Steering Emergent Behavior in Multi-Agent LLM Systems via Information Networks
Authors: Sepehr Ilami, Babak Heydari
Organizations: Department of Mechanical and Industrial Engineering, Northeastern University, Boston, MA, USA · Network Science Institute, Northeastern University, Boston, MA, USA
Large language model (LLM) agents increasingly operate in groups, each pursuing an individual objective while jointly determining system performance. If an operator controls which agent observes which other agent, but not their payoffs or messages, how should observations be structured to raise collective welfare? We make the information network (who observes whom) a design object, decoupled from the fixed interaction network over which agents act. Ten identical LLM agents play repeated group-decision games; by default each observes only an anonymous summary of the group's previous-round outcome, and a manager with a fixed budget adds directed observation links that reveal peers' past actions. Across 1,080 episodes and six games spanning coordination, anti-coordination, social dilemmas, and discovery, we ask how much information to share and, at a fixed budget, who should observe whom. Disclosure creates behavioral coupling: observation acts as a contagion channel whose welfare effect depends on the alignment between individually attractive behavior and the collective objective. First, the value of information is alignment-contingent: more peer information tends to help or be neutral when individual behavior is aligned with the collective goal, but under free-riding or congestion externalities it amplifies collectively harmful behavior; in a public-goods game the group does best with no information at all. We read this sign separation as a cross-game regularity within one controlled LLM population, not a model-invariant law. Second, at a fixed budget, who observes whom matters primarily when coupling is harmful: a modular "barbell" that lets sub-populations entrench harmful behavior is reliably the worst wiring, reversing the network-reciprocity intuition from human evolutionary game theory. These results argue against a default of full transparency for agent collectives.
Figures & tables
Figure 1: RQ1 — is more information better? Discounted welfare (mean ±1 SEM over 12 seeds) against the information budget K . Top: relatively aligned games—more information helps. Bottom: misaligned games (free-riding and congestion)—more information hurts. All effects are non-monotone; the sign tracks the alignment between individually attractive behavior and the collective objective. Full trend statistics are in Table S2.
Figure 2: RQ2 — who observes whom, at matched budget ( E=40 ). Top: the four canonical topologies (undirected skeletons) with spectral gap and modularity Q . Bottom: welfare relative to each game’s mean (mean ±1 SEM); misaligned games (left) vs. aligned games (right). The modular barbell (red) is the worst wiring for the misaligned games; for the aligned games topology is effectively irrelevant. Per-game rankings and metric profiles are in Tables S4–S5.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Game
Class
Action
Payoff / welfare
System-level feedback (peer-anonymous)
sum_guessing
coordination (coop.)
integer 0 – 50
−∣sum−θ⋆∣ (proximity to a hidden target θ⋆∈[0,500] )
the group sum was too high / too low / correct
stag_hunt
coordination (coop.)
stag / hare
stag 20 , hare 12 , failed stag 0 ; hunt succeeds only if all 10 choose stag
the hunt succeeded or failed, and how many chose hare
bandit
discovery (coop.)
explore / exploit
exploit safe 10 ; explore H−8 (hidden H=24 ), value shared across all agents
the best reward discovered so far, or that no arm is yet explored
the shared pool’s total and its shortfall below the maximum
threshold_pgg
dilemma (comp.)
contribute 0 – 20
0 until cumulative ≥800 , then reward 400 ( R/N=40 /agent) split every remaining round
the threshold was met (unlocked), or how many tokens short
el_farol
anti-coordination (comp.)
bar / home
bar 10 if uncrowded, home 1 , bar −2 if crowded ( ≥7/10 attend)
the bar was overcrowded or under capacity, and the attendance
Appendix
Table S1: The six games. Each isolates a distinct strategic tension. By default an agent sees only its own history and the peer-anonymous feedback in the final column—never a specific peer’s action or payoff.
game
class
W(K=0)
W(K=90)
shape
Spearman r ( p )
JT p
Cliff δ
verdict
bandit
discovery
1823
2031
increasing
+0.89 ( <.001 )
.0001
+1.00
sig. ( + )
sum-guessing
coordination
−8640
−5665
increasing
+0.35 ( .001 )
.001
+0.44
sig. ( + )
stag-hunt
coordination
2082
2408
rise then plateau
+0.06 ( .58 )
.44
−0.17
ns (control)
pgg
dilemma
9299
5514
peak (best at K=0 – 10 )
−0.60 ( <.001 )
.0001
−0.85
sig. ( − )
threshold-pgg
dilemma
3628
2793
valley ( K≈40 ); net −
−0.24 ( .029 )
.013
−0.78
sig. ( − )
el-farol
anti-coord.
