Summary
Flint Ashery, Aiello and Baronchelli put homogeneous populations of live LLM
agents through a naming game. Random pairs each pick a name and are rewarded
for matching. Each agent sees only its own last five interactions, and the
authors state that the prompt does not say the agent is part of a population. In all four models, populations of 24 reach a
single shared name. With two names, single agents with empty memory show no
statistically significant preference, yet in three of the four models the
populations land on one of the two more often than chance. A memory-by-memory reading of one model's choices
shows the asymmetry appearing by an agent's third interaction. A population in
consensus is then exposed to a committed minority that always plays the other
name. The minority flips the whole population once it reaches a critical size.
That size depends on the model and on whether the population holds the
favoured name or the disfavoured one: one agent in 48 for one model, 16 in 24
for another.
This is the predecessor of the group-size paper,
by the same group, and it instantiates collective dynamics
at fixed N. It adds two things that the follow-up lacks. Its agents are live
model calls rather than cached policy tables, and it introduces a committed
minority: a fixed-strategy faction set by the experimenter, where flag game's
unconvertible agents arise from private evidence. It does not vary population size as an experimental
variable.
Method
Each agent's prompt describes a two-player game, the payoff (+100 for a match,
−50 for a mismatch) and the agent's last H interactions: its own name, its
partner's, the payoff and its running score. The model is asked, as an outside
observer, to predict a player's next action, and that prediction is taken as
the agent's move. The prompt does not say the agent is part of a population or
how partners are chosen. The name pool is W single letters, shuffled at every call.
Defaults are N=24, W=10, H=5. Agents are sampled live at temperature 0.5 with
top-K 10 (Flint Ashery et al. 2025).
Populations are homogeneous, and each run uses one model. A meta-prompting
check asks the model comprehension questions about the rules, the history and
the score. The authors report accuracy nearly always above 0.8. The exception is
Llama-2-70b-Chat, which falls below 0.8 on counting how often a name was played.
Individual bias is the share of each name chosen by an agent with empty
memory, tested against uniform (binomial for W=2, chi-square for W=10).
Collective bias is the share of runs that end in consensus on each name. For
W=2 the name a population favours is called "strong" and the other "weak".
Critical mass is measured from a population already in consensus, in which
every agent's memory holds only that name. Committed agents that always play
the other name are added. A flip is counted when 95% of the last 3N
interactions succeed on the new name, within 30 population rounds.
Llama-3-70B-Instruct is run at N=48 and H=3 under a different time rule. Ten
runs are made per condition, three for Llama-3-70B-Instruct.
Key results
Convention emerges. At N=24 and W=10, all four models' populations reach a
state in which every agent outputs the same name. The authors report that
consensus is established by population round 15 for every model except
Llama-2-70b-Chat. They report convergence at N=200 and at W=26 for
Llama-3-70B-Instruct, but those curves are figure-only. Once reached, a
consensus on either name holds for three of the four models. For
Llama-3.1-70B-Instruct only the strong name's consensus holds (Fig. S7).
Collective bias without individual bias. At W=10, individual
Llama-2-70b-Chat and Claude-3.5-Sonnet agents show no significant preference
among the ten names (chi-square P=0.100 and 0.410), while both Llama-3 models
do. The authors report that consensus nonetheless concentrates on particular
names for all four, shown only as a distribution in Fig. 2A. At W=2 with
{Q, M}, no model's individual preference is significant (P=0.068, 0.116,
0.757 and 0.849 for Llama-3, Llama-3.1, Claude and Llama-2). Of 40 runs each,
populations converge on the strong name 40 times for both Llama-3 models, 36
for Llama-2 and 26 for Claude (Table S2). On this entry's arithmetic, Claude's
26 of 40 is not distinguishable from chance by a two-sided binomial test
(P≈0.08). The other three are. For Llama-3-70B-Instruct the empty-memory count
leans slightly, though not significantly, toward the name the population then
never chooses (2,435 against 2,565 of 5,000). Across four pools, Llama-3-70B-Instruct
populations converge on one name in 40 of 40 runs for {Q, M}, {X, Y} and
{Alice, Bob}, and in 24 of 40 for {F, J} (Fig. S6 caption).
