CompCog: Integrating retrieval and prediction in sentence processing using attention-based LLMs

Understanding language is one of the most remarkable abilities of the human mind. Every conversation, text message, news article, or book requires people to remember information, make predictions about what comes next, and rapidly combine this information to understand what is being communicated. Both memory and prediction are essential for language understanding, yet scientists do not know how they work together. This project uses recent advances in artificial intelligence (AI) to address that question and provide new insights into the nature of human intelligence and cognition. By comparing how people and AI systems process language, the project will help the public better understand what today’s AI systems can and cannot tell us about the human mind, as well as the capabilities and limitations of AI technologies that are becoming increasingly common in everyday life. This project tests the hypothesis that memory retrieval and prediction in human language processing can be jointly explained through the computational mechanisms of attention that underlie modern large language models. The research evaluates two major theoretical frameworks that explain processing difficulty during language comprehension: memory-based accounts, which emphasize retrieval and interference among stored representations, and expectation-based accounts, which emphasize probabilistic prediction. Using a comprehensive suite of attention-based and prediction-based metrics derived from large language models, the project systematically compares model behavior with human behavioral data across a broad range of established language processing phenomena involving syntactic structure, semantic interpretation, and retrieval interference. Analyses focus on determining when model-derived attention patterns successfully capture the behavioral signatures of human processing and what these patterns reveal about the mental representations, memory cues, and cognitive mechanisms that support language comprehension. By identifying points of convergence and divergence between computational models and human behavior, the project seeks to develop a unified account of memory and prediction in language processing and to generate new insights into the cognitive architecture underlying human language. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2612265 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Daniel Parker | Institution: OHIO STATE UNIVERSITY, THE, COLUMBUS, OH | Award Amount: $459,039 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2612265 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2612265.html