Guiding emergent behavior in multi-agent LLM systems through information networks

Core question: when an operator can control which agent observes which agent but cannot change payoffs or messages, how should the observation structure be designed so that the group obtains higher collective welfare.

• Scenario and motivation: the authors say LLM agents increasingly act in groups (negotiation, resource allocation, coordination and competition), and their collective behavior is not directly designed by anyone; the interaction network (who transacts with whom) is usually exogenous and cannot be rewired, whereas the information network (who observes whom) is decoupled from the interaction network and is a lever still available to the operator.
• Shortcomings of existing governance tools: according to the paper, existing levers mostly act on the agents themselves or their coupling—personas and roles, fine-tuning strategies, communication protocols, incentives and payoffs; the interaction network, though the most studied, is often not modifiable. The authors say the information-network lever has received relatively little attention.
• Core observation (behavioral coupling): the authors propose that observation creates behavioral coupling—after seeing peers' past behavior, agents adjust their own choices more strongly based on it, so observation becomes a contagion channel; its welfare effect depends on whether individually attractive behavior is aligned with the collective target, so more information is not necessarily better.
• What the paper does: 10 identical LLM agents play 6 repeated group decision-making games (coordination, anti-coordination, social dilemma, discovery), and the manager adds directed observation links under a budget constraint; it asks two questions in turn—RQ1 how much information to share, RQ2 who observes whom under a fixed budget—plus an exploratory disclosure-timing analysis.
• Setup highlights: by default each agent sees only its own history and an anonymous summary of the group's previous-round results; links reveal only actions from completed rounds, not payoffs; the manager submits the network after round 1 and then fixes it. The paper says results from 1,080 episodes support one design rule: argue against default full transparency.

Multi-agent LLM systems

• Multi-agent debate by Du et al. (2024), role-based software teams by Hong et al. (2024) and Qian et al. (2024), the AutoGen conversation framework by Wu et al. (2024), social simulation by Park et al. (2023) and Gao et al. (2023); surveys in Wang et al. (2024), Xi et al. (2025), Guo et al. (2024).
• The paper notes that these systems treat messages as the coordination medium, whereas this paper closes the communication channel and takes only the observation network as the design object, in order to isolate the role of information structure from that of message content.

Emergent collective behavior in LLM populations

• Ashery, Aiello, and Baronchelli (2025) document emergent conventions and collective bias; Riedl (2026) studies emergent coordination; Ren et al. (2024) studies norm formation; Piatti et al. (2024) studies sustainable cooperation under scarcity; Maini, Goldstone, and Tiganj (2026) document differences between LLM and human group behavior.
• The paper states that these works establish that genuine emergence exists in LLM populations, and that this paper asks a complementary design question: how to shape the information substrate on which emergence rests.

Value of information and social learning on networks

• Classic results of Hirshleifer (1978) and Morris and Shin (2002): in strategic settings, more public information can reduce welfare.
• On the mechanism side, it draws on models of imitation and social influence on networks (DeGroot 1974; Golub and Jackson 2010; Friedkin and Johnsen 1990; Flache et al. 2017), and on findings that social influence can synchronize errors and harm collective accuracy (Lorenz et al. 2011; Centola and Macy 2007); also the exploration/exploitation trade-off on networks (Lazer and Friedman 2007; Mason and Watts 2012) and group decision-making research (Stasser and Titus 1985; Woolley et al. 2010; Roberts and Goldstone 2011).

Information network design for agentic AI

• Closest to this paper's framing is work treating information flow as a governance mechanism for multi-agent AI: Chen et al. (2026) adaptively regulate what agents can observe, and Ebrahimi et al. (2026) study how network opacity drives strategy escalation.

Baselines and benchmarks: this paper is not set up as proposing a new method and comparing it with baselines; the comparison is a condition comparison within the same controlled population—RQ1 uses K=0 (no information) as the baseline and K=90 (fully directed graph) as the other end, and RQ2 compares four topologies (hub, regular, expander, barbell) at E=40; the evaluation environment is six games built by the authors (sum_guessing, stag_hunt, bandit, pgg, threshold_pgg, el_farol, Table S1).

The authors position the paper as: transplanting value-of-information and social learning theory to LLM populations, giving a controlled, cross-game characterization of how information amount and observation topology map to welfare, with the temporal question as an exploratory extension.

Overall approach: treat the information network (who observes whom) as the sole manipulated variable, perform causal identification on a controlled population with a fixed interaction structure and fixed agent policies, and answer two questions in turn: information volume (RQ1) and topology (RQ2).

