Interpreting token compositionality in LLMs: A robustness analysis