LLMs Show Strong Order Biases in High-Stakes Decisions, Picking Inferior Options

Large language models deployed to guide consequential decisions in hiring and university admissions are susceptible to significant biases based on the order in which candidates or options are presented—a problem that can lead them to recommend objectively worse choices.

In a new study submitted to arXiv, researchers conducted the first systematic investigation of how position affects LLM judgment across multiple models and realistic scenarios. The work reveals that these biases operate differently depending on the quality of the options being compared, and identifies a previously unreported form of discrimination tied to applicant names.

How Order Reshapes LLM Preferences#

The team tested position effects in two domains: resume comparisons simulating real hiring decisions, and color selection tasks designed to eliminate confounding variables. Their findings show a quality-contingent pattern: when all candidates are highly qualified, models tend to favor whoever appears first, but shift to preferring later options when quality is lower overall.

This behavior diverges from simple alphabetical or positional anchoring. Instead, the models exhibit what researchers characterize as a fragile preference structure—meaning their selections rest on shallow reasoning that collapses when the presentation order changes, rather than reflecting genuine underlying judgment about candidate quality.

A Hidden Name Bias#

Beyond order effects, the research uncovered name-based favoritism, where particular names systematically received preferential treatment independent of any demographic information. This bias persisted even after researchers controlled for signals typically associated with protected characteristics, suggesting the models learned associations between certain names and perceived competence or fit during training.

Distinguishing Genuine Judgment from Order-Dependent Noise#

The researchers extended established rational choice theory to classify pairwise preferences into three categories: robust (consistent across presentation orders), fragile (sensitive to order but internally consistent), and indifferent (genuinely ambivalent). This framework allowed them to pinpoint cases where order effects don't merely break ties but actually flip models toward objectively worse candidates—a failure mode distinct from documented human biases.