A controlled study finds that while state-of-the-art large language models appear robust to task-irrelevant context at the aggregate level, this stability masks significant per-example instability. Prepending semantically meaningless pseudo-words to benchmark questions causes marked shifts in model predictions on a small fraction of examples.

  • The two-sided effect degrades performance on some examples while improving it on others, holding consistently across a wide range of models and datasets.
  • Affected examples are largely model-specific and the instability is modulated by context type, context length, test-time compute, and model development stage.

These findings reveal context-induced tail risks concealed by aggregate accuracy, motivating per-example reliability evaluation of language models.