Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures

This study introduces a framework utilizing Shapley values to analyze the impact of adjectives on the performance of various Large Language Models (LLMs), including o3, gpt-4o-mini, phi-3, llama-3-70b, and deepseek-r1, evaluated on the MMLU benchmark. Key findings reveal that certain adjectives serve as powerful levers with effects contingent on their syntactic roles, and larger models exhibit complex interaction effects, complicating the steering of model behavior. These insights underscore the necessity for tailored prompting strategies and alignment techniques as model architectures evolve and their interpretative capabilities become more nuanced.