192
125
peak then decline
−0.54 ( <.001 )
.0001
−0.62
sig. ( − )
Appendix
Table S2: RQ1 (amount): welfare-vs-budget trend tests. Discounted welfare at the extreme budgets, the signed curve shape, Spearman r (with p ), Jonckheere–Terpstra trend p , and Cliff’s δ between lowest and highest K . Sign is ≥0 for every relatively aligned game and <0 for every misaligned game.
spectral gap
concentration
overlap
game
partial r ( p )
partial r ( p )
partial r ( p )
bandit
+0.10 ( .35 )
+0.12 ( .27 )
−0.05 ( .64 )
sum-guessing
+0.12 ( .27 )
−0.08 ( .49 )
−0.07 ( .50 )
stag-hunt
−0.09 ( .40 )
−0.17 ( .12 )
−0.07 ( .53 )
pgg
−0.22 ( .047 )
+0.15 ( .17 )
−0.09 ( .44 )
threshold-pgg
−0.11 ( .31 )
−0.16 ( .16 )
−0.05 ( .67 )
Appendix
Table S3: RQ1 mechanism: partial correlations of welfare with three directed structural metrics, controlling for the budget K , under the random- K sweep. Holding density fixed removes almost all single-metric structure signal, motivating the fixed-budget architecture (RQ2) and timing designs.
Figure S1: RQ1 (time-resolved): cumulative discounted welfare over (round, K ) per game (mean over 12 seeds). The K -driven welfare gap, and its sign, opens progressively over the episode: aligned surfaces (top) tilt up with K , misaligned surfaces (bottom) tilt down. The round- 20 slice of each surface is the corresponding curve in main-text Fig. 1 .
KW
hub
regular
expander
barbell
best / worst
game
p
(deg-het max)
(deg-het min)
(mod min)
(mod max)
pgg
.0008
8870±248
8761±491
8032±526
6189±552
hub / barbell
el-farol
.0037
163±10
124±10
151±15
110±9
hub / barbell
threshold-pgg
.015
3567±28
3135±133
3425±39
3110±130
hub / barbell
bandit
.019
2007±4
1987±9
1990±5
1976±8
(negligible, <2% )
stag-hunt
.076
2432±49
2431±24
2428±45
2536±20
ns
Appendix
Table S4: RQ2 (architecture): Kruskal–Wallis omnibus across the four architectures and mean ± SEM discounted welfare of each ( n=12 seeds). Bold marks the best (hub) and worst (barbell) in the misaligned games. Architecture is a lever only in the misaligned games; the barbell is uniformly the worst wiring where it matters.
architecture
clus.
deg. var.
Gini
mod. Q
assort.
sp. gap
hub
0.60
6.40
0.34
0.12
−0.66
0.34
regular
0.50
0.00
0.00
0.19
n/a
0.44
expander
0.65
3.60
0.27
0.10
−0.29
0.40
barbell
0.86
0.20
0.05
0.45
−0.01
0.07
Appendix
Table S5: RQ2: structural-metric profile of the four architectures at E=40 directed edges (20 undirected). Architectures realize, respectively, degree-heterogeneity max (hub) / min (regular) and modularity min (expander) / max (barbell). Classical metrics (clustering, degree variance, degree Gini, modularity Q , degree assortativity) are computed on the undirected skeleton; the spectral gap is the directed mixing rate of Eq. ( 2 ) in the main text. The barbell’s high modularity co-occurs with the lowest spectral gap—a mixing bottleneck.
game
K
A
B
C
B − A ( p )
C − A ( p )
B − C ( q )
TOST equiv.?
