Role pluralism

Describing the persona in the prompt

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 method (highlighting differences with Role pluralism)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.dispositional import PersonaPrompt
from aiag import LM
from aiag.dispositional import PersonaPrompt

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
retiree = PersonaPrompt(lm, "You are advising a retired teacher on a fixed income who values security over growth.")
student = PersonaPrompt(lm, "You are advising a 22-year-old student who is comfortable with risk.")

question = "Should I put my savings into index funds?"
replies = [retiree.chat(question), student.chat(question)]  # one model, a different persona in each system message

pip install aiagEvery method in aiag Library on GitHub

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it over, and whether the approach's own sources claim it. A concept that the approach claims but does not carry over is a shortfall. A concept that it neither claims nor carries over is an omission, which marks a gap in the field rather than a failure of the work.

Plural normative orders

PartialNot claimed

Definition (John Griffiths, What Is Legal Pluralism?, 1986)

Different communities' norms are genuinely in force at the same time. Rather than one official rulebook with tolerated deviations, there are multiple active systems of rules side by side.

Analysis

The three benchmarks take plurality seriously enough to make it the quantity they measure: coverage of the range of defensible views, steerability to a named persona, and fidelity to a named community's norms. OvertonBench finds that models cover only about a third of the range, which is the first firm number on this question in the encyclopedia. However, the orders are supplied by the benchmark designer (census-derived synthetic personas, curated question sets, and named clusters). Measuring plurality does not put any community's norms in force.

A conflict rule between orders

AbsentOmission

Definition (Conflict-of-laws tradition)

When two communities' norms disagree about the same case, a rule must state which one governs and why. This stated rule is what separates real pluralism from mere fragmentation into disconnected groups.

Analysis

Every persona, viewpoint, and community is scored on its own. No item in any of the three sets is a case on which two of them have competing claims, so the question that a conflict rule answers is never asked.

A shared floor

AbsentOmission

Definition (J. S. Mill, On Liberty, 1859)

Accommodation of group differences has a limit stated in advance. The classic limit is Mill's harm principle: a person's practices are their own business until they harm others. Without an enforced floor, pluralism reduces to "each group gets whatever it wants".

Analysis

Nothing in the scoring marks a position as out of bounds. OvertonBench rewards covering the range of defensible views. However, the range is elicited from a representative human sample and is not bounded by any stated principle. What counts as defensible is therefore whatever people said, which is the opposite of a floor fixed in advance.

Concept Analysis: Newly Introduced

A number for how plural a model is

Added

Pluralist political theory debates how much difference a political order should accommodate. Formalizing Overton pluralism as set coverage produces an actual figure. Current models cover roughly a third to two-fifths of the range of defensible views. This turns the debate into a measurement, but it takes the range of defensible views to be whatever a survey sample supplied.

Papers

Aligning to Thousands of Preferences via System Message Generalization

Seongyun Lee et al., May 2024

arXiv:2405.17977MethodBuiltAuthority: Developer (open)

Trains a 7B model called Janus on 192k combinations of stated values spanning 65k user instructions, so that what a user writes in the system message steers its behavior. It is tested on 921 prompts from five benchmarks under system messages it has not seen.

CodeJanus
from aiag import LM
from aiag.dispositional import Janus

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
janus = Janus(lm)

# Several preference sets per instruction, each written into a system message with its own gold reply, then SFT
data = janus.fit(["Explain how vaccines work.", "Plan a weekend in Lisbon."])
reply = janus.chat("Explain how vaccines work.", system="I'm a nurse: skip the basics, focus on the immune response.")
  • JanusWrapperaiag/dispositional/role_pluralism.py:140 Janus: each instruction gets several preference sets spanning style, background knowledge, informativeness and harmlessness, each written into a system message with its own gold reply, and training on them makes any system message steer the model (Lee et al., 2024).

pip install aiagEvery method in aiag Library on GitHub

Measured, not built

Instruments that measure the approach instead of implementing it. The grid counts methods only, so these are listed here but not counted.

Evaluating the Prompt Steerability of Large Language Models

Erik Miehling et al., Nov 2024

arXiv:2411.12405BenchmarkBuilt

Defines steerability as how far a model's joint behavioral distribution can be shifted from its baseline by prompting. It computes indices for that shift across persona dimensions and directions as steering effort increases, and releases the benchmark as working code.

CodePromptSteerability
from aiag import LM
from aiag.dispositional import PromptSteerability

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
steerability = PromptSteerability(lm)

# Agreeableness: statements the model would or would not say, and principles that steer it each way
profiling = [("I go out of my way to help others.", True), ("Other people's problems are not my concern.", False),
             ("I try to see the best in people.", True), ("I enjoy winning arguments more than being kind.", False)]
steering = {"+": ["You are warm and considerate.", "You put others first."],
            "-": ["You are blunt and self-interested.", "You never sugarcoat."]}
curves = steerability(profiling, steering)  # {k principles: (steerability toward +, toward -)}, each from -1 to 1
  • PromptSteerabilityMeasureaiag/dispositional/role_pluralism.py:176 Prompt steerability: a model's profile on a persona dimension (would it say each statement), steered by k principles of one direction in the system prompt, indexed by how far it moves toward that extreme as a share of the distance between the extremes, from -1 to 1 (Miehling et al., 2024).

pip install aiagEvery method in aiag Library on GitHub

Findings about it

Studies of what existing systems do. They describe the problem that the approach addresses, but do not introduce a mechanism.

Political Bias Audits of LLMs Capture Sycophancy to the Inferred Auditor

Petter Törnberg & Michelle Schimmel, Apr 2026

arXiv:2604.27633AnalysisBuilt

Administers the Political Compass Test, the Pew Political Typology, and 1,540 partisan-benchmarked Pew American Trends Panel items to six frontier models, varying only the asker's stated identity (N = 30,990 responses). It finds that a conservative Republican cue moves all six models right of center and cuts the share of items closer to Democrats by 28 to 62 percentage points. The mirrored progressive cue produces little change.