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

arXiv cs.CL Papers

Summary

This paper compares semantic search dynamics between humans and LLMs using verbal fluency data, finding that humans exhibit more variable and exploratory search patterns that current models fail to reproduce.

arXiv:2607.12195v1 Announce Type: new Abstract: Semantic memory retrieval can be conceptualized as navigation through conceptual space. We compared semantic search dynamics between humans and three large language models (GPT-4o, Gemini-2.5-Pro, Claude-Sonnet-4.5) using verbal fluency data. By applying trajectory-based NLP metrics to the items generated by 82 human participants and LLM output 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). Humans exhibited higher entropy, larger semantic steps and broader dispersion than all LLMs, indicating more variable and exploratory search. Temperature tuning produced only partial alignments, as individual metrics matched between humans and LLMs at specific settings, but no configuration reproduced the complete human profile (in all dimensions). These findings suggest that human semantic search implements a distinctive balance between local exploitation and global exploration that current model architectures fail to reproduce.
Original Article
View Cached Full Text

Cached at: 07/15/26, 04:22 AM

# Comparing Semantic Navigation in Humans and Large Language Models using Natural Language Processing
Source: [https://arxiv.org/abs/2607.12195](https://arxiv.org/abs/2607.12195)
[View PDF](https://arxiv.org/pdf/2607.12195)

> Abstract:Semantic memory retrieval can be conceptualized as navigation through conceptual space\. We compared semantic search dynamics between humans and three large language models \(GPT\-4o, Gemini\-2\.5\-Pro, Claude\-Sonnet\-4\.5\) using verbal fluency data\. By applying trajectory\-based NLP metrics to the items generated by 82 human participants and LLM output across eight temperature settings, we quantified three complementary dimensions: entropy \(step size predictability\), distance to next \(successive semantic steps\), and distance to centroid \(global dispersion\)\. Humans exhibited higher entropy, larger semantic steps and broader dispersion than all LLMs, indicating more variable and exploratory search\. Temperature tuning produced only partial alignments, as individual metrics matched between humans and LLMs at specific settings, but no configuration reproduced the complete human profile \(in all dimensions\)\. These findings suggest that human semantic search implements a distinctive balance between local exploitation and global exploration that current model architectures fail to reproduce\.

## Submission history

From: Felipe Toro Hernández \[[view email](https://arxiv.org/show-email/de8702d8/2607.12195)\] **\[v1\]**Mon, 13 Jul 2026 22:38:58 UTC \(635 KB\)

Similar Articles

How Human-Like Are Large Language Models? A Register-Aware Linguistic Evaluation Framework

arXiv cs.CL

This paper introduces a register-aware linguistic evaluation framework to assess how human-like large language models (LLMs) are by comparing the distribution of 67 lexico-grammatical features between human and LLM-generated texts using Maximum Mean Discrepancy. Experiments across seven instruction-tuned open-source models and five registers show that no model perfectly matches human baselines, and closeness to human language varies by register rather than model size.

Same Patient, Different Words, Different Diagnosis? Evaluating Semantic Stability in Clinical LLMs

arXiv cs.CL

This paper proposes a semantic verification framework using Natural Language Inference (NLI) to evaluate the sensitivity of clinical LLMs to meaning-preserving prompt variations, introducing metrics such as MVS, ΔC, and WCI. Results show that domain specialization does not consistently improve robustness, with both domain-specific and general-purpose models showing mixed performance.