Where the asymmetry comes from (Table 1, Llama-3.1-70B-Instruct). In the
first interaction, an agent picks M 50.8% of the time. In the second, averaged
over equally likely memories, it picks M 48.7% of the time. Both are reported as
neutral. After a success the agent keeps its name 99.4% of the time, and after a
failure it switches 97.3% of the time. By the third interaction the choice is no
longer symmetric under swapping the labels: from the memory {1: M, Q; 2: Q, M}
the agent picks M with probability 0.848, and from the mirror-image memory it
picks Q with probability 0.451. The aggregate probability of choosing M rises to
0.563. In the SI, the authors report that for the base Llama-3.1-70B model, run on
an initially neutral pair of random strings, strong collective bias is already
present at the second interaction. From this they say they cannot attribute the
collective bias to fine-tuning. That table did not survive conversion and is not checked here.
Committed minorities (Table S3). The text gives the extremes: a critical
mass as small as 2% (Llama-3-70B-Instruct) and as large as 67%
(Llama-2-70b-Chat). Table S3 prints agent counts. On this entry's
arithmetic, the shares are:
| Model | N | To flip consensus on strong name | To flip consensus on weak name |
|---|---|---|---|
| Llama-3-70B-Instruct | 48 | 6 (12.5%) | 1 (2.1%) |
| Llama-3.1-70B-Instruct | 24 | 10 (41.7%) | 0 |
| Claude-3.5-Sonnet | 24 | 5 (20.8%) | 5 (20.8%) |
| Llama-2-70b-Chat | 24 | 16 (66.7%) | 11 (45.8%) |
The zero for Llama-3.1 is the authors' report that a population in consensus on
its weak name switches to the strong name with no committed agents at all. The
weak consensus is not a steady state for that model (Fig. S7). Below the
critical mass, the authors report that the population settles into a mix
of the two names, since the committed agents keep playing theirs. The paper's
account, in dynamical-systems terms, is that the strong convention attracts
more configurations and holds them more firmly. Agents near a strong
consensus reach memory states that stop exploring, and agents near a weak one
keep exploring. In their introduction the authors cite prior critical-mass estimates
of 10–40% from theoretical models and 25% from human experiments. They do not
set their own thresholds against those figures.
Why it matters
The collective dynamics concept holds
that a population outcome can belong to no single agent. This paper is the
cluster's earliest live-agent demonstration of that for a convention. For W=2,
three of the four models show a collective preference that their single agents
do not. Table 1 locates where it comes from: label-asymmetric responses to
particular memories, which appear only once agents have interaction histories.
The group-size paper extends
the same design to amplification and reversal, to socially loaded word pairs and
to N from 1 to beyond 10⁴. It does so with cached policy tables. The live-agent
runs here are the evidence that the effect is not an artefact of that
surrogate, at N=24 and on letter names.
The committed-minority result is the part that is new to the cluster. A
population's convention can be overturned by a faction that never updates, and
how large that faction must be is a property of the model and of the name being
displaced, not a constant. The spread runs from about a fiftieth to two-thirds
of the population. On this entry's reading, the theoretical and human-experiment
estimates the authors cite in their introduction fall within that range. On this entry's reading, the asymmetry matters for safety
framing more than the extremes do. The convention a population settles into
unaided is the one hardest to dislodge (strong-name thresholds equal or exceed
weak-name thresholds for every model). A minority pushing toward the
population's own collective bias needs few agents, or none.
On the open disagreement in the concept's scope note, this paper sides with
consensus, and it does not test size. With no committed agents, every model's
population converges, at N=24, and by the authors' report at N=200. It has no
private evidence and no signed ties. The committed minority is the one
persistent heterogeneity it introduces, and below the critical mass it leaves
the population in a mixed state rather than a consensus. On this entry's
reading, that is consistent with the concept's reconciling reading:
non-consensus appears when some agents cannot be moved, not because the
population is large. It is not a test of that reading. There is only one
committed camp, against a majority that can update, and the manipulated
variable is the minority's share, not N. Flag game's large-N polarization comes
from two camps of evidence-induced zealots, each unable to convert the other.
