Summary
El, Paeng, Ganguli, Zou and colleagues run 32-agent communities of
language-model agents over a signed communication graph — each edge marking the
receiving agent as friendly or unfriendly to the sender — for eight rounds of
message exchange on either a MATH problem recast as a binary choice or a
political statement. Across 9,600 such communities (over 10,000 including the
generalization and asynchronous runs) the dynamics turn out to be compressible.
Fitting an Ising energy with Glauber updates — peer pressure along the signed
graph plus a per-agent intrinsic field regressed from persona and question
embeddings — predicts held-out one-step transitions at 75–86 balanced accuracy
and rollouts at 61–77, beats every baseline, transfers to graph families it was
never fitted on, and reproduces the population-level distribution of collective
outcomes. The fitted parameters then do interpretive work: the communities sit
below their own critical temperature, which is why conviction builds; the
concordant coupling dominates the discordant one, which is why they converge
rather than split; and neighbours currently holding the correct answer exert
more pull than neighbours holding the wrong one, which is why accuracy improves.
This is the wiki's second population-scale multi-agent entry and the second to
carry a statistical-mechanical treatment, arriving a month before the
flag game and cited in its own related-work
section. The two disagree in an informative way. Flag game varies population
size against controlled private evidence and finds polarization taking over as
N grows; this paper holds N at 32, varies the interaction graph instead, and
finds the repulsive channel too weak to sustain a split. Where flag game
supplies zealots who cannot be converted, this one supplies a measured coupling
asymmetry that makes consensus the generic outcome. The wiki still has no
concept for this cluster and this entry does not create one.
Method
Each community is N=32 agents with distinct personas, a shared binary question,
and a symmetric signed adjacency J ∈ {−1, 0, +1} with zero diagonal, drawn from
four graph families (random matrices, low-rank graphs, square lattices,
triangular lattices). On objective questions — competition problems from
MATH rewritten as two-option multiple choice — a persona is the worked
ground-truth solution to a different MATH problem, so agents differ in
expertise. On subjective questions — political statements from a
bias-evaluation dataset — a persona is a free-text demographic, political and
personality profile from TWIN-2K-500, so agents differ in preference.
Rounds are synchronous, t = 0…8. Each agent samples its vote five times and
averages, giving an opinion on a six-value grid in [−1, 1]; it then writes a
short message stating its position and one reason. Messages are routed by edge
sign into a friendly inbox or an unfriendly inbox, and updates are Markovian —
inboxes hold only the latest round, and no interaction history is carried. The
empirical analysis covers 9,600 communities: four backbone models × 60
questions × ten graphs × four episodes.
Two classification schemes. Individuals are typed by how many times they switch
sides: frozen (0), switcher (1), intermittent (2), oscillator (>2). Groups are
typed by their net opinion relative to a split band of ±0.2 — the resolution of
a single agent's five-vote average — at t=0 and t=8: persistent split and
convergence for communities with no initial majority, persistent majority,
divergence and majority switch for those with one. Conviction is the mean
squared opinion, which separates a near-zero net opinion produced by
polarization from one produced by indifference.
The model assumes agents move to reduce social pressure. An Ising energy over
the signed graph plus a personal-pressure term yields a logistic update whose
argument is the peer field plus an intrinsic field gᵢ, itself parameterized as a
linear function of concatenated persona and question embeddings. The single
coupling β is split into three — β⁺ on concordant edges, β⁻ on discordant
edges, β₀ for the bare fact of a connection — and fitted in two stages by
cross-entropy on observed one-step transitions. Because opinions rarely change
and the two labels are unbalanced, scoring uses balanced accuracy over the four
flip/stay × sign cells. Baselines: persistence, an interaction-free rule
(personal pressure only) and a mean-field rule (social pressure toward the
global average, graph structure discarded).
Key results
Communities do not merely lock in the initial majority. Frozen agents
dominate the individual archetypes, with oscillators second and switchers and
intermittents comparatively rare. At group level, divergence and majority
switch — the initial majority weakening or being overturned — reach 11–12% for
GPT-4o-mini and Qwen3.5-9B. Conviction rises over rounds in every model–regime
cell, with early rounds dominated by indifference and communities moving into
consensus or polarization as the run proceeds.
