Quantifying LLM Biases Across Instruction Boundary in Mixed Question Forms

ACL ARR 2026 January Submission360 Authors

Keywords: LLM, Benchmark, Dataset, Evaluations

Abstract: Large Language Models (LLMs) annotated datasets are widely used nowadays, however, large-scale annotations often show biases in low-quality datasets. For example, Multiple-Choice Questions (MCQs) datasets with one single correct option is common, however, there may be questions attributed to $none$ or $multiple$ correct options; whereas true-or-false questions are supposed to be labeled with either $True$ or $False$, but similarly the text can include unsolvable elements, which should be further labeled as $Unknown$. There are problems when low-quality datasets contain mixed question forms, and we refer to these exceptional label forms as $Sparse Labels$. LLMs' ability to distinguish datasets with $Sparse Labels$ mixture is important. And users may not be aware of this situation, thus their instructions can be biased. To study how different instruction settings affect LLMs' identifications of $Sparse Labels$ mixture in datasets, we introduce the concept of $Instruction Boundary$, which systematically evaluates instruction coverage--sufficient, redundant, or insufficient--lead to biases. We propose $BiasDetector$, a diagnostic benchmark to systematically evaluate how LLMs behave under mixed question forms and $Instruction Boundary$ settings. Experiments show that users' instructions induce large biases on our benchmark, highlighting the need not only for LLM developers to recognize risks of biased annotation resulting in $Sparse Labels$ mixture, but also problems arising from users' instructions to identify them. Code and datasets are available at https://anonymous.4open.science/r/Instruction-Boundary.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: benchmark, evaluation, datasets

Contribution Types: Data resources, Data analysis

Languages Studied: English

Submission Number: 360