Nothing here corresponds to that.
interpretive tensions
Are the committed agents LLM agents? The abstract describes the
committed minorities as groups of adversarial LLM agents. The methods describe agents that
"follow a fixed strategy and use the alternative convention at all times",
irrespective of history. On this entry's reading, nothing in the result depends
on a model generating their moves. What is measured is how the LLM majority
responds to a fixed-strategy faction.
Collective bias without individual bias rests on non-rejection.
"Unbiased" means the binomial or chi-square test fails to reject uniformity at
5%. Llama-3-70B-Instruct's P=0.068 and Llama-3.1's P=0.116 are near-misses.
The positive case is the claim that the population outcome is more lopsided
than the individual preference, which Table S2 supports for three models. For
Claude it is weak on this entry's arithmetic. The paper's own mechanism (Table
- is not an absence of individual bias in any case. It is an individual bias that is
conditional on memory and absent only from the empty-memory test.
"Strong" is defined by the outcome. The strong name is the one the
population converges on, and the critical-mass asymmetry is then reported
relative to it. The ordering is close to built in: the convention a population
reaches unaided being harder to leave is what makes it the population's
preference. Claude's equal thresholds (5 and 5) are the one case where the
asymmetry is absent. Its collective bias is also the weakest of the four.
Few runs, uneven protocols. The critical masses come from ten runs per
condition, three for Llama-3-70B-Instruct. That model is run at N=48 and H=3
under a different stopping rule, so its 2% is not on the same footing as the
other rows. Collective bias is 40 runs per model at a single N. No confidence
intervals are printed for either.
Forecasting a player, not playing. Each move is the model's prediction, as
an outside observer, of a player's next action. The follow-up uses the same
framing. The authors treat the forecast
as de facto play. Whether the same model addressed as the player would choose
the same way is not tested.
Letters, not norms. The conventions are single letters and random strings,
chosen to be meaningless. The authors list socially loaded conventions as
future work. The follow-up supplies them, at the cost of live agents.
concepts
- Collective dynamics — the
live-agent, fixed-N result for convention, and the concept's first
composition manipulation by faction: a committed minority whose critical
share depends on model and on which convention it displaces. It supports the
concept's claim that individual-level evaluation does not predict the
population outcome. It does not bear on the size dependence, which it does
not vary.
cross-references
- Group size effects and collective misalignment
— the follow-up, which cites this paper for the neutral-agents result and
uses its live-agent procedure as the reference when validating the policy
tables. It shares the game, prompt framing and payoff. It swaps live agents for cached
policy tables, letters for socially loaded word pairs, and fixed N for a
sweep. It has no committed-minority experiment. Llama-3.1-70B-Instruct is the
only model in both. - Physics of Agents — cites this paper as part
of the line asking whether LLM communities obey sociodynamic laws. Both find
consensus as the default at N=24 to 32. Physics of Agents gets there with signed
graphs and per-agent intrinsic fields, this paper with an unstructured,
identical population. - Flag game — the cluster's other population with
agents that cannot be converted. Flag game's zealots are produced by private
evidence and come in two opposed kinds. This paper's committed agents are set
by the experimenter and come in one kind. The paper does not test what two
opposed committed camps would do. - Attractor dynamics — adjacent, not
instantiating. The paper's basin-of-attraction language belongs to a population's
memory-state dynamics in an explicit game, not to the content of a
dialogue. The same boundary is drawn for the follow-up.
sources
- Flint Ashery, A., Aiello, L. M., & Baronchelli, A. (2025).
Emergent social conventions and collective bias in LLM populations.
Science Advances 11(20) eadu9368, 14 May 2025; preprint arXiv:2410.08948
(v1, 11 Oct 2024). The published main text was read via PubMed Central. The SI
was read from arXiv v2.