Interaction helps on objective questions and drifts rightward on subjective
ones. Over 6,400 objective group trajectories, incorrect→correct majority
switches outnumber correct→incorrect for all four models. On political
statements three of the four drift right across eight rounds: Gemma-3n-E4B from
75% to 96% right-leaning communities, Qwen3.5-9B from 52% to 67%, GPT-4o-mini
from 30% to 37%; Llama-3.1-8B-Instruct starts near 53% and ends marginally
lower. GPT-4o-mini is the sharp case — strongly left-leaning at t=0 and still
moving right under interaction, so the starting distribution and the direction
of drift are separate properties. A label-bias check on the label-balanced
objective questions finds Llama's share of +1 answers falling from 39% to 16%,
which cannot be truth-seeking.
Three couplings are what make the fit work. The three-coupling discrete
update is the best method in every column of the main table, at 75–86 one-step
balanced accuracy and 61–77 in rollout. With a single β the same rule drops
near chance in places (53.9 for Llama on subjective, 50.6 for Qwen on
objective). The mean-field baseline runs from the mid-50s to the low-70s; the
interaction-free baseline is near chance for three of four models, reaching
59–61 only for Gemma-3n-E4B, which the authors read as some of its agents being
predictable from persona and question alone. In-distribution and
out-of-distribution columns differ by at most 2.4 points for the three-coupling
rule. On three graph families held out entirely it is best in 15 of 16 columns
(85.0–97.8 one-step, 59.7–89.3 rollout), with the triangular lattice the
hardest case in rollout. A 64-agent mixed community — the same 32 personas
served once by GPT-5.6-sol and once by DeepSeek-V4-Flash — reaches 81.9 (70.4
rollout) on subjective and 80.0 (67.4) on objective questions.
The rollouts reproduce the outcome mix, not the individual. Sampling rather
than thresholding the fitted probability and rolling out from real initial
opinions, group-archetype shares deviate by about 3 points (objective) and 5
points (subjective) mean absolute deviation, with the largest single gap around
10 points. The individual oscillator count is inflated, and this largely cancels
at group level. The authors state the limit explicitly: the fitted dynamics
produce the right mix of outcomes, not the right archetype for a given group.
Four episodes differing only in sampling randomness frequently disagree on the
group archetype, which is the ceiling that motivates evaluating at the level of
the distribution.
Below the critical temperature. Sweeping 41 log-spaced temperatures from
0.05 to 20 and rolling out 500 steps, the variance of the absolute net opinion
peaks at a community-specific critical temperature marking the finite-size
analogue of a transition between the high-conviction regimes (consensus,
polarization) and indifference. The fitted operating point 𝒯=1 lies below that
peak in every dataset–model cell, which is the model's account of why
indifference decays and conviction builds.
Attraction beats repulsion, so consensus beats polarization. β⁺ exceeds β⁻
for all four models. The effective weight on a concordant edge runs 0.99 to
3.03 across models and question types, while the effective weight on a
discordant edge never exceeds 0.73 and is −0.47 (GPT-4o-mini, subjective) and
−0.01 (Qwen3.5-9B, subjective) in two cells. A stable split needs repulsion
comparable to attraction; the measured asymmetry says these communities do not
have it. The authors note the exception their own model allows: strong,
oppositely signed intrinsic fields can still sustain polarization.
Correct neighbours pull harder. A five-coupling variant splits each channel
by whether the neighbour holds the correct answer at the current step, so an
agent moves between the two groups as it changes its mind. In all four models
correct neighbours pull harder on the concordant channel and incorrect
neighbours push harder on the discordant one — and since discordant influence
is repulsive, both asymmetries point toward the truth. This is the fitted
account of the observed accuracy gain, and it is defined for objective
questions only.
Why it matters
The wiki's multi-agent findings divide into transmission channels — persuasion
in mind viruses, majority–minority affect
in chain-of-affective dynamics — and
population-scale accounts of what a collective does with what is transmitted.
The flag game opened the second group and this
entry is its second member, which is the main reason to file it now. It also
answers a different question than flag game does: not how many agents, but
how they are wired. The signed graph is the manipulated variable, and the
result is that its structure is recoverable as two numbers whose ratio decides
whether the community converges or splits.
The predictive claim is the load-bearing one. A fitted rule that generalizes to
unseen questions and to graph families it never saw, and that holds up on a
mixed community of two frontier models it was not fitted on, is a stronger
statement about compressibility than the wiki has for any other multi-agent
phenomenon. What it buys, on the authors' own framing, is anticipating a
multi-agent system's failure modes without running it at scale. What it costs
is stated in the same paper: the model discards message content entirely,
influence enters only as the sign of a neighbour's stance, and the rollouts get
the distribution of outcomes right while misstating how often a given agent
changes its mind.
Two results cut against comfortable readings. The truth-seeking asymmetry is
real but is a property of the setting, not the models: on objective questions
with verifiable answers, correct neighbours happen to pull harder. On subjective
questions the same machinery produces systematic ideological drift in three of
four models — the same mechanism, a directional bias amplified rather than an
error corrected. And the Llama label-bias result is a reminder that a community
can move together toward an answer option for reasons that have nothing to do
with the option's content. Any read of multi-agent deliberation as
self-correcting has to survive both.
interpretive tensions
Consensus-favouring here, polarization-favouring in the sibling. The flag
game reports polarized endpoints rising to
dominance as population grows; this paper measures couplings under which a
stable split cannot sustain itself and consensus is generic. The two are not
directly comparable — flag game sweeps N from 4 to 128 with an unsigned
protocol and evidence-induced zealots, this one fixes N=32 and supplies signed
edges — but they point opposite ways on the question a reader would most want
answered, namely whether scaling a swarm tends toward agreement or toward
camps. This paper's own escape hatch is the footnote that strong oppositely
signed intrinsic fields can sustain polarization, which is close to what flag
game's zealots are. Whether the two accounts reconcile at that point is not
settled by either paper.
Fitted parameters are not measured mechanisms. The three interpretive
claims — below critical temperature, attraction over repulsion, correct
neighbours pull harder — are readings of coefficients in a model fitted to
predict transitions, not interventions on the system. The paper says so in its
future work, where controlled interventions on edges, messages or initial
opinions are listed as the test of whether the learned parameters predict what
happens under perturbation. Until that is run, the parameters explain the
observations they were fitted against. This is the dimension on which flag
game, which predicts which agent matters and then patches it, is the stronger
design.
Whose social pressure? The personas on subjective questions are profiles of
real individuals, the questions are political statements, and the dynamics are
described in the vocabulary of opinion formation. The authors caution
explicitly against over-extrapolating from persona-conditioned agents to human
communities and call any resemblance a hypothesis. The caution is worth keeping
attached to the rightward-drift result in particular, which reads as a claim
about a population but is a claim about four models' conditional distributions
under a fixed prompt format.
A binary opinion is a narrow window on a psychological phenomenon. Every
result here is computed on a sign. Whether an agent's stated reason changed,
whether it conceded a point, whether it was persuaded or merely outvoted, is
invisible to the analysis by construction. The paper is candid that discarding
message content is its main limitation; the entry is filed on the understanding
that what is modelled is the stance trajectory, not the deliberation.
concepts
Collective dynamics — the population-structure result, and one of the concept's two mechanistic accounts. Filed concept-less while the editor held the concept decision; re-homed on 2026-09-24. Its apparent disagreement with the flag game about the generic outcome is carried in the concept's Scope note.
cross-references
- Persona selection — the intrinsic field
gᵢ is a persona's predisposition on a question, regressed from persona and
question embeddings, and the interaction-free baseline measures how far that
alone gets you: near chance for three of four models, 59–61 balanced accuracy
for Gemma-3n-E4B. Read against the concept, that is a negative result about
persona as a predictor of behaviour under social pressure rather than an
instantiation of the selection mechanism. - Sycophancy — the fitted concordant coupling is
conformity to peers with no user in the loop, and it dominates the repulsive
channel in all four models. But the paper never probes against known-correct
private evidence the way a sycophancy study would, and never uses the term;
the asymmetry is a fitted coefficient, not a demonstrated capitulation. Noted
as adjacent because the flag game's
memory-conflict probe raises the same peer-directed question with a design
that can answer it. - Attractor dynamics — adjacent, not
instantiating. The ordered low-temperature state a community settles into is a
convergence result, but it is convergence of a finite population under an
explicit signed protocol with a fixed question, not the destination of an
unconstrained dialogue. The mechanism is a coupling asymmetry that can be
measured and, in principle, changed by rewiring. - Mind viruses and chain-of-affective
dynamics — the wiki's transmission-channel
findings. This paper's messages are a channel too, but reduced to a sign; what
it adds is the topology the channel runs over. The chain-of-affective result
on majority–minority structure is the nearest precedent for treating
population composition as the variable. - Probine et al. fit population opinion trajectories of networks that include a Llama 3.1 8B model (not demonstrably the same checkpoint) and find a uniform innate bias dominating, which on the wiki's mapping is roughly the persona-independent part of this paper's intrinsic field: a non-social route to the agreement this paper reads from its couplings.
sources
- El, B., Paeng, J., Dinc, F., Su, S., Erdogan, M., Pappu, A., Ye, H., Zhao, W.,
Ganguli, S., & Zou, J. (2026). Physics of Agents: Statistical Mechanics
Predicts Collective Behavior of AI
Agents.
arXiv:2608.16578.