A paper questions pairwise comparisons in AI alignment

Proportionality, egalitarianism and equal treatment: according to a paper published on arXiv on July 7, 2026, these priorities are inherently global, and pairwise comparisons, the standard tool for learning human preferences in AI alignment and participatory design, fail to capture them. The work, available as arXiv:2607.02672, formalizes a concept the authors call internal pluralism, from which two concrete failures of this type of data follow.

The method in question is everywhere. When an annotator chooses between two model responses, when a platform asks which of two options you prefer, or when a participatory process asks people to rank alternatives two at a time, local pairwise comparisons are being used. The abstract itself places the problem in two contexts: participatory design and the alignment of automated systems. On top of that data sit the reward models that guide techniques such as RLHF, used to train today's leading language models.

Two assumptions almost nobody questions

The authors point out that this method assumes two things without saying so: 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. The paper investigates what happens when both assumptions collide with internal pluralism: the idea that a single person evaluates decision rules according to several priorities they consider legitimate at the same time.

You do not have to look far to see it. Someone may want an allocation system to be proportional to merit and, at the same time, want nobody to fall below a minimum. Both priorities are reasonable, both are theirs, and in many concrete cases they conflict.

Where local comparisons fail

With a formal model of these pluralistic preferences, the paper identifies two distinct failures. The first: priorities such as proportionality, egalitarianism or equal treatment are global by nature. What they imply in one case depends on what happens in the rest of the cases, so no amount of local comparisons is enough to capture them. The second: even when priorities can be represented locally, tension between strongly held priorities generates internal conflict, and forcing an answer produces potentially costly behavioral distortions in the data.

What this means for LLM alignment

The reading for the language model ecosystem is direct. Preference-based alignment pipelines, from RLHF to variants with automated judges, inherit these limitations if their raw material is forced binary choices. Today's models, including those from Anthropic and other labs, are trained in part with comparisons of this kind. If some annotators answer under internal conflict, the reward model is not learning a clean preference but the residue of a tension the question format does not allow them to express.

This does not invalidate current methods, but it does bound what they can promise. And it connects with a broader conversation on pluralistic alignment that the research community has been opening for some time: how to design systems that represent multiple, conflicting values instead of flattening them into a single reward function.

Who this is useful for

The paper is relevant to alignment researchers, but also to more applied profiles: teams collecting user feedback with thumbs up and down, products running A/B tests on assistant behavior, and participatory design or civic tech projects that aggregate citizen preferences. In all those cases, the format of the question determines which values make it into the data and which are left out.

At ElephantPink we work daily with systems tuned through human preferences, and this kind of theoretical work seems necessary to us: before discussing whether a model is well aligned, it helps to know what the data used to align it can and cannot measure.