July 28, 2025

FERE-CRS Phase III Dynamic Cognitive Control and the Emergence of Artificial Fluid Reasoning

Key Points

The study introduces the Meta-Cognitive Context Classifier, which accurately classifies cognitive demands with over 95% accuracy.
Fluid Agents outperformed specialist control agents, offering higher-quality explanations during complex tasks.
The framework demonstrates dynamic strategy-switching between convergent and divergent reasoning styles.
Findings suggest a move towards more adaptable and robust artificial intelligence systems capable of managing diverse cognitive tools.

Abstract

A fundamental, yet elusive, component of general intelligence is cognitive control—the ability to fluidly shift reasoning strategies to meet the evolving demands of a complex problem. While artificial intelligence has excelled at creating specialists, a unified mechanism for dynamic strategy-switching within a single agent remains a grand challenge. This paper presents a significant architectural and empirical advance in the Fluid Emergent Reasoning Engine (FERECRS) project, directly tackling the problem of cognitive control. We ground our approach in Active Inference (AIF), proposing that true cognitive flexibility arises from a hierarchical process of meta-cognition, or inference about how to conduct inference. We introduce a novel architectural component, the Meta-Cognitive Context Classifier (MCCC), which functions as an internal control-switch. The MCCC is trained to analyze the cognitive demand of a sub-problem and dynamically load a corresponding, pre-learned set of motivational parameters (Cognitive Resonance Score weights), enabling the agent to switch between a "cautious logician" and an "exploratory creative" stance on a step-by-step basis. We test this architecture using a rigorous three-stage experimental design, culminating in a complex, multi-phase artifact analysis task. The results provide strong empirical support for our hypotheses. First, the MCCC achieved high accuracy (95%) in classifying cognitive demands, validating it as a reliable control mechanism. Second, the fully-equipped "Fluid Agent" demonstrated superior performance, achieving significantly higher-quality explanations than two specialist control agents. Analysis of the agent's reasoning trace reveals a clear, justifiable pattern of switching between convergent and divergent stances, providing the first quantitative signature of artificial fluid reasoning in this framework. This work demonstrates a viable, theoretically-grounded mechanism for moving beyond static AI "personalities" to agents that can learn to manage a repertoire of cognitive tools, representing a tangible step toward more general, robust, and adaptive artificial intelligence.

Mark Helpful

Bookmark

Relay