Problem setup and formalization

• N=10 identical agents play T=20 rounds; in each round agent i chooses action ati and receives payoff uti. By default it sees only its own history and an anonymous summary of the group's previous-round results, and never sees specific peers' actions.
• The manager chooses a directed network G=(N,E), where a link i→j means i observes j's past actions, with a budget of at most K links; links reveal only actions from completed rounds, not payoffs; the network is submitted after round 1 and then fixed.
• The objective is to maximize the group's discounted cumulative welfare W(G)=𝔼∑_t=1^Tγ^t−1∑_i u_i^t(G), with γ=0.95, i.e., the earlier high welfare is reached the better; for pure coordination games, proximity to the target is used as welfare.
• All agents use the same instance of the same open-weight model and the same base prompt, with only the information block differing; the authors say this is to attribute structure in collective behavior to the information network rather than to preference differences, and that the model is a "controlled stochastic policy substrate" rather than a representative sample of all LLMs.

Behavioral coupling mechanism

• Adopt the standard model of imitation on graphs (DeGroot 1974; Golub and Jackson 2010): each agent adjusts toward the average action of the peers it observes, while also being pulled by a feedback signal toward the game-specific target.
• The linear approximation is a_t+1=ηθ^⋆ 1+((1−η−κ)I+κ M)a_t, where κ is imitation strength, η is feedback strength, and M is the row-normalized observation matrix; the mixing rate is determined by the spectral gap of M, and the sign is determined by the alignment between individually attractive behavior and the collective target.
• This yields a unified hypothesis to be tested (rather than assumed): the welfare effect of information ≈ speed of behavioral mixing × the sign of whether the propagated behavior is aligned; the authors emphasize that the sign is an emergent property of the population under coupling, not derived from the game label.

Six games

• Grouped by alignment properties rather than by "cooperative/competitive" labels: relatively aligned are bandit (discovery, where exploration is a public good), sum-guessing, and stag-hunt (coordination); misaligned are pgg (public goods dilemma, free-riding) and el-farol (anti-coordination/congestion); threshold-pgg is a hybrid of dilemma and coordination; stag-hunt is a deliberately saturated control (under a consensus threshold almost any disclosure can coordinate, and the information-volume gradient is flat by construction).
• At each round's choice, others' current actions are not visible, and the manager's links are the only channel for learning others' behavior; the full action space, payoffs, feedback, and verbatim prompts are in the SI (Table S1 and §Game Specifications).

Experimental design

• RQ1 (information volume): randomly add links and sweep the budget K ∈ {0, 10, 25, 40, 55, 70, 90}; K=0 is the no-information baseline, and K=90=N(N−1) is the complete directed graph; random edge addition separates quantity effects from structure effects.
• RQ2 (topology): fix E=40 directed edges and compare four structural extremes—hub and regular (maximum/minimum degree heterogeneity), barbell (two communities plus one bridge, maximum modularity), and expander (good mixing, minimum modularity); the spectral gaps are 0.34/0.44/0.40/0.07 and the modularities are 0.12/0.19/0.10/0.45. The authors say that because structural metrics covary across the four topologies, these serve only as locally fixed joint probes, with no single-metric attribution.
• Exploratory temporal analysis: fix the total edge-round budget and compare three disclosure schedules—constant, front-loaded, and back-loaded—at two budget anchors (K=25, 70); because the effects are statistically fragile, this is reported only as a supplementary finding.

Data scale and statistics

• 1,080 episodes in total: seven budget levels for RQ1, four topologies for RQ2, and the temporal analysis (2 schedule comparisons × 2 anchors), each crossed with 6 games and 12 seeds; model qwen2.5:7b, temperature 0.7, locally deployed, about 216k model calls; all figures show the mean ±1 SEM over 12 seeds.
• Tests match the questions: RQ1 uses Spearman trend and Cliff's δ (Table S2); RQ2 uses a Kruskal–Wallis omnibus plus per-topology rankings (Table S4); the temporal analysis uses seed-paired bootstrap CIs, Benjamini–Hochberg FDR correction, and TOST equivalence tests (Table S6). The authors explicitly write that sample size is a genuine limitation and report by cell.

Experimental setup

• Model under test: only qwen2.5:7b (temperature 0.7, 256-token limit, local deployment via Ollama, NVIDIA A100); all 10 agents are identical instances of the same model.
• Evaluation environments: six self-built games (Table S1)—sum_guessing, stag_hunt (coordination), bandit (discovery), pgg, threshold_pgg (dilemma), el_farol (anti-coordination); N=10, T=20, γ=0.95.
• Conditions: RQ1 sweeps K ∈ {0,10,25,40,55,70,90}; RQ2 compares hub/regular/expander/barbell at E=40; the temporal analysis compares constant/front-loaded/back-loaded schedules at K=25, 70.
• Metrics: group discounted cumulative welfare (mean ±1 SEM); statistical tests are Spearman r, Jonckheere–Terpstra, KW omnibus, Cliff's δ (Table S2), and cross-game sign separation uses a game-level one-sided Mann–Whitney test.
• Sample size: 12 seeds per cell, 1,080 episodes in total, about 216k model calls; the paper reports no cross-model replication.