bandit
25
1976
1962
1969
−14 ( .064 )
−7 ( .27 )
−7 ( .47 )
yes
bandit
70
2014
2030
2016
+16 ( .016 )
+2 ( .97 )
+14 ( .19 )
yes
sum-guessing
25
−7545
−7910
−7432
−365 ( .85 )
+112 ( .79 )
−477 ( 1.0 )
no
sum-guessing
70
−6239
−6150
−6407
+89 ( .91 )
−168 ( 1.0 )
+257 ( 1.0 )
no
stag-hunt
25
1857
2019
1247
+163 ( .65 )
−609 ( .10 )
+772 ( .19 )
no
stag-hunt
70
2411
2405
2353
−6 ( .09 )
−58 ( .75 )
+52 ( 1.0 )
yes
Appendix
Table S6: Timing analysis: mean discounted welfare for constant (A), front- (B), and back-loaded (C) disclosure at matched budget, with seed-paired contrasts. p -values are permutation tests over 12 seed-pairs; q is BH-FDR over the 12 -cell family for the B − C contrast. No B − C contrast survives FDR. The only mechanism-consistent effect is that anti-coordination ( el-farol ) is hurt by delay (C − A <0 at both budgets).
Figure S2: Timing (discounted cumulative): running (cumulative) discounted welfare (mean ±1 SEM over 12 seeds) under constant (A), front-loaded (B), and back-loaded (C) disclosure at matched budget ( K=70 solid, K=25 dashed); the shaded round-1 column is the homogeneous baseline. This is the discounted running total the manager optimizes. In the coordination and discovery games (top) the three schedules are nearly indistinguishable; separation appears only in the public-goods and congestion games (bottom) and, at the low budget, in stag-hunt . Per-round trajectories are in Fig. S3 and full contrasts in Table S6 .
Figure S3: Timing (per-round): per-round welfare (mean ±1 SEM over 12 seeds) under constant (A), front- (B), and back-loaded (C) disclosure at matched budget ( K=70 solid, K=25 dashed). The shaded round-1 column is the homogeneous baseline; dotted lines mark the window boundaries (rounds 2 – 7∣8 – 14∣15 – 20 ). Within its active window a schedule briefly separates, but the per-round differences are small and largely cancel; the discounted cumulative view (Fig. S2 ) is the integrated result.
Figure S4: Timing (undiscounted cumulative): running, undiscounted total of per-round welfare through round t (cumulative sum of the per-round welfare), for constant (A), front-loaded (B), and back-loaded (C) disclosure at matched budget ( K=70 solid, K=25 dashed; mean ±1 SEM over 12 seeds). This is the un-discounted companion to Fig. S2 , which shows the discounted cumulative welfare the manager optimizes. Within each game panel the three schedules share the same accrual baseline, so any A/B/C separation is the pure timing effect; per-panel y -axes keep that separation visible despite the common downward accumulation of the negative-welfare games. As in every other view, the schedules nearly coincide except in the misaligned games (bottom), confirming that timing is second-order and that this is not a discounting artifact.
Figure S5: Timing (endpoint summary): final (round- 20 ) discounted welfare by disclosure schedule — constant (A, the reused RQ1 run), front-loaded (B), back-loaded (C) — at each budget ( K=25 left group, K=70 right group) within every game panel (mean ±1 SEM over 12 seeds). This collapses the trajectories of Fig. S2 and Fig. S3 to the single quantity the manager optimizes, so where timing is second-order the three bars are equal, and where it moves the outcome a gap appears; the per-cell mark is a Kruskal–Wallis test across the three schedules ( ∗=p<0.05 , “ns” otherwise). The bars are the graphical form of the A/B/C columns of Table S6 : the few visible gaps ( stag-hunt at K=25 , el-farol , threshold-pgg at K=70 ) are exactly the cells flagged there, and none survive FDR correction.
game
class
bobs ( p )
bnon ( p )
bobs−bnon ( p )
bandit
aligned
+0.16 ( 10−15 )
+0.07 ( 10−8 )
+0.09 ( 10−3 )
sum-guessing
aligned
+0.49 ( 10−50 )
+0.14 ( 10−7 )
+0.35 ( 10−10 )
stag-hunt
aligned
+0.66 ( 10−73 )
−0.04 ( 10−2 )
+0.70 ( 10−59 )
pgg
misaligned
+0.38 ( 10−38 )
−0.01 ( .57 )
+0.39 ( 10−22 )
threshold-pgg
misaligned
+0.49 ( 10−31 )
+0.10 ( 10−4 )
+0.39 ( 10−9 )
el-farol
misaligned
−0.10 ( 10−11 )
−0.46 ( 10−124 )
+0.35 ( 10−30 )
Appendix
Table S7: Direct coupling test: OLS coefficients of Eq. ( S1 ), per game, SEs clustered by run ( ∼ 26k agent-rounds and 156 runs per game; 0<K<90 ). bobs is the response to observed peers’ prior actions, bnon to non-neighbors’, controlling for the agent’s own prior action. bobs is significant and bobs>bnon in every game; the sign of bobs is positive (imitation) except in the anti-coordination game el-farol , where it is negative (avoidance).
