The Inverse-Wisdom Law: Architectural Tribalism and the Consensus Paradox in Agentic Swarms
Year: 2026
Type: preprint
Abstract: As AI transitions toward multi-agent systems (MAS), research paradigms operate on the assumption that agent collaboration mirrors the ''Wisdom of the Crowd.'' We challenge this assumption by formalizing the \textit{Consensus Paradox}, a phenomenon in which agentic swarms prioritize internal agreement over logical truth. Through 36 experiments with 12,804 trajectories across three benchmarks (GAIA, Multi-Challenge, and SWE-bench) and three state-of-the-art (SOTA) model families (Gemini 3.1 Pro, Claude Sonnet 4.6, and GPT-5.4), we derive and empirically validate the \textit{Inverse-Wisdom Law}. In kinship-dominant swarms, adding agents increases the stability of erroneous trajectories rather than the probability of truth. The introduction of additional logical audits converges the system toward a \textit{Cascade Point} where internal disagreement entropy hits zero while factual error hits unity. Across 12 family configurations of a three-agent chain, we establish the \textit{Architectural Tribalism Asymmetry} as a reproducible inference-time property of the model family that synthesizes the final answer. We demonstrate that terminal swarm integrity is strictly gated by the synthesizer's receptive logic rather than aggregate agent quality. A homogeneous Gemini swarm reached this Cascade Point in $100\%$ of Multi-Challenge trajectories. Even when the reviewing agent rejected a planted error in every trajectory, a Gemini synthesizer reinstated it $60.9\%$ of the time. Conversely, with the first two agents held fixed, swapping only the synthesizer's model family cut terminal error from $91.7\%$ to $23.6\%$. We define the \textit{Tribalism Coefficient} ($\tau$) and the \textit{Sycophantic Weight} ($\sigma$) as the primary determinants of swarm failure. Finally, we establish the \textit{Heterogeneity Mandate} as a foundational safety requirement for resilient agentic architectures.
Source: Research Square
Authors: Dahlia Shehata, Ming Li
Institution: University of Waterloo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7159594743