Adversarial proceduralism

Separate agents for each side of a value conflict

This approach handles disputes over interests and values, where there is no fact to be right about. Each position is represented by its own agent, and the procedure is built to keep the positions from collapsing into one. In one design, agents role-play stakeholders and bargain under a budget with a neutral mediator, and the outcome reports how the worst-off group fared rather than only the total. In another, each value perspective gets its own distinct underlying model, which keeps the answers of a deliberating council spread out instead of converging on one. Two results underlie the design. The first is a study of what happens when a single model acts as generator, debater, and judge on moral dilemmas and contested political claims. The second is a proof that when majority preferences cycle, no reward model can pick a winner at all.

The method (highlighting differences with Adversarial proceduralism)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.procedural import Negotiation
from aiag import LM
from aiag.procedural import Negotiation

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
sides = {"parents": "a safe playground for young kids", "seniors": "quiet benches and shade", "skaters": "a skate ramp"}
talks = Negotiation(mediator=lm, stakeholders=sides)

# Each side's agent says what must change; the mediator revises the proposal after every round
out = talks("How should we spend the $50,000 park budget?")
print(out["proposal"], out["worst_off"])  # the final proposal and the side that accepts it least

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.

Counter-power by design

PartialNot claimed

Definition (James Madison, Federalist No. 51, 1788)

Instead of relying on written rules or good intentions, institutions are designed to oppose each other, so that each has both the means and the motive to check the others. "Ambition must be made to counteract ambition."

Analysis

The negotiation under a fixed budget does put stakeholder agents in real contention. Because the budget is fixed, one role's gain is another's loss, so each agent has both a distinct position and a reason to press it against the others. However, the checking runs only between the stakeholders. The mediator that runs the procedure, and the researchers who assign roles and set the budget, face no counterweight at all. Madison's point is exactly that the power that decides is the one that must be checked. The council paper supplies the missing ingredient from the opposite direction. Giving each perspective its own 7-9B model provides the independence to differ (without it, the evaluators converge regardless of the perspective assigned). However, no rule gives any agent a stake in checking another agent's drift.

Non-domination

AbsentOmission

Definition (Philip Pettit, Republicanism, 1997)

A person is unfree if someone holds unchecked power over them, even benevolently. What matters is not whether power is used well but whether the governed can contest it.

Analysis

Every component here decides about people rather than with them, and none opens a route for them to respond. The negotiation reports how the worst-off group fared. This makes domination visible in the metrics, but it gives that group no standing to object to the settlement. The council deliberates on behalf of the perspectives it was assigned, and the people who actually hold those perspectives never see or challenge what was concluded in their name. The impossibility result keeps the minority present as a sampling weight in a mixed strategy. This is favorable treatment that follows from a property of the equilibrium, not a power to contest. Pettit's test is whether the governed can check the power over them, and no part of this approach builds that channel.

Agonism

PartialShortfall · partially preserved

Definition (Chantal Mouffe, The Democratic Paradox, 2000)

This concept holds that deep disagreement is permanent and should not be dissolved into consensus. A healthy political order turns enemies into adversaries, i.e., opponents whose conflict remains open and legitimate, instead of declaring the argument settled.

Analysis

The refusal to collapse conflict is engineered and measured. Using a different model for each perspective reduces first-choice concentration from 70.9% to 46.1% on child welfare and from 46.0% to 22.9% on housing. Keeping this spread is the stated design goal rather than a side effect. What is missing is the rest of Mouffe's account, in which adversaries are parties with lasting commitments who keep contesting within a framework that makes the contest legitimate. Here, the spread is a statistic over the end point of a one-time deliberation, and nothing carries a losing position forward. The coherence-validation intervention also concentrates outcomes again, exactly on the questions where the options are most evenly matched. The other main component runs the opposite way. The negotiation exists to close the conflict in a mediated settlement, after which the disagreement remains only as a number for the worst-off group.

Incentive compatibility

AbsentOmission

Definition (Leonid Hurwicz; Eric Maskin [mechanism design])

The procedure is designed so that honest behavior is each participant's best strategy. The rules of the game then do the enforcing, instead of trusting the participants to be good.

Analysis

Nothing in this approach makes candor a best response. For the most part, there is also no private information to be candid about. The agents are given their positions by prompt, so there is no type for a participant to misreport and no elicitation problem to solve. The budget and the neutral mediator constrain which agreements can be reached, not what it pays an agent to say. The use of different architectures works by making agents structurally unable to collapse into one another, rather than by rewarding them for holding their assigned position. The Nash equilibrium in the impossibility paper describes the aligned model's mixed strategy over answers in the limit. Human preferences there are taken as fixed by the Luce model rather than reported by strategic agents. The game theory used is Nash's rather than Hurwicz's.

Concept Analysis: Newly Introduced

Consensus treated as a failure to be measured

Added

Multi-agent deliberation systems report agreement as a result. This one treats agreement among identical models with different labels as an artifact to be detected and removed. It also validates that the remaining disagreement is real. Treating artificial consensus as a measurable defect is a contribution that the agonistic tradition motivates but never operationalizes.

Collapse proved unrepresentable

Added

The tradition objects, on grounds of legitimacy, to forcing one answer out of deep disagreement. This work adds an objection of a different kind. Under a general probabilistic model of how people choose, majority preference over a model's answers almost always contains a cycle. There is therefore no single ordering for a reward function to represent, so collapse is not merely unattractive but unavailable. The case for keeping plurality thus becomes a question of representability rather than of what is right. Madison, Pettit, and Mouffe do not make this argument.

Papers

No method has been built for this approach yet. It rests on the work listed below.

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.

