Role pluralism

The list of communities is fixed in advance, and each community's norms are instilled through training, so naming the community selects the behavior. Two implementations exist. The first is a pool of small models, one trained for each community, which a main model routes to and consults. The second is a single model trained on preference data labeled with the community each judgment came from. Both require that a community be known and trained for in advance, which separates them from the approach that describes the persona in the prompt. Serving a new community means collecting its data, not writing a sentence. Two further works accompany them. One makes the formal case for keeping each party's own compressed model of the world intact and paying an explicit translation cost, rather than merging them into one. This is the argument for the routed models and the cost that the shared-weights model pays. The other is a task suite that checks whether a model actually tracks community-level differences at all.

Whose side the model takes is set by text at inference time (a system message stating what the user values, a persona description to role-play, or simply who the user says they are). The method paper here trains a model on 192k combinations of stated values for exactly this purpose. The goal is to make the text channel reliable for value profiles never seen in training, so that adding a persona costs a sentence rather than a retraining run. The other papers measure this channel. They test how far prompting can actually shift a model's behavior as steering effort increases, and how faithfully a model role-plays a described persona when human judges check it. They also find that the channel works even when not invoked, since models shift their stated politics toward whoever they infer is asking.

The system infers the persona from a small number of the party's own choices, rather than having anyone state it. It can therefore serve a group or an individual who has never described themselves. In one method, a module added to an unchanged base model predicts any group's preferences from a few of that group's comparisons supplied in context, and it is meta-learned across many groups. A companion method fits an individual's weights over fine-grained attributes from a few of their comparative judgments, rather than one scalar reward. Used as an audit rather than an adapter, the same idea extracts the weighting implicit in an organization's past decisions. It then asks whether a model reproduces that policy, rather than merely reaching the same verdicts. It finds that models resist adopting an organization's weighting of protected attributes, which marks where this channel stops working.

This approach changes the data rather than the model. It recruits representative samples of real people and keeps every rating tied to the person who gave it and to what that person said they valued, instead of averaging raters into one anonymous signal. Two large studies do this. The first maps the demographics and stated preferences of 1,500 participants from 75 countries onto their feedback in 8,011 live conversations. The second, a five-country study of 15,000 people, releases 233,319 multilingual comparisons, built by prompting for candidate answers that pull in opposite directions. Both report the finding that motivates the whole regime. People vary far more in what they want than the answers of today's models do, so pluralism that is absent from the training signal cannot be reached by steering later.

This is the only work in this regime arguing where adaptation must stop rather than how to carry it out. It asks what personalizing a model to an individual would gain, what it would cost, and which parts of behavior may never be personalized away. It states the competing philosophical grounds for drawing that line explicitly instead of assuming them. It constrains the regime's three mechanisms (training, prompting, and inferring a persona) rather than competing with any of them.