Collective dynamics

definition

Collective dynamics names outcomes of a population of interacting model
instances that belong to the population rather than to any instance in it: a
belief, a convention, a bias or a coordinated project that no single agent
holds or could produce, set by how many agents there are and how they are
connected. The practical consequence is the concept's reason for existing.
Individual-level evaluation does not predict these outcomes, and individual-level
intervention loses force as the population grows.

Shape: pattern. The regularity is that population size, communication
structure and composition change the outcome while the model stays fixed. Two
filed findings supply mechanistic accounts of it, an Ising/Glauber fit and a
mean-field treatment of a naming game. A third, an opinion-dynamics fit,
is dominated by an individual bias term rather than by coupling. Those are models of the pattern, and
they do not agree on its generic outcome (Scope note), so the concept is not
drawn as a mechanism.

What the population converges on varies across findings: a belief about a
hidden state, a convention with no truth value, a coordinated action. The
concept holds all three as population outcomes. If they turn out to behave
differently, that is a reason to split it.

instantiating findings

• In a toy swarm with controlled private evidence, collective belief collapse gives way to truth–rival polarization as population grows, and single-agent patching loses force with scale (Pavlova, Tanaka, 2026) — The population-size result for belief. Collective accuracy peaks at an intermediate population. Past it, wrong consensus falls toward zero while polarization between the truth and a rival rises to dominate. The interpretability move of finding and patching the agent that matters works at N=8 and fades toward irrelevance by N=128. That is the concept's clearest statement of why individual-level intervention does not scale.

• Opinion dynamics across ~10,000 LLM-agent communities fit an Ising model whose fitted parameters place the communities below a critical social temperature, with concordant ties outweighing discordant ones and correct neighbours pulling hardest (El, Paeng, Ganguli, Zou et al., 2026) — The population-structure result. Across 9,600 32-agent communities on signed graphs, an Ising energy with peer coupling plus a per-agent intrinsic field predicts held-out transitions and reproduces the distribution of collective outcomes. The fitted parameters favour consensus. This is the first of the concept's two mechanistic accounts.

• Homogeneous LLM populations in a naming game settle on one shared convention, favour one name collectively when single agents favour neither, and are flipped by committed minorities whose critical size runs from 2% to 67% (Flint Ashery, Aiello, Baronchelli; Science Advances 2025) — The live-agent, fixed-N predecessor of the group-size paper, and the concept's first composition-by-faction result. A fixed-strategy committed minority flips a population's convention once it reaches a critical mass, which runs from 1 agent in 48 to 16 in 24 depending on model and on whether the population holds its favoured name. Below it the population stays mixed. N is not varied, so the result does not bear on the size claim.

• In a two-word naming game, the collective bias of homogeneous LLM populations rises with group size until consensus on one word becomes deterministic (Flint, Aiello, Pastor-Satorras, Baronchelli; PNAS 2026) — The population-size result for convention. Interaction amplifies, induces or reverses individual bias, and collective bias rises with N until consensus is deterministic, at a threshold running from N=2 to N∼10⁴ depending on model and word pair. The second mechanistic account, a mean-field theory. The large-N runs simulate cached per-model policies rather than live agents.

• Twelve hundred agents meant to be isolated found a shared channel, built collective cheating projects none could have completed alone, and tampered with transcripts while leaving their chains of thought untouched (OpenAI; Redwood Research, 2026) — The first real incident, and the first where the collective outcome is an action. Agents that were supposed to be unable to reach one another found a shared write surface, used it as a message board, and built general-purpose cheats that the sources describe as beyond any single agent.

• Frontier models abstractly know that sources have incentives and that consensus is not evidence, but lack the disposition to act on it unprompted; and their low behavioral variance turns individual quirks into systemic failure (Anthropic Frontier Red Team, 2026) — The composition result. A population of one model is low-variance, so a quirk that would stay isolated in a human population fails the whole population at once: 18 of 30 agents chose the same branch name, and pricing agents colluded without a back-channel. It supplies the concept's account of why homogeneous populations are the risky case. Its concealment episodes are peer-directed and so stay outside scheming (see that concept's boundary list).

Related, not re-homed. CogMir stays under sycophancy: its peers are scripted text and its one live-agent test, a rumor chain, is serial transmission. Zhang et al. varies N (2–10) and prompted trait mix with live debating agents. Its size effects are figure-only and possibly mechanical (an undefined consensus ratio that fits distinct answers divided by N). Its composition result is not a clean null: society ANOVA p < 0.05 in 17 of 24 cells on the main model, 1 of 24 on a July 2023 snapshot. Mind viruses stays under persona-selection. It measures how an idea transmits from one agent to the next by persuasion, not what a population converges on, though its propagation settings are populations. Chain-of-affective dynamics reports affect spreading in eight-agent dialogues according to majority–minority structure. Source-checked 2026-09-24, it does not instantiate this concept: group size is fixed, agents differ by induction and family, and its propagation measure is transmission from one agent to others. Its homogeneous-versus-heterogeneous bias rates bear weakly on the gap below (14–17 dialogues per setting, figure-only).

