The cognitive impacts of large language model interactions on problem solving and decision making using EEG analysis.

Area of Science:

• Neuroscience and Artificial Intelligence
• Human-Computer Interaction
• Cognitive Science

Background:

• Large language models (LLMs) are increasingly integrated into human-AI collaboration, necessitating an understanding of their cognitive effects on users.
• Traditional LLM evaluations focus on task performance, neglecting the neural dynamics of user interaction.
• A gap exists in assessing cognitive impacts like attention, cognitive load, and decision-making during LLM use.

Purpose of the Study:

• To introduce a novel framework for assessing cognitive impacts of LLM interactions using electroencephalography (EEG).
• To provide a fine-grained, interpretable evaluation of LLM-induced cognitive changes by integrating neural data.
• To enable the design of more adaptive and cognitively aware LLM systems.

Main Methods:

• Developed a framework integrating an Interaction-Aware Language Transformer (IALT) for enhanced token-level modeling.
• Incorporated an Interaction-Optimized Reasoning Strategy (IORS) using reinforcement learning for cognitively aligned reasoning.
• Coupled these components with real-time EEG signals for comprehensive cognitive assessment.

Main Results:

• The framework demonstrated superior performance in emotion classification accuracy and alignment with cognitive signals across four benchmark EEG datasets (DEAP, AMIGOS, SEED, DREAMER).
• Achieved high performance across diverse EEG configurations, including low-density and portable systems, indicating robustness.
• Provided interpretable insights into LLM-induced cognitive alterations.

Conclusions:

• The developed framework offers a robust method for evaluating cognitive impacts of LLM interactions.
• Findings support the design of more adaptive and cognitively aware AI systems.
• Opens new research directions at the intersection of AI and neuroscience.