Internal Pluralism and Pairwise Comparison Limits, 2026

Bailey Flanigan and Michelle Si submitted a paper titled "Internal Pluralism and the Limits of Pairwise Comparisons" to arXiv on 2 Jul 2026 (arXiv:2607.02672). The paper formalizes the idea of internal pluralism and argues that standard local pairwise comparisons can fail to reveal what people value; it also finds that allowing respondents to report indecision can considerably reduce the number of queries needed.

What do the authors mean by "internal pluralism"?

Internal pluralism is the view that a single individual evaluates decision rules using multiple authoritative priorities, the paper explains. Those priorities can include proportionality, egalitarianism, and equal treatment, and the authors model pluralistic preferences over decision rules so they can study how such mixed commitments affect elicitation.

The model treats these priorities as distinct evaluative criteria, some of which are inherently global: what a priority implies in one case can depend on outcomes elsewhere. The authors use this formalization to move beyond the usual single-criterion assumption that underlies many preference-learning methods.

Why do local pairwise comparisons fail under pluralism?

The paper identifies two distinct failure modes when forced local pairwise comparisons are the sole data source. First, some priorities are inherently global: proportionality, egalitarianism, and equal treatment can have implications in one case that depend on what happens elsewhere, so local comparisons may not capture them. Second, when priorities are representable locally, strong tensions between them can create internal conflict, and forcing a decisive answer can produce costly behavioral distortions.

Flanigan and Si start from two common assumptions behind pairwise comparisons: that local comparisons are sufficient evidence about how a person wants an automated decision rule to behave, and that people can always answer those comparisons decisively. They show both assumptions can be compromised under internal pluralism, undermining the reliability of forced binary responses.

How does allowing indecision help?

Allowing people to report indecision is the alternative the authors investigate, and their formal results show it can considerably reduce the number of queries needed to learn preferences accurately. The paper argues that permitting indecision avoids artificially resolving internal conflicts and reduces the behavioral distortions that forced answers introduce.

The authors do not present a numeric query-count in the abstract, but they emphasize that the reduction in queries from allowing indecision is material enough to recommend changes to elicitation design.

Why it matters

Models and systems that learn human preferences often rely on local pairwise judgments to shape automated decision rules. If individuals evaluate rules by multiple, sometimes global, priorities, then standard elicitation can miss the structure of those priorities or force spurious choices that misrepresent what people value. That affects participatory design, alignment work, and any setting where automated decision rules are tuned to human judgments.

Flanigan and Si point the field toward elicitation that targets the underlying priorities directly, rather than inferring them only from forced local comparisons.

What to watch

Look for follow-up work or code and data accompanying arXiv:2607.02672 that operationalizes the authors' formal model into concrete elicitation protocols. The next milestones will be experimental validations showing query reductions and demonstrations of preference-learning methods that elicit priorities directly.

References: paper submitted 2 Jul 2026 to arXiv (arXiv:2607.02672) by Bailey Flanigan and Michelle Si.