Collective AI can amplify tiny perturbations into divergent decisions

Hajime Shimao; Warut Khern-am-nuai; Sung Joo Kim

cs.AIMar 10, 2026

Collective AI can amplify tiny perturbations into divergent decisions

Authors: Hajime Shimao, Warut Khern-am-nuai, Sung Joo Kim

Organizations: The Pennsylvania State University, Malvern, PA, USA · McGill University, Montreal, QC, Canada · American University, Washington, DC, USA

Abstract

Large language models are increasingly deployed not as single assistants but as committees whose members deliberate and then vote or synthesize a decision. Such systems are often expected to be more robust than individual models. We show that iterative multi-LLM deliberation can instead amplify tiny perturbations into divergent conversational trajectories and different final decisions. In a fully deterministic self-hosted benchmark, exact reruns are identical, yet small meaning-preserving changes to the scenario text still separate over time and often alter the final recommendation. In deployed black-box API systems, nominally identical committee runs likewise remain unstable even at temperature 0, where many users expect near-determinism. Across 12 policy scenarios, these findings indicate that instability in collective AI is not only a consequence of residual platform-side stochasticity, but can arise from sensitivity to nearby initial conditions under repeated interaction itself. Additional deployed experiments show that committee architecture modulates this instability: role structure, model composition, and feedback memory can each alter the degree of divergence. Collective AI therefore faces a stability problem, not only an accuracy problem: deterministic execution alone does not guarantee predictable or auditable deliberative outcomes.

Figures & tables

Figure 1: Deterministic hosted committees remain sensitive to nearby initial conditions. Hosted runs were executed in a fixed deterministic environment, and exact reruns of the same full 20-round committee pipeline were artifact-identical. ( A ) Mean pairwise trajectory distance

D(t)

in HL-01 under meaning-preserving perturbations to the scenario text, shown separately for Roles and NoRoles . ( B ) Cross-scenario decision fragility under deterministic perturbations, defined as one minus the plurality share across 20 variants. ( C ) Cross-scenario perturbation-divergence slope

\hat{\lambda}_{\mathrm{pert}}

. The variation shown therefore reflects controlled perturbations of initial conditions rather than stochastic reruns. Operationally, a fragility value of 0.60 means the plurality recommendation appears in only 40% of nearby variants, so the same underlying case can yield materially different committee recommendations under tiny wording changes.

Figure 2: Architecture modulates deployed instability (HL-01, $T=0$ ). ( A ) Mean pairwise trajectory distance $D(t)$ for three conditions: uniform NoRoles (positive but lower-divergence baseline), uniform Roles , and mixed NoRoles . Adding roles in a homogeneous committee increases divergence, and heterogeneous model composition without roles also increases divergence. ( B ) 2 $\times$ 2 matrix crossing roles and model composition. The mixed+roles cell is less unstable than mixed+no-role, showing non-additive interaction. Realized sample sizes are shown in-cell (uniform+roles has slight attrition at $n=19$ ; all other cells are $n=20$ ).

Figure 3: Chair channel within role-structured deployed committees. ( A ) HL-01 role-ablation effects on

\Delta\hat{\lambda}=\hat{\lambda}_{\text{full roles}}-\hat{\lambda}_{\text{ablated}}

. Chair ablation yields the largest reduction among role mandates (with a no-role reference bar shown for context). ( B ) Scenario-level Chair effect across IM-01, HL-01, CL-01, SP-03, and AI-01, reported as

\Delta\hat{\lambda}=\hat{\lambda}_{\text{full roles}}-\hat{\lambda}_{\text{ablate Chair}}

. Point estimates are positive in all five scenarios, with strongest and best-resolved effects in IM-01 and HL-01.

Figure 4: Feedback-memory interventions attenuate instability. ( A ) Intervention: reducing memory window from

k=15

k=3

attenuates divergence across four scenarios. ( B ) Falsification target: collapsing memory to

k=1

lowers

\hat{\lambda}

in IM-01 and CL-01 relative to baseline Roles . Together these tests support a feedback-memory amplification pathway.

LLM Decision-Making

LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.

See all 242 papers

Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding
May 26, 2026Wajdi ZaghouaniAI Accountability Human-in-the-Loop AI
Northwestern University in Qatar
The Wisdom of Artificial Deliberative Crowds
Sep 18, 2026Federico Barrera-Lemarchand, Mariano Sigman, Joaquin NavajasMulti-Agent Deliberation Multi-Agent Consensus
Laboratorio de Neurociencia, Escuela de Negocios, Universidad Torcuato Di Tella, Buenos Aires, Argentina · Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Buenos Aires, Argentina
Multi-agent discussion gains less when dissent is withheld
Sep 29, 2026Chand Sahil Mansuri, Xin Wang, Mengying Li +4Multi-Agent LLM Systems Opinion Dynamics
School of Systems Science and Industrial Engineering, Binghamton University, Binghamton, NY, USA · School of Management, Binghamton University, Binghamton, NY, USA · IBM T. J. Watson Research Center, Yorktown Heights, NY, USA

May 26, 2026J/K move · Enter open · S save

Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding

Wajdi Zaghouani

Northwestern University in Qatar

AI Accountability Human-in-the-Loop AI Deliberative Democracy

Open paper AskPDF arXiv

Large language models (LLMs) can support democratic deliberation at scales previously constrained by turn-taking and facilitation bandwidth. Recent work shows that LLM-generated group statements are often preferred over human-mediated outputs, while theoretical analyses argue that LLMs relax the simultaneity constraints limiting collective intelligence. Yet pure LLM mediation risks collapsing pluralism, over-optimizing for agreement, and undermining legitimacy when participants cannot contest how they are represented. We propose a symbiotic human-AI framework organized into three layers: observation and diversity amplification, facilitation with clause-level provenance, and human primacy for ratification. Our contributions include graded coverage, diversity, and erasure metrics with salience-aware weighting; a provenance pipeline combining cross-encoder similarity with causal knockout diagnostics; preference-conditioned trade-off control; equity-aware contestability workflows; adversarial robustness tests; and an evaluation protocol with ablation designs informed by evidence of LLM-as-judge limitations. The result is a testable blueprint for deliberation technology that scales collective intelligence while preserving agency and legitimacy.

Abstract

LLM Decision-Making

Explore similar work

Accountable Human-AI Deliberation with LLMs: Scaling Collective Intelligence through Symbiotic Scaffolding

The Wisdom of Artificial Deliberative Crowds

Multi-agent discussion gains less when dissent is withheld