Summary
Piao, Lu, Gao, Xu, Hu, Santos, Li and Evans run about 1,000 GPT-3.5 Turbo
agents that argue about one US political issue at a time: partisanship, gun
control or abortion. Each agent decides every round whether to keep its current
conversation partner or be handed a random new one. Starting from a
centrally peaked opinion distribution, the neutral share collapses, the
population splits into a left camp and a right camp, and the interaction
network sorts itself into two like-minded communities. Two controls carry most
of the weight for this wiki. With the network held static or randomized, no
balanced split forms and one camp ends up holding about three quarters of the
agents. On a factual question (flat Earth) the population reaches consensus
within ten rounds.
This is a collective dynamics finding on
the structure side. It is the concept's first entry where agents rewire their
own network, and its first polarization result produced without controlled
private evidence from a network the agents themselves rewire. It is the polarization-side counterpart to
Physics of Agents. Its value for the concept's
open disagreement, between Physics of Agents and the
flag game, lies in its controls more than in its
headline. It varies network rewiring and question type, which both of those
hold fixed. Its static-network control is the nearest it comes to Physics of
Agents' fixed graph, and there one camp dominates, which is compatible with
consensus-favouring couplings; with the flag game it shares almost no
conditions (Interpretive tensions).
Method
Agents hold one of five opinions, from left to right, and self-report it.
Opinions are initialized near-Gaussian, with 40% neutral, on a Watts–Strogatz
graph with rewiring probability 0.001. Each round has three LLM-driven stages.
An agent writes a short message giving its reasons. It decides whether it
would enjoy continuing with its current partner; if not, it is assigned a
random new one. It then writes a persuasion message, and every agent updates
its opinion and reasons from what its connections sent. Message history with a
partner is carried forward. No rule for homophily, in-group preference or
opinion averaging is coded in. Main runs use GPT-3.5 Turbo at temperature 1,
zero-shot, for about 40 rounds (the paper gives no run length; this is read
from its t=35–40 and 36–40 windows).
A pairwise test pairs agents holding the same opinion for one round, so any
opinion change is not social influence. A "self-regulation" variant has agents
check each message, reason and update for consistency with their current
opinion and regenerate until consistent. The robustness checks are as follows.
GPT-4o, ChatGLM and Llama-3 run as full systems; GPT-4o and DeepSeek-V3 also
replace GPT-3.5 midway through a run. The checks add temperatures 0.5 and 1.5,
four initial graph families, a centralized start with 80% neutral, a
2,000-agent run of the self-regulated variant, immigration, and flat Earth. The individual-mechanism
experiments use 100 agents and instruct a varying share of them to show
selective exposure or confirmation bias. Five interventions are applied from
t=35 to t=40 on a polarized partisanship system.
The headline "polarization level" is mean distance from neutral. It does not
distinguish a split population from a unanimous extreme one, so the bimodality
claims rest on the reported distributions and camp shares. Several of the
statistical tests (those on mechanism and intervention effects) treat timesteps within one run as samples (n=5–6), and no replicate runs
are reported.
Key results
Neutrality collapses and camps form. The neutral share falls from 40% to
22.5% (partisanship), 0.4% (gun control) and 5.1% (abortion), and the rest
divide into left and right camps. Under GPT-3.5 the split is left-skewed. The
pairwise test traces this to individual behaviour: right-leaning agents paired
with like-minded partners sometimes switch left, and left-leaning ones do not
switch right. The authors call this self-inconsistency, and it is more frequent
among right-leaning agents, by 0.068 in proportion. Self-regulation
cuts the inconsistency measure by 9.4–52.2% across issues, and the resulting
split is balanced. GPT-4o and ChatGLM skew left too; Llama-3 skews right.
The network sorts. Same-camp interactions rise by 156.8–382.7% over a run.
Cross-camp contact makes up 7.4–10.2% of interactions. Agents rate same-camp
agents more favourably than opposing ones. Exposure to more extreme like-minded
neighbours is associated with larger moves toward the extremes; cross-camp
exposure does not reliably moderate, a pattern the authors match to the
backfire effect.
Structure decides the outcome shape. On static or random networks the
paper reports that "no balanced polarization pattern forms" and one camp takes
about 75% of agents. Lowering temperature to 0.5 cuts partner changes to 1.32%
of edges per round, less than half the rate at 1.0 or 1.5. Polarization and
clustering both drop with it. Making 10% of tie-severing decisions fail leaves
polarization nearly unchanged but substantially weakens community separation.
Issue type decides it too. On flat Earth every agent ends up strongly
opposed within ten rounds. Immigration polarizes like the other political
issues. A start with 80% neutral polarizes more slowly but does polarize. All
four initial graph families end in two communities.
Composition moves it. Seeding 10% of agents with chain-of-thought prompting
raises the moderate-or-neutral share from 8% to 38%. It lowers polarization by
15% and community modularity from 0.43 to 0.25. Neutral broadcast influencers
reduce polarization by 28.5% and non-neutral ones raise it. Agents instructed
to show selective exposure or confirmation bias raise polarization as their
share grows.
