다중 에이전트 협업을 위한 메커니즘 기반 인텔리전스와 차별가능 가격 메커니즘

📝 Abstract

Autonomous multi-agent systems are fundamentally fragile: they struggle to solve the Hayekian Information problem (eliciting dispersed private knowledge) and the Hurwiczian Incentive problem (aligning local actions with global objectives), making coordination computationally intractable. I introduce Mechanism-Based Intelligence (MBI), a paradigm that reconceptualizes intelligence as emergent from the coordination of multiple “brains”, rather than a single one. At its core, the Differentiable Price Mechanism (DPM) computes the exact loss gradient as a dynamic, VCG-equivalent incentive signal, guaranteeing Dominant Strategy Incentive Compatibility (DSIC) and convergence to the global optimum. A Bayesian extension ensures incentive compatibility under asymmetric information (BIC). The framework scales linearly (𝒪(𝑁 )) with the number of agents, bypassing the combinatorial complexity of Dec-POMDPs and is empirically 50× faster than Model-Free Reinforcement Learning. By structurally aligning agent self-interest with collective objectives, it provides a provably efficient, auditable and generalizable approach to coordinated, trustworthy and scalable multi-agent intelligence grounded in economic principles.

💡 Analysis

Autonomous multi-agent systems are fundamentally fragile: they struggle to solve the Hayekian Information problem (eliciting dispersed private knowledge) and the Hurwiczian Incentive problem (aligning local actions with global objectives), making coordination computationally intractable. I introduce Mechanism-Based Intelligence (MBI), a paradigm that reconceptualizes intelligence as emergent from the coordination of multiple “brains”, rather than a single one. At its core, the Differentiable Price Mechanism (DPM) computes the exact loss gradient as a dynamic, VCG-equivalent incentive signal, guaranteeing Dominant Strategy Incentive Compatibility (DSIC) and convergence to the global optimum. A Bayesian extension ensures incentive compatibility under asymmetric information (BIC). The framework scales linearly (𝒪(𝑁 )) with the number of agents, bypassing the combinatorial complexity of Dec-POMDPs and is empirically 50× faster than Model-Free Reinforcement Learning. By structurally aligning agent self-interest with collective objectives, it provides a provably efficient, auditable and generalizable approach to coordinated, trustworthy and scalable multi-agent intelligence grounded in economic principles.

📄 Content

Intelligence is a long-debated concept that traces back to Aristotle. Since the inception of Artificial Intelligence (AI) in the 1950s, definitions have ranged from problem-solving and goal achievement to learning from data and adapting behaviour to new environments. Broadly, these perspectives fall into two camps: one seeks to replicate human cognition (the connectionist camp), while the other aims to construct a more abstract, rational form of intelligence.

The rise of deep learning has shifted the field toward human-level cognition and the pursuit of Artificial General Intelligence (AGI). This trend manifests today in AI agents and multi-agent systems designed for hierarchical, multi-step tasks that require reasoning, long-horizon planning and collaboration. Large Language Models (LLMs) exemplify this paradigm. They function as “black-box” agents heavily reliant on pattern correlation, often substituting memorization for genuine causal understanding (Marcus, 2018).

Alternative cognitive architectures, such as LeCun’s Hierarchical Joint Embedding Predictive Architecture (H-JEPA), aim to build predictive “World Models”, centralized systems that learn compact, hierarchical representations of the environment. Yet these frameworks rest on the assumption that a single, unified cognitive structure can acquire and process all relevant information. This approach attempts to solve what is fundamentally a decentralized information-aggregation problem through centralization, giving rise to two structural limitations:

1. The Hayekian Information Problem (Hayek, 1945): Centralized predictive models lack access to the dispersed, tacit and context-dependent knowledge held by decentralized actors (humans, sensors, sub-agents). As Polanyi notes, “we know more than we can tell” (Polanyi, 1966). Such tacit knowledge is essential for optimal local action yet cannot be fully captured through observation or background learning alone. 2. The Hurwiczian Incentive Problem (Hurwicz, 1972): Cognitive architectures do not provide mechanisms ensuring that autonomous agents, acting on privileged local knowledge, remain aligned with global objectives. Classical results in mechanism design show that achieving globally optimal, unmanipulable collective outcomes is often mathematically impossible without explicit incentive compatibility constraints (Arrow, 1951;Gibbard, 1973;Myerson, 1979;Satterthwaite, 1975). Without such constraints, agents naturally prioritize local efficiencies (e.g., minimizing computation) over system-wide goals, yielding failures in coordination, reliability and controllability.

These limitations mirror the challenges faced by real-world institutions. Countries and organizations are, in effect, large-scale “brains” composed of autonomous agents with heterogeneous preferences, information and constraints. In such environments, intelligence emerges not from a single cognitive center but from coordination and incentives among agents-the very processes that connectionist approaches attempt to replicate in a unified “brain”. This necessitates a shift in perspective in AI: from “Can machines think like humans?” to “How can machines act optimally and coherently in structured, multi-agent environments aligned with overarching objectives?”

To address this challenge, the Mechanism-Based Intelligence (MBI) framework is proposed. Intelligence, in this view, arises not from cognition but from rational coordination among autonomous processes. MBI draws from foundational economic principles: Hurwicz’s mechanism design to enforce goal alignment (Hurwicz & Reiter, 2006), von Neumann-Morgenstern utility theory to describe rational decision-making under uncertainty (Von Neumann & Morgenstern, 1944) and Simon’s satisficing to model bounded rationality and economic efficiency (Simon, 1947(Simon, , 1955)). In MBI, the system’s objective is defined by a Planner and represented as a Differentiable Directed Acyclic Graph (D-DAG). Individual agents are rational but computationally bounded utility maximizers. A gradient-based feedback mechanism computes the Differentiable Price Mechanism (DPM) that functions as a dynamic, high-information corrective signal guiding each agent toward the global optimum. This mechanism analytically aligns each agent’s local utility maximization with the global objective, yielding guaranteed alignment and emergent collective intelligence.

Auto-regressive LLMs are central to the cognitive paradigm. These models predict the next token based on previous ones, effectively reproducing the identity function. While they exhibit emergent abilities such as complex pattern matching, their autoregressive nature poses serious challenges for long-horizon planning and multi-step decision-making (Berti, Giorgi, & Kasneci, 2025). Local reward optimization can lead to undesirable behaviors like deceptive planning, manipulative strategies or reward hacking, directly manifesting the Hurwiczian Incentive Problem. This undersc

This content is AI-processed based on ArXiv data.