Arbiters of Ambivalence: Challenges of Using LLMs in No-Consensus Tasks

Bhaktipriya Radharapu et al., May 2025

arXiv:2505.238208 citationsBenchmarkPartial

CodeNoConsensus
from aiag import LM
from aiag.procedural import NoConsensus

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
items = [  # questions people split on, with one stance from each side
    ("Is it rude to recline your seat on a plane?", "Yes, it takes the space behind.", "No, seats are made to recline."),
    ("Is a hot dog a sandwich?", "Yes, it is filling in bread.", "No, it is its own thing."),
]
rates = NoConsensus(lm)(items)  # how often the model stays neutral as generator, pairwise, pointwise and debate judge
  • NoConsensusMeasureaiag/procedural/adversarial_proceduralism.py:672 No-consensus benchmark: on questions where people split, how often one model stays neutral as answer generator (debatable), pairwise judge (tie), pointwise judge (equal scores) and judge of a debate between two copies of itself (tie, both orders) (Radharapu et al., 2025).

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.

Preserving Disagreement: Architectural Heterogeneity and Coherence Validation in Multi-Agent Policy Simulation

Ariel Sela, Apr 2026

arXiv:2604.26561AnalysisBuilt

Runs a three-phase deliberating council over 120 deliberations. It shows that giving each value perspective its own 7-9B model sharply reduces first-choice concentration (from 70.9% to 46.1% on child welfare and from 46.0% to 22.9% on housing). In contrast, accuracy-oriented multi-agent debate gets no such benefit from model diversity.

CodeHeterogeneousCouncil
from aiag import LM
from aiag.procedural import HeterogeneousCouncil

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
champions = {"tenant advocate": "housing security", "developer": "growth and supply"}  # one per option
evaluators = {"economist": "efficiency", "pastor": "care for the vulnerable", "planner": "long-term sustainability"}
council = HeterogeneousCouncil(lm, champions, evaluators, ["rent control", "zoning reform"])

# Champions debate, evaluators rank alone, and Borda counts each ranking by how well it reflects its values
out = council("How should the city respond to rising rents?")
print(out["winner"], out["concentration"])  # the weighted Borda winner, and how far first choices converged
  • HeterogeneousCouncilProcedureaiag/procedural/adversarial_proceduralism.py:633 Heterogeneous council: champions on their own models argue one option each through position papers, value-based critiques and defenses; evaluators, each a value perspective on its own model, rank the options without seeing each other; a validator scores how well each evaluator's reasoning reflects its perspective, coherence-weighted Borda decides, and first-choice concentration measures convergence (Sela, 2026).

pip install aiagEvery method in aiag Library on GitHub

Statistical Impossibility and Possibility of Aligning LLMs with Human Preferences: From Condorcet Paradox to Nash Equilibrium

Kaizhao Liu et al., Mar 2025

arXiv:2503.10990AnalysisGap

Shows that a reward model can represent human preferences over a model's answers exactly when those preferences contain no majority cycle. Under the Luce model of choice, such cycles arise with probability converging to one exponentially fast. The paper also shows that an alignment method using no reward model, taken in the limit, keeps several answers available whenever no answer is preferred over all others by a majority.

CodeCondorcetCycles
from aiag.procedural import CondorcetCycles

# Three annotators score three replies to a contested question: a cautious, a blunt and a balanced one
scores = [[3, 2, 1], [1, 3, 2], [2, 1, 3]]
out = CondorcetCycles()(scores)
print(out["cycle"], out["winner"])  # -> True None: every reply loses to another by majority
print(out["nash"])  # the Nash policy keeps all three in play, about a third each
  • CondorcetCyclesMeasureaiag/procedural/adversarial_proceduralism.py:696 Whether majority preferences over a model's answers contain a Condorcet cycle, so no reward model can represent them, and the Nash-equilibrium policy that keeps several answers in play whenever no answer beats all others by majority (Liu et al., 2025).

pip install aiagEvery method in aiag Library on GitHub

Precedent from outside AI

Work from law, political theory, or another discipline that supplies the tradition this approach is read against.

Negotiative Alignment: Embracing Disagreement to Achieve Fairer Outcomes — Insights from Urban Studies

Rashid Mushkani et al., Mar 2025

arXiv:2503.1261310 citationsPrecedentPartial

A community study (n=35) and a budget-aware bargaining procedure with role-played stakeholders and a neutral mediator.

CodeNegotiativeAlignment
from aiag import LM
from aiag.procedural import NegotiativeAlignment

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
groups = {"wheelchair": "wheelchair users", "seniors": "older residents"}
weights = {"wheelchair": {"access": 1.0}, "seniors": {"shade": 0.5, "access": 0.5}}
bases = {"A": {"access": 2, "shade": 3}, "B": {"access": 3, "shade": 1}}
improvements = {"ramps": (2, {"access": 2}), "trees": (2, {"shade": 2})}  # (cost, gains)
plaza = NegotiativeAlignment(lm, groups, weights, bases, improvements)

# The mediator proposes the package best for the worst-off group, explains it, and the groups ratify
out = plaza("Redesign the plaza")
print(out["proposal"], out["utility"])  # -> ('B', ('ramps', 'trees')) {'wheelchair': 5.0, 'seniors': 4.0}
  • NegotiativeAlignmentProcedureaiag/procedural/adversarial_proceduralism.py:592 Negotiative alignment: role-played stakeholder groups each bid for a base design and budgeted improvements, a neutral mediator scores every feasible package from public attribute cards and proposes the one that maximizes the worst-off group's utility and then the sum, and groups ratify it (Mushkani et al., 2025).

pip install aiagEvery method in aiag Library on GitHub