A critical step towards artificial general intelligence is the creation of agents that can adapt their cognitive strategies to solve diverse problems. While our previous work introduced the Fluid Emergent Reasoning Engine (FERE-CRS), a cognitive architecture grounded in Active Inference, it relied on hand-tuned, static parameters that defined its problem-solving "personality." This paper details a significant evolution of that work, moving from a configured to a learning agent. We address the core research question: Can an agent learn to dynamically adapt its own internal motivations—specifically, the weights of its Cognitive Resonance Score (CRS) heuristic—by experiencing a curriculum of fundamentally different problem types? We introduce a "cognitive curriculum" methodology, training the agent on two opposing problem sets: a logical/convergent task (Sudoku) and a creative/divergent task (the Alternative Uses Test). A simple reinforcement learning mechanism was implemented, allowing the agent's Meta-