This study conceptualizes semantic memory retrieval as navigation through conceptual space. We compared the semantic search dynamics of 82 human participants with three large language models (GPT-4o, Gemini-2.5-Pro, Claude-Sonnet-4.5) using trajectory-based NLP metrics across eight temperature settings.
We quantified three complementary dimensions: entropy (step size predictability), distance to next (successive semantic steps), and distance to centroid (global dispersion). Results indicated that humans exhibited higher entropy, larger semantic steps, and broader dispersion, suggesting a more variable and exploratory search process.
Temperature tuning produced only partial alignments, with specific metrics matching between humans and LLMs under certain settings, but no configuration reproduced the complete human profile across all dimensions. These findings suggest that human semantic search implements a unique balance between local exploitation and global exploration, which current model architectures fail to replicate.
Blogger's Review: This study provides an in-depth comparison of semantic navigation between humans and large language models, highlighting human superiority in exploratory and variable search behavior. This insight could pave the way for future model improvements, especially in simulating human-like semantic search processes. It is hoped that subsequent research will further bridge this gap.