When you ask a language model to judge many things in one prompt, like grading answers or flagging harmful comments, you might assume each item is judged on its own. But the model reads all items together, so its verdict on one can be influenced by the others and by its position.
Researchers tested this by giving the same items individually and in different random groups, and by changing the neighbor items, their order, and the slot. They also swapped in text and verdicts from other groups to see how the effect travels.
They found that about one in seven decisions changed when the group changed, even though overall accuracy stayed similar. Batching makes judgments unstable, especially for borderline items. The practical tip is to grade risky items one at a time or check how much your batch changes results.