The Wording Effect: Quantifying Two-Way Drift in LLM Benchmark Performance
This paper introduces BenchDrift, a framework that quantifies how meaning-preserving rephrasings of benchmark problems cause significant two-way performance fluctuations in LLMs, revealing that phrasing sensitivity persists across model strengths and that fragility stems more from specific rephrasings than from individual model weaknesses.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a judge at a talent show, but instead of watching a performance, you are grading a robot's ability to solve puzzles. In the world of artificial intelligence, these puzzles are called "benchmarks." Think of a benchmark as a standardized test, like a final exam for a computer. For years, scientists have believed that if a robot gets a high score on this test, it truly understands the material. But there's a catch: every puzzle on the test is written in just one specific way. It's like asking a student, "What is 2 plus 2?" and giving them a perfect score, but never asking, "Can you tell me the sum of two and two?" or "If you have two apples and find two more, how many do you have?"
This paper dives into a corner of science called Natural Language Processing, which studies how computers understand human words. The key idea here is that computers are surprisingly sensitive to how a question is asked. Just like a human might get confused if a teacher uses a weird accent or a confusing sentence structure, a computer might stumble if the words change, even if the math or logic stays exactly the same. Why does this matter? Because if a robot's score depends entirely on the specific wording of the test, we can't really trust that score to tell us how smart the robot actually is. It might just be good at guessing the right answer to a specific riddle, not solving the problem itself.
The researchers behind this study, working at IBM, decided to investigate this "wording effect" with a new tool they built called BenchDrift. Imagine you have a magic photocopier that can rewrite a test question in dozens of different ways without changing the answer. You could ask the same math problem as a polite request, a dramatic story, a list of bullet points, or even a series of smaller questions. The team used BenchDrift to take real test questions and rewrite them along four different "axes": Linguistic (changing the words and style), Referential (swapping names or numbers), Pragmatic (changing the tone or setting, like pretending the robot is an artist), and Structural (changing how the information is organized).
They ran these rewritten questions through eight different AI models, ranging from smaller to very large ones. What they found was a bit of a shocker. They discovered a phenomenon they call "drift." Sometimes, rewriting a question helped the robot get the answer right when it had previously failed (positive drift). But often, rewriting a question made a robot fail a problem it had previously solved correctly (negative drift).
The most surprising discovery was that the smarter the robot, the more fragile it seemed to be. The strongest models, which usually get the highest scores, actually lost more correct answers when the questions were rephrased than they gained. It's as if the top students were so used to the specific way the teacher asked questions that they couldn't handle a slightly different phrasing. In fact, for the best models, a single test score was often an overestimate of their true ability. The researchers found that the gap between a model's "best-case" score (when it gets lucky with the wording) and its "worst-case" score (when the wording trips it up) was massive—on average, a swing of 74.7 percentage points. For example, a model that scored 93.4% on the original test dropped to 38.2% when the questions were rephrased in a certain way.
The team also ruled out a few easy excuses. They proved that the robots weren't just failing because the questions became harder or longer; sometimes, making a question shorter or longer caused the same amount of trouble. They also found that even when a robot was extremely confident in its answer (giving it a high probability score), a simple rephrasing could still break its logic. This suggests that the fragility isn't just about the robot being unsure; it's about the robot relying too heavily on the specific shape of the words.
Furthermore, the researchers noticed that different models tended to fail on the same types of rephrasings. For instance, breaking a problem into a series of smaller, explicit questions (a structural change) caused the most failures across the board, while simply adding extra spaces or changing the font caused very few. This means the problem isn't unique to one specific robot; it's a flaw in how these systems handle certain types of language structures.
In the end, the paper argues that a single benchmark score is like a snapshot of a moving car; it tells you where the car was at one exact moment, but not how stable it is. The researchers suggest that instead of just reporting one number, we should report a range: the worst score the model could get, the best score, and the average. This gives a much clearer picture of what the AI can actually do. They found that to get a reliable picture of a model's true strength, you need to test it on many different versions of the same problem, not just the one original version. If we don't do this, we might be celebrating robots that are actually quite brittle, just because they happened to get the wording of the test just right.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.