Interventions after the fact are small. On a system already polarized,
restricting agents to moderate opposing voices shifts opinions by about 2%.
Instructed open-mindedness and neutral influencers reduce polarization most, by
11.8% and 8.8%. Instructing agents to seek diverse partners barely moves a
network that has already settled.
Why it matters
The concept's simulation entries vary size, wiring and private evidence, but
none lets agents choose their own contacts. This one does, and
its static- and random-network controls suggest that choice as the condition
under which the runs split rather than tip to one side, from one control whose
issue and run count are not stated. That
adds a candidate to the concept's list of what can hold camps apart, alongside
the flag game's evidence-induced zealots and Physics of Agents' opposed
intrinsic fields: homophilic self-sorting of the network. That reading is the
wiki's, assembled from the paper's controls. The paper itself frames its result
as matching argument-accumulation models of human polarization.
It also gives the concept its clearest contrast between kinds of question in
one system. The same agents, prompts and network rules reach consensus on a
factual question and polarize on contested ones. That is the same split Physics
of Agents draws between objective and subjective questions, though there the
fitted couplings favour consensus in both.
The pairwise and self-regulation results are individual-level, and the paper
keeps them that way. The left skew comes from single agents drifting from an
assigned opinion toward the model's own lean, the authors call it
self-inconsistency, and reducing it leaves the split intact while rebalancing
it. The population-level claim is the split, not its direction.
interpretive tensions
Where it overlaps Physics of Agents. Both run live agents on political
questions, exchange reasoned messages, and condition agents on assigned
personas or stances. They differ on the variables that matter here. Physics of Agents uses 32
agents on a fixed signed graph for eight Markovian rounds. This paper uses
1,000 agents, an unsigned graph the agents rewire, carried message history and
about 40 rounds. Its static-network control is the nearest it comes to Physics
of Agents' fixed graph, and there one camp dominates, which is the outcome
consensus-favouring couplings predict. So the two do not conflict where their
conditions meet. But the control's graph is unsigned, its issue, backbone and
run count are not stated, and three quarters is a majority, not the ≥85%
consensus the flag game uses as a threshold. The drift directions also differ:
GPT-4o-mini drifts right under Physics of Agents' protocol, while GPT-3.5 and
GPT-4o skew left here. Different prompts and models make that a note, not a
contradiction.
Where it overlaps the flag game. Very little. The flag game's polarization
is between the truth and a rival, sustained by private evidence, and grows with
N from 4 to 128. This paper has no private evidence, and on its one factual
question it reaches consensus. Its N runs from 100 to 1,000, with a single
2,000-agent check (self-regulated variant) reported as unchanged, so it does not test the size claim.
Both papers get polarization from something that stops opposing agents from
converting each other: decisive evidence in the flag game, and network
insulation here. Whether these are the same mechanism, as the concept's
"persistent heterogeneity" reading would need, neither paper tests.
What "polarization" measures. The paper's polarization level rises with
extremity, so a population unanimous at one pole scores at least as high as an
evenly split one. Its interventions are scored on that metric. The bimodality
claims come from the distributions, and the static-network control shows the
metric's blind spot: a one-camp outcome there is reported as no balanced
polarization, not as low polarization.
Persona, not belief. Agents are told to take a stance, then write
persuasion to themselves and others. The left skew comes from agents not
holding the assigned stance. The self-regulation fix imposes consistency
through a regeneration loop that runs on more than 60% of opinion updates
(averaged across five issues). The wiki's reading is that that loop may itself
make opinions stickier, which could help sustain a split; the paper does not
test this, and its 0.5-temperature run, where agents keep their ties, shows
less polarization.
Evidence base. The paper reports no replicate runs, so the timestep-level
tests measure stability within one trajectory, not variation across runs. The
paper calls its intervention results hypothesis-generating. The SI's opening
ranking of interventions also differs from the main text (see the stub).
concepts
Collective dynamics. It instantiates the
population-structure side, and it is the first entry where the network is the
agents' own product. The outcome shape (split, one camp dominant, or consensus)
changes with rewiring and with question type while the model stays fixed.
cross-references
- Persona selection: adjacent, not
instantiating. Self-inconsistency is an assigned political stance giving way
to the backbone's default lean, and its direction follows the model (left for
GPT, right for Llama-3). The paper does not measure persona. - Group size and collective bias
and social conventions: those fix the
network as well-mixed random pairing and find convergence. Their results are
consistent with this paper's random-network control, where one side
dominates. - Sycophancy: not engaged. The influence here is
peer-to-peer among live agents, not a user or scripted peers. - Probine et al. — the individual-level drift this paper calls self-inconsistency is the kind of term their uniform bias fits in a different setup (fixed graphs, small open models). There, homophily is weighting by opinion distance on a fixed graph, not rewiring, and it adds nothing to held-out fit once the bias is present.
sources
- Piao, J., Lu, Z., Gao, C., Xu, F., Hu, Q., Santos, F. P., Li, Y., & Evans, J.
(2026). Emergence of polarization in networks of large language model
agents. Nature
Communications (accelerated article preview); arXiv:2501.05171, v1 January
2025, v3 read.