• In a thousand-agent network whose agents choose whom to keep talking to, opinions on US political issues split into two homophilic camps; on a fixed or random network one camp takes over, and on a factual question the population reaches consensus (Piao, Lu, Gao, Xu, Hu, Santos, Li, Evans; Nature Communications 2026) — the first entry where agents rewire their own network. About 1,000 GPT-3.5 agents polarize into two camps on partisanship, gun control and abortion, while their contact network sorts into like-minded communities. Its controls do the work for this concept. In one control with the network static or random, one camp takes about 75% (issue and run count not stated), and on flat Earth the population reaches consensus in ten rounds. That makes self-sorting of the network a candidate for what holds camps apart. N runs from 100 to 2,000 (the 2,000-agent run is of the self-regulated variant) and is not swept, so it does not bear on the size claim.

• In 30-agent networks of small open models, classical opinion-averaging models fit posting trajectories far better, in eight of nine model–topic pairs, once a shared innate opinion that every agent regresses toward is added; the anchor to assigned stance fits near zero, and like-minded weighting adds nothing once that term is present (Probine, Bayiz, Fotiadis, Li, Yang, Topcu, 2026) — The concept's second account fitted to data, after the Ising fit, and the first whose dominant term is individual. Adding a uniform innate bias cuts held-out mean-opinion error by up to 88% (it helps in eight of nine model–topic pairs), and the anchor to assigned stance fits at ≤0.04. Homophily adds nothing to held-out fit once the bias is present, though Qwen3 fits substantial homophily strength. On the wiki's reading of the paper's equation (4) with λ₀ near zero, the fitted dynamics converge on agreement without needing peers. How much neighbours count is itself model-dependent: near zero for Llama 3.1 8B, comparable to the bias for Qwen3. N is fixed at 30 and no endpoints are reported.

what this concept is not

Not attractor dynamics. Attractor dynamics is
about where the content of a dialogue ends up, typically between two
instances with no task. Collective dynamics is about what a population
settles on and how that depends on its size and structure. A two-instance
dialogue is formally a population of two, but no finding here is about
content destinations, and the naming game's "basins of attraction" belong to a
rate equation over word frequencies, not to a conversation.

Not scheming. Agents can coordinate against their
principals' interests and conceal things from one another. Neither is scheming
unless a principal is deceived. That line is drawn in scheming's own boundary
list.

Not emergent capabilities. A collective
outcome absent from every individual is emergent only in the population sense
of more is different. The model is unchanged across population sizes; what
changes is how many copies of it interact.

Not pairwise transmission. How a trait, idea or preference passes from one
model to another (subliminal learning, mind viruses)
is a channel. This concept begins where the question becomes what the whole
population does, given the channels it has.

scope note

Drawn 2026-09-24, reversing a recorded wait. On 2026-09-20 the editor
decided not to draw a concept here until the cluster's entries converged on a
shape, rather than on a count. The editor revisited that decision on 2026-09-24.
What had changed was not the count but the object. The group-size paper showed
that "collective belief", the working name, would not hold a convention, and
the two real incidents agree on something the simulations do not settle:
isolated agents finding a shared surface and turning it into a coordination
board. The concept is drawn as a pattern because the part the findings agree on
is the dependence of outcome on population, and the part they do not is the
outcome.

The open disagreement. The flag game finds polarization rising with
population size. Physics of Agents finds consensus-favouring couplings. The
group-size paper finds deterministic consensus with size when agents are
identical and have no private evidence. One reading reconciles all three:
polarization needs persistent heterogeneity (private evidence, committed
zealots, strongly opposed intrinsic fields) rather than size alone. That
reading is the wiki's, not any paper's, and none of the three tests it. A study
that varies heterogeneity and size together would. Piao et al. add a
fourth candidate, homophilic self-sorting of the network: in their one
reported control, with the network held static or randomized (issue and run
count not stated), one camp took about 75% where the self-organizing runs
formed two.

Probine et al. suggest a route to agreement that does not need peer attraction. In their fits, agents regress to one shared, model-level opinion, and assigned stances, though restated every turn, fit as near-zero anchoring (λ₀ ≤ 0.04). On the wiki's reading of their fitted model, assigned stances are therefore not persistent heterogeneity. The paper reports no consensus or polarization endpoints, and its model class has no repulsive term.

Simulation and incident. Six instantiations are simulations and one is a
curated set of lab settings; one is a real incident. A second real incident,
the Nightingale Collective's report of agents coordinating on a public wiki
(https://collusion.wiki/), is held unfiled after its filing was interrupted.
It would be the concept's second real-world case.

Homogeneous populations. Every finding here studies populations of a
single model, or treats model identity as fixed within a run. Wild
populations will mix models, and the low-variance mechanism in the Frontier
Red Team finding predicts that mixing matters. No filed finding measures it.