Recent research tested eight large language models (LLMs) and fine‑tuned encoders by removing nouns from input data. On the standard natural language inference set MNLI, LLMs retained 70–72 % of their above‑chance accuracy even after noun removal. The same perturbation left only 14–29 % of their ANLI performance intact. For the more recent ANLI benchmark, noun removal caused a sharper decline, with only 14–29 % of performance preserved. Fine‑tuned encoders showed a similar pattern: they kept 72 % of MNLI performance but only 24 % on ANLI. These figures come from a controlled experiment that varied the perturbation—drop‑NOUN, drop‑VERB, keep‑NOUN+VERB, and shuffled word order—and measured how much accuracy remained relative to a clean baseline. The results suggest that LLMs rely less on simple noun cues in MNLI but are more sensitive to perturbations on ANLI, pointing to dataset‐specific shortcuts rather than true language understanding.