Prefrontal multiscale entropy and state transitions distinguish large language model-assisted from search-assisted...

Learning is fundamental to human development and relies critically on cognitive processing. Increasingly, individuals use external tools, particularly search engines and large language models (LLMs), to enhance learning. Although both tools can improve learning outcomes, they may shape neurocognitive processing through distinct pathways. The present study examined prefrontal neural dynamics, using multiscale entropy and latent state transitions, across three conditions: LLM-assisted, Search-assisted, and an Unassisted condition. Participants completed an identical cognitive test while prefrontal hemodynamic activity was recorded using functional near-infrared spectroscopy (fNIRS). Multiscale entropy was used to quantify the multiscale temporal irregularity in prefrontal signals, whereas Gaussian Hidden Markov Model (GHMM)-derived transition ratios and dwell times were used to characterize latent prefrontal state transitions. Both LLM-assisted and Search-assisted conditions improved test performance relative to the Unassisted condition. However, they exhibited dissociable neural patterns. Compared with search assistance, LLM assistance was associated with lower entropy, fewer state transitions, and longer dwell times. These findings indicate that comparable behavioral performance may be supported by distinct patterns of prefrontal signal complexity and latent state transitions. More broadly, these findings may highlight the potential importance of considering tool-specific cognitive processing dynamics when selecting and integrating external tools in educational settings.