Interpreting token compositionality in LLMs: A robustness analysis [2410.12924]