Figure S6: Direct coupling test: per-game response coefficients to observed peers ( bobs , green) vs. non-neighbors ( bnon , grey) from Eq. ( S1 ), ± 95% CI (SEs clustered by run). The observed-peer response dominates in every game; it is positive (imitation) everywhere except the anti-coordination game el-farol , where it is negative (avoidance)—the game-appropriate sign of coupling.
The performance of multi-agent LLM systems depends not only on the capabilities of the underlying models but also on information-sharing policies, which govern what information each agent observes and at what level of detail. But does providing AI agents with more detailed information always improve collective outcomes, or can transparency paradoxically degrade efficiency? To investigate this, we introduce empirical information design, an experimental framework that systematically compares disclosure policies across economic markets populated by LLM agents, while holding interaction rules and models fixed. We evaluate ten frontier LLMs in multi-resource trade, bilateral bargaining, and double auctions, using ground-truth utilities to decouple identified opportunities from realized outcomes. We find that finer information disclosure yields highly model-dependent results. In trade, revealing exact valuations rather than ordinal preference rankings reduces realized gains for GPT-6 Sol (from 98% to 91% of maximum) yet increases them for Claude Sonnet 5 (from 72% to 86%). In bargaining, exact information leads Grok 4.7 to produce less balanced agreements, whereas Claude Opus 5.5 divides gains symmetrically. In double auctions, full disclosure inflates price demands, triggering market breakdown and lowering welfare for Gemini 3.1 Pro (97% to 78%), Claude Opus 5.5 (94% to 64%), and Grok 4.7 (97% to 62%). Overall, these results suggest that model selection and information design must be co-optimized, as LLMs vary in how their strategic responses to information impact system objectives.
Mete Erdogan, Mirac Suzgun, Aneesh Pappu +7
Stanford University · All Purpose AI · University of Maryland, College Park
Large language model (LLM) agents are increasingly deployed in interacting populations, raising the question of what such populations come to believe collectively. Whether a population aggregates genuine knowledge or collapses into a false consensus directly affects how much such systems can be trusted. Classical social-network models assume that the network itself determines how beliefs combine. This assumption breaks down for LLM agents, whose limited attention takes in only part of what they are exposed to, so these models overstate how much information a population actually pools and cannot tell genuine consensus from herding. We introduce SNLA, a framework that models how much each agent actually influences others, rather than merely how the network connects them. This influence depends on each agent's position in the network and on how sharply attention focuses. Theoretically, we show on a tractable proxy that narrow attention causes herding, where the effective sample size stays bounded regardless of population size, while wide attention recovers wisdom-of-crowds behavior only when the exposure graph is undirected and degree-regular. Empirically, a controlled testbed validates these predictions directly, and the herding-wisdom transition reproduces on operator-controlled variants of three multi-agent LLM benchmarks.
Kaixuan Liu, Guojun Xiong, Weinan Zhang +1
Department of Computer Science Emory University Atlanta, GA 30322 · School of Computer Science Shanghai Jiao Tong University Shanghai, China
LLM-powered AI assistants acting on behalf of users can produce poor collective outcomes at scale. We introduce a framework for evaluating their emergent behaviour in social dilemmas, applied to three iterated games (Public Goods, Collective Risk, Common Pool Resource). We prompt each model to produce a natural-language strategy, then have the same model translate it into code. This aims to isolate strategic reasoning from input-parsing, enables pre-deployment inspection, and scales to populations of hundreds of agents. We propose three analyses: behavioural fingerprinting via exhaustive evaluation over opponent histories; self-play robustness across mixtures of a model's strategies with either a Selfish or Collective disposition; and cultural evolution under payoff-biased imitation. Applied to three state-of-the-art LLMs, we find substantial cross-model differences in self-play welfare, and that cultural evolution converges to low-welfare, Selfish-dominant equilibria in larger groups.
Richard Willis, Jianing Zhao, Yali Du +1
King’s College London London, United Kingdom · King’s College London · Turing Institute +1