Tag
The paper tests label-free strategies to reduce positional bias in multiple-choice benchmarks but finds they do not reliably improve accuracy or reduce bias. Only showing all options with an LLM matcher preserves baseline performance, while cyclic permutation sometimes helps.