarizing short stories via different sensory modalities—image-only, text-only, and both—to structural MRI captured across human subjects over a lifetime. We analyze the complexity of language and story development through semantic embeddings and assess the embeddings' correlation with human neural patterns. Our findings indicate that the model's ability to adapt its linguistic outputs to replicate age-specific developmental stages varies significantly with the input modality, showing substantial correlations between the LLM's text outputs and structural brain changes in humans. Furthermore, we found that brain-network changes were specific to the input modality chosen. For example, developmental LLM image input response changes were more closely tied to visual networks than when the input was text. This work has implications for our understanding of LLMs as well as our understanding of the developmental linguistic and sensory changes that occur over a lifetime.
Keywords: Large Language Models Semantic Embeddings Age Development MRI