Main results

Discounted welfare at extreme budgets and trend tests for each game in RQ1 (per Table S2):

Swipe horizontally to read the table

Game	Category	W(K=0)	W(K=90)	Spearman r (p)	Cliff's δ	Verdict
bandit	Discovery	1823	2031	+0.89 (<.001)	+1.00	sig. (+)
sum-guessing	协调	−8640	−5665	+0.35 (.001)	+0.44	sig. (+)
stag-hunt	协调	2082	2408	+0.06 (.58)	−0.17	ns（对照）
pgg	困境	9299	5514	−0.60 (<.001)	−0.85	sig. (−)
threshold-pgg	困境	3628	2793	−0.24 (.029)	−0.78	sig. (−)
el-farol	反协调	192	125	−0.54 (<.001)	−0.62	sig. (−)

• 作者报告符号分离在全部六个博弈上成立：对齐博弈 r 非负、错位博弈 r 为负（game-level Mann–Whitney 单侧 p=0.05，3 对 3 博弈下可达的最小值），六个中五个单独达到统计显著；作者将其读作单一受控 LLM 人群内的跨博弈经验规律，而非跨模型不变的定律。
• Figure 1 显示每条曲线都非单调（"更多不是更好"），作者称报告的是整体跨博弈秩模式而非逐点单调；threshold-pgg 呈浅谷：从 K=0 的 3628 降至 K≈40 的 3057，K=55–70 恢复到 3572，K=90 降至最低 2793，作者解读为搭便车传染与阈值协调两股力量交替占优，且中段种子间 SEM 约增为四倍。
• stag-hunt 是唯一不显著的格子，作者称其为按构造梯度平坦的饱和对照。

RQ2：固定预算下的拓扑

E=40 下四种拓扑的 KW omnibus 与折扣福利（据 Table S4，n=12 seeds）：

Swipe horizontally to read the table

博弈	KW p	hub	regular	expander	barbell
pgg	.0008	8870±248	8761±491	8032±526	6189±552
el-farol	.0037	163±10	124±10	151±15	110±9
threshold-pgg	.015	3567±28	3135±133	3425±39	3110±130
bandit	.019	2007±4	1987±9	1990±5	1976±8
stag-hunt	.076	2432±49	2431±24	2428±45	2536±20
sum-guessing	.79	−7590±1133	−7369±1288	−7181±1228	−6644±1294

• 作者报告拓扑只在三个错位博弈中显著，两个协调博弈不显著；bandit 名义显著（p=0.019）但跨拓扑福利差距不足 2%，作者称可忽略。
• 在架构起作用的错位博弈中，模块化 barbell 稳定是最差布线，作者称其相对最优拓扑损失 13–33% 福利；hub 在全部三个错位博弈中最好，但作者称这一次级趋势不如 barbell 惩罚干净。
• 作者指出 barbell 惩罚的符号与人类演化博弈论的网络互惠直觉相反，并给出解释：网络互惠依赖基于适应度的选择，而 LLM 智能体是较软的反馈驱动耦合，模块化"孵化区"在错位博弈中放大的是有害行为；因四个结构指标共变，作者只报告 KW omnibus 与 barbell 对比，不做单指标回归（Table S5、§Structural Metrics）。

机制的直接检验

• 论文在决策日志层面检验行为耦合假设（Eq. S1，Table S7、Fig. S6）：以一步动作变化对观察同伴上一轮平均动作、非同伴上一轮平均动作和自身上一轮动作做 OLS 回归，标准误按 run 聚类（每博弈约 26k 智能体-轮、156 次运行，限 0<K<90）。
• 作者报告被观察同伴系数 bobs 在全部六个博弈中高度显著，且 bobs−bnon 均为正并显著（p 从 8×10⁻⁴ 到 2×10⁻⁵⁹）；五个模仿型博弈 bobs>0 (stag-hunt 0.66, sum-guessing 0.49 strongest), and in el-farol bobs=−0.10 is negative, which the authors call the avoidance rewarded by the anti-coordination game, with direction consistent with the mechanism.

Exploratory temporal analysis

• Fixing total edge-round exposure, compare constant (A), front-loaded (B), and back-loaded (C) schedules at K=25, 70 (Table S6): none of the B−C comparisons in the 12 cells survived BH-FDR correction, and the authors present this as an exploratory finding rather than a formal conclusion.
• The authors name three raw effects: el-farol is harmed by back-loading under both budgets (C−A=−41, p=0.005; −51, p=0.009); threshold-pgg at K=70 is harmed by any deviation from constant disclosure (B−A=−844, p=0.002; C−A=−637, p=0.012); bandit at K=70 slightly prefers front-loading (B−A=+16, p=0.016, magnitude <1%). The authors say the two dilemmas point in inconsistent directions, and timing, unlike information volume, does not appear to be determined by strategy type.

