Customary pluralism

Engineering a plurality of conventions in artificial agents

Rather than collecting customs that already exist, the papers in this approach create more than one convention among artificial agents and arrange for several of them to survive. One lever is the communication network, applied at run time to already-trained agents. Who is allowed to talk to whom decides whether a population of language models settles on a single shared name for things or splits into cliques that each keep their own, even when everyone talks every round. The other lever is the payoff, applied during training. An agent is rewarded for coordinating well with copies of itself and badly with the conventions already discovered, so each training run has to invent a different way of doing the same task. The two differ in where the plurality ends up. In the first, one population carries several conventions at once. In the second, a library of separately trained policies each holds one convention. This library is built as a set of partners with different conventions, so that coordination with an unfamiliar human is reliable.

The method (highlighting differences with Customary pluralism)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.customary import Conventions
from aiag import LM
from aiag.customary import Conventions

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
conventions = Conventions(lm, k=3)  # each new convention must be unlike the ones already found

found = conventions("Two delivery robots meet in a narrow hallway and must decide who goes first.")
print(found)  # up to three workable conventions for the same task

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.

Prescription

AbsentOmission

Definition (Edmund Burke, Reflections on the Revolution in France, 1790)

Long usage itself confers authority. A practice that has organized life for generations carries compressed, tested wisdom, even when nobody can articulate its justification.

Analysis

Both mechanisms create conventions within a single run (in Saab and Abdallah (2026), a naming game played over rounds among open-weight agents; in Sarkar et al. (2023), a training loop). As a result, nothing has been in use long enough for use itself to give it standing. Neither paper treats a surviving convention as better for having survived. The winning label in the naming game is arbitrary. Sarkar et al.'s (2023) own complaint about self-play is precisely that what it converges on is arbitrary, which is why it adds a penalty that forces different answers. Real human usage does enter at one point, when the trained policy is paired with a human partner. There, the agent conforms to the partner's existing convention to gain the coordination payoff, and not because long practice has given that convention authority.

The unwritten constitution

PartialShortfall · partially preserved

Definition (A. V. Dicey, Law of the Constitution, 1885)

Some of a country's most binding norms are never written down. For example, the UK's constitutional conventions bind governments without being codified anywhere. Being binding and being codified are therefore different properties.

Analysis

The approach fully carries over the uncodified part of the concept. The orders never exist as a written rule. They exist only as emitted labels and first-token preference distributions in Saab and Abdallah (2026), and as weights that produce coordinated play in Sarkar et al. (2023). Agents coordinate on them without explicit communication. Dicey's other part is missing. Here, binding means only that an agent who deviates loses coordination reward. Nothing sanctions departure, and no agent holds another to the norm. When Saab and Abdallah (2026) rematches partners, the label history simply repairs coordination rather than enforcing anything on the defector. Both papers use the word convention and assume that it governs joint behavior without codification, which borrows the tradition's vocabulary. However, neither engages the distinction between binding and codified, which is Dicey's actual point.

Organic amendment

PartialNot claimed

Definition (Edmund Burke, Reflections on the Revolution in France, 1790)

Norms change gradually from within a community's practice, through adjustment rather than by decree from outside. The community itself is the author of these changes.

Analysis

In Saab and Abdallah (2026), the content of a convention is settled from within the population. Nobody designates which name wins, and agents converge on one through repeated local exchanges between pairs. When contacts are rewired, the retained partner-label history lets the group readjust step by step rather than be reset. This is the gradual, decentralized adjustment that Burke describes. However, everything the approach is actually about is set from outside. The routing graph decides whether the population unifies or splits apart. In Sarkar et al. (2023), the entire plurality comes from a cross-play term that pushes each new policy away from the ones already found. The designer adds a mixed-play correction when that term starts producing bad-faith play. No community holds a convention as its own, so nothing here is the author of its own drift. The drift is a controlled output of the interaction graph and the objectives.

Concept Analysis: Newly Introduced

Fragmentation as a control variable

Added

Customary traditions treat plural orders as a historical fact to be understood. Here, whether the orders stay plural is set by the experimenter's routing graph, i.e., by how the agents are connected. This makes plurality something an operator can produce or dissolve, which is a possibility no customary tradition contemplates.

Papers

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

Findings about it

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

Graph Feedback Controls Consensus and Clique Formation in Open-Weight Language-Model Populations

Samer Saab & Chaouki Abdallah, Jul 2026

arXiv:2607.12077AnalysisBuilt

Runs a naming game across open-weight agents with 1.1B to 32B parameters. It shows that similarity-based routing can isolate an emerging convention and sustain fragmentation, even when every agent interacts in every round. Matched controls rule out uneven participation and model-family effects as explanations.

CodeSimilarityRouting
from aiag.customary import SimilarityRouting

# feedback=1.0 pairs each agent with the free agent whose recent names most resemble its own; 0.0 pairs at random
routed, mixed = SimilarityRouting(feedback=1.0), SimilarityRouting(feedback=0.0)  # lm=... for LLM agents
routed(60), mixed(60)

print(routed.cliques())  # -> {'C': 0.4166666666666667, 'A': 0.3333333333333333, 'B': 0.25}, cliques that keep their own names
print(mixed.cliques())   # -> {'B': 1.0}, one convention for everyone
  • SimilarityRoutingProcedureaiag/customary/customary_pluralism.py:111 Similarity-routed naming game: every round every agent plays, but each is paired with the free agent whose recent names most resemble its own, a graph feedback that can wall an emerging convention off into cliques that each keep their own (Saab & Abdallah, 2026).

pip install aiagEvery method in aiag Library on GitHub

Background

Work this approach builds on that falls outside the inclusion criteria (e.g., a single-objective baseline). It is listed for reference and not counted anywhere.

Diverse Conventions for Human-AI Collaboration

Bidipta Sarkar et al., Oct 2023

arXiv:2310.15414MethodBuilt

Trains a collection of agents that each learn a different way of coordinating in a cooperative game. Each agent is rewarded for playing well with copies of itself and badly with the ways of coordinating already found.

CodeCoMeDi
from aiag.customary import CoMeDi

comedi = CoMeDi()

# Each new convention maximizes self-play and minimizes cross-play with the closest one already found
comedi(k=3)
print(comedi.crossplay().round(decimals=1))  # -> 3.0 self-play on the diagonal, at most 1.0 cross-play off it
partner = comedi.partner()  # a policy trained to do well with every convention found
  • CoMeDiProcedureaiag/customary/customary_pluralism.py:56 CoMeDi: conventions trained one after another, each maximizing self-play while minimizing cross-play with the most compatible convention already found, plus mixed-play episodes that hand over from that partner to itself mid-game so it cannot buy low cross-play with a handshake that sabotages strangers (Sarkar et al., 2023).

pip install aiagEvery method in aiag Library on GitHub