Comparing Semantic Navigation in Humans and Large Language Models using Natural Language Processing

This paper treats semantic memory retrieval as navigation through conceptual space and compares verbal-fluency trajectories from 82 human participants with outputs from GPT-4o, Gemini-2.5-Pro, and Claude-Sonnet-4.5. Using entropy, distance to the next item, and distance to the centroid, the authors report that humans show less predictable transitions, larger semantic steps, and broader global dispersion. Eight temperature settings produced only partial metric-level matches; no tested configuration reproduced the full human profile across all three dimensions. The findings point to a difference between model sampling variability and human semantic exploration.