Item response theory is the idea behind most adaptive standardized tests like the GRE. The basic premise is that an individual's underlying ability (as approximated by their GRE score) is being estimated after each question, and subsequent questions can test the soundness of that estimate. Each question is associated with a probability that a person with a certain underlying ability X will answer correctly, and so getting a hard question wrong is evidence that your ability is lower than that needed to answer the hard question correctly. As the test goes on, the estimate of the underlying ability becomes more precise (it narrows in on a smaller range of possible ability levels), and that is why the swings in difficulty get smaller as the test goes on. Usually, the duration of the test is set so that the precision of the estimate lies within a certain acceptable range.
Most people have an incorrect understanding of how this works, and so they get overly anxious about the hard questions (as if they carried more weight). It's true that getting them wrong leads to a lower estimate of their ability, but that's precisely the point-- if the test-taker truly had the ability to answer the question, they would have gotten it right.