Accuracy and Order Sensitivity Diverge Under Label-Free Strategies
This paper demonstrates that while preventing large language models from seeing option labels during multiple-choice answering successfully eliminates positional bias, it fails to improve accuracy and often degrades performance, revealing that the primary bottleneck is the withholding of options rather than the matching mechanism.
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 taking a multiple-choice test in a classroom. You know the answer, but the teacher has shuffled the order of the choices on the page. Suddenly, you pick the wrong letter, not because you forgot the fact, but because your brain got confused by the new arrangement. This is the strange world of Large Language Models (LLMs)—super-smart computer programs that chat and write like humans. Scientists use multiple-choice questions to test how much these models "know," but they've discovered a glitch: the models often act like nervous students who are too sensitive to where the answers are sitting on the page. If the choices are listed A, B, C, D, the model might get it right; shuffle them to B, A, D, C, and it might get it wrong. This makes the test scores unreliable, because we can't tell if the model is actually smart or just good at guessing the pattern of the letters. Researchers want to fix this so we can trust what these AI models are really saying.
In this paper, two scientists from Queen's University Belfast, Karl Hanna and Chen Feng, decided to test a clever idea to fix this "shuffling problem." They wondered: what if we made the model write its answer in plain English first, without seeing the multiple-choice options at all? Then, in a second step, we could ask the model to match its own written answer to the correct letter. It's like asking a student to write the answer on a scrap of paper, and then handing them the test sheet to circle the right bubble later. The hope was that by hiding the options during the thinking phase, the model couldn't be influenced by their order. They also tried a second method where the model looked at each option one by one in isolation, like tasting a single ingredient before deciding on a recipe, hoping this would remove any bias caused by the order of the ingredients.
The results, however, were a bit of a letdown. The researchers tested these methods on six different AI models across two major question sets (MMLU and ARC-Challenge). They found that neither of their "label-free" strategies reliably made the models smarter. In fact, the "write first, match later" method actually made the models worse at getting the right answer in almost every case. It turned out that hiding the options was the real problem; without seeing the choices, the models couldn't figure out how to phrase their answers correctly, leading to a lot of confusion and errors. The only version that worked as well as the standard test was one where the model could still see the options while thinking.
Even more surprisingly, the scientists discovered that making the models less sensitive to the order of the answers didn't automatically make them more accurate. For example, one model became much less confused by the shuffled letters, but its test score still dropped. It's like a student who stops guessing based on the letter order but still doesn't know the material well enough to get a better grade. The paper concludes that simply removing the "positional bias" (the confusion caused by where answers are placed) doesn't guarantee a better score. In fact, the traditional method of shuffling the answers and taking a vote (called cyclic permutation) actually worked better than these fancy new tricks. The main takeaway is that while we can fix the bias, it doesn't always mean the model gets smarter, and sometimes trying to be too clever with the test design just breaks the model's ability to answer at all.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.