Other ablations and analyses

• RQ1 mechanism (Table S3): under the random K sweep, after controlling for K, almost no single structural metric retains an independent association with welfare (only two nominal signals, pgg∼spectral gap partial r=−0.22, p=0.047 and el-farol∼concentration r=+0.26, p=0.016, neither passing correction).
• Temporal evolution (Fig. S1): the K-driven welfare gap and its sign widen gradually with rounds rather than appearing suddenly at a particular round.
• Supplementary temporal views (Fig. S2–S5): discounted and undiscounted cumulative, per-round trajectories, and endpoint bar charts; the authors say the conclusions are consistent across views—schedule differences are visible only in misaligned games and are all second-order effects.

Limitations and future directions noted by the authors

• Single model (Limitations (i)): the experiments establish information-network effects only within one controlled LLM population, not a law invariant across models; the authors propose that the contagion account predicts transfer to populations exhibiting similar observation-conditioned imitation, but cross-model transfer requires direct testing, and they suggest that even replicating RQ1 with a second model on a few representative games and budgets would tighten the scope of the regularity.
• Only one anti-coordination game (Limitations (ii)): the sign pattern on the congestion side is a single point rather than a family.
• Temporal results are exploratory only (Limitations (iii)): with 12 seeds per cell, no schedule comparison survived FDR correction, so it can be reported only with an equivalence caveat, and can be treated neither as a confirmed effect nor as a confirmed absence of effect.
• Structural metrics covary (Limitations (iv)): at N=10 multiple structural metrics move together, so only the KW omnibus and the barbell comparison are reported, with no single-metric attribution.
• Future directions (end of Limitations): the authors say each of the above is a natural target for higher-repetition, multi-model follow-up work; the Synthesis section also proposes that the temporal results could motivate adaptive, state-dependent disclosure-timing interventions.

Scope of experimental coverage

• Only one model under test, qwen2.5:7b (temperature 0.7, 256-token limit), with all 10 agents being identical instances of the same model (§Parameters, Configuration, and Computing).
• Evaluation environments are six self-built games: two coordination, one discovery, two dilemma, one anti-coordination (Table S1).
• Condition coverage: seven budget levels for RQ1, four topologies for RQ2, and two budget anchors × three schedules for the temporal analysis; 12 seeds per cell, 1,080 episodes in total (§Experimental Design).
• Observed content includes only peers' past actions, not payoffs or current actions; there is no messaging channel between agents (§Setup and Objective).
• The paper reports no replication on any other model family; the authors explicitly write that transfer across model families is a separate empirical question (§Setup and Objective, Limitations).

An analytical paper on multi-agent LLM governance: it decouples the information network (who observes whom) from the fixed interaction network and treats it as a governance lever that operators can design.

• Problem: the collective behavior of agent populations is not directly designed, and the interaction network and incentives are often unchangeable; the authors ask how much information should be shared and who should observe whom when operators can control only the observation structure.
• Method: 10 identical qwen2.5:7b agents play 6 repeated games for 20 rounds, and a manager adds directed observation links that reveal only past actions under a budget; seven budget levels are swept (RQ1), four topologies are compared at E=40 (RQ2), totaling 1,080 episodes with 12 seeds per cell; the mechanism is that observation creates behavioral coupling, and the welfare effect ≈ mixing speed × sign of alignment.
• Information amount results: sign separation holds in all six games—welfare rises with budget in aligned games and falls in misaligned games (Table S2); pgg has the highest welfare at K=0 (9299), dropping to 5514 under full disclosure; stag-hunt as a saturation control is not significant.
• Topology results: at a fixed budget, topology matters only in misaligned games (pgg KW p=0.0008), and the modular barbell is worst in all three misaligned games; the authors state it loses 13–33% welfare and that this is contrary to the network reciprocity intuition from human evolutionary game theory (Table S4).
• Mechanism and additional results: log-level regressions show agents respond more strongly to peers they actually observe, with direction consistent with game type (Table S7); no effect of disclosure timing survives FDR correction, and it is reported only as exploratory (Table S6).
• Authors' conclusion: information is a real governance lever, but its direction and magnitude depend on the game and the budget; before designing, one should first ask what behavior greater visibility will amplify, and therefore they argue against full transparency as the default for agent collectives; the authors also emphasize that this is an empirical regularity within a single controlled population, and transfer across models requires separate testing.