ABCD: All Biases Come Disguised

Multiple-choice question (MCQ) benchmarks have been a standard evaluation practice for measuring LLMs' ability to reason and answer knowledge-based questions. Through a synthetic NonsenseQA benchmark, we observe that different LLMs exhibit varying degrees of label-position-few-shot-prompt bias, where the model either uses the answer position, the label in front of the answer, the distributions of correct answers present in the few-shot prompt, or a combination of all to answer each MCQ question. We propose a simple bias-reduced evaluation protocol that replaces the labels of each question with uniform, unordered labels and prompts the LLM to use the whole answer presented. With a simple sentence similarity model, we demonstrate improved robustness and lower standard deviation between different permutations of answers with a minimal drop in the LLM's performance, exposing the LLM's capabilities under reduced evaluation artifacts, without any help from the prompt examples or the option labels. Across multiple benchmarks and models, this protocol substantially improves the robustness to answer permutations, reducing mean accuracy variance $3\times$ with only a minimal decrease in the model's mean performance. Through ablation studies on various embedding models and similarity functions, we show that the method is more robust than the standard ones.

When we evaluate large language models, we usually ask them multiple-choice questions: pick A, B, C, or D. The scores from these tests drive everything — from everyday choices to important decisions. We wanted to know whether those scores actually measure reasoning. To find out, we built a deliberately absurd benchmark called NonsenseQA. Every question is a jumble of random words. The four answer options are random words too. We pick one at random to call "correct." There is nothing to reason about — a model should score 25%, the same as guessing. When we ran thirteen leading open-source models on it, several scored above 95%. Why? They were not understanding the questions. They were picking up on patterns in the test format itself: which letter tended to be the correct answer in previously shown examples (e.g., if the previous five answers were all "A," the model leans toward "A" again), which position correct answers usually sat in, even the implicit ordering of A through D. These shortcuts are good enough to look like genuine intelligence on standard benchmarks. If models can game a benchmark made of nonsense, what is stopping them from gaming the real ones? We propose a small change to fix this issue. Instead of labeling options A, B, C, D, we label them all with a dash. Instead of asking the model to pick a letter, we ask it to write out the full answer in its own words, then match that against the four candidates using a small text-similarity model. We call this Matched-and-Dashed. It requires no retraining, no internal access to the model, and almost no extra computation. Under our protocol, most of the models that scored 95% on nonsense fall back to roughly 25%, where they should be. On real benchmarks — commonsense reasoning, science questions, graduate-level material, and four languages — score variance drops three-fold while actual performance barely changes. The reasoning that was always there shows up more cleanly once the shortcuts are removed. We hope this becomes a standard part of LLM evaluation.