What question did this study set out to answer?

June 26, 2026Open Access

Distributed Cognitive Architecture (DCA) — Theory I: Atomic Agents · Fractal Composition · Convergence

Key Points

The aim is to establish a formal convergence theory for memory-augmented multi-agent systems using frozen Foundation Models.
Developed atomic WMC-Agents that incorporate a World Model and Memory Controller.
Established convergence signals across five measurement families: geometric, semantic, structural, statistical, consensus.
Deployed the architecture in the DocVQA 2026 challenge to assess performance at competition scale.
The DCA architecture achieved first place in the >35B-parameter category of the DocVQA 2026 challenge.
Demonstrated effective intra-agent memory dynamics and inter-agent dynamics with finite-termination and bounded-accumulation.
Validation of framework through measurable convergence signals and architectural composition.

Abstract

Distributed Cognitive Architecture (DCA) — Theory I develops the formal convergence theory for memory-augmented, multi-agent systems built around frozen Foundation Models. Genuine task-solving intelligence requires two dynamics that current Foundation Models lack: task-adaptive memory access — which knowledge enters working context at each step, beyond static chat histories and one-shot retrieval — and task-adaptive architectural composition — how a task decomposes into sub-tasks dispatched to specialists, beyond static workflow graphs. DCA supplies both: atomic WMC-Agents (a World Model coupled with a Memory Controller) composed fractally into multi-agent hierarchies. The central abstraction is the convergence signal — an observable quantity that is bounded, decreases in expectation under task progress, and is grounded in a system-level goal. Five families of such signals (geometric, semantic, structural, statistical, consensus) span the measurement modalities of the architecture, and a single signal substrate serves all three run-time consumers: the Memory Controller's Context Retrieval Policy, the Orchestrator's Orchestration Policy, and the Convergence Monitor that aggregates them into a Lyapunov-style measure for termination. This substrate-sharing is what makes intra-agent memory dynamics and inter-agent multi-agent dynamics one theory rather than two, with finite-termination, bounded-accumulation, and calibration guarantees that hold uniformly across measurement modalities. The framework was deployed end-to-end at the DocVQA 2026 challenge (ICDAR 2026), where the architecture placed first in the >35B-parameter category of the official, externally juried leaderboard — an existence proof that it operates at competition scale. Theory I is the formal-theory member of the DCA paper family — companion to DCA — Foundations (the biological motivation) and to a planned Theory II. Detailed empirical results appear in the companion technical report (DOI: 10.5281/zenodo.20707289).

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper