Probing Semantic Alignment, Lexical Invariance, and Syntactic Influence in LLM Metaphor Processing

This study presents a diagnostic analysis of large language models (LLMs) in metaphor processing, probing semantic alignment, lexical invariance, and syntactic influence. The analysis uses geometric probing to assess semantic attribute alignment, context-varying substitutions to evaluate lexical stability, and controlled syntactic perturbations to determine sensitivity in metaphor detection. The findings indicate that while LLMs perform well on metaphor benchmarks, their interpretations may exhibit semantic drift and sensitivity to syntactic changes, suggesting a complex interplay of signals that necessitates careful interpretation of their metaphor processing capabilities for practitioners.