Semantic diversity is the degree to which outputs differ in meaning, not just in wording. Two responses can look varied on the surface while still repeating the same claim, reasoning path, or conclusion. In LLMs, low semantic diversity often means the system is sampling many phrases from a narrow argument space instead of exploring distinct ideas.
This matters in cyber security because narrow outputs can hide gaps in analysis. If a model is used for threat intelligence, policy drafting, phishing review, or incident summaries, low semantic diversity may miss uncommon tactics, edge cases, or alternative hypotheses. Attackers can also exploit the opposite problem by generating semantically diverse phishing lures that evade simple text filters. Defenders measure semantic diversity to test whether multiple model samples truly add coverage, improve robustness, or are only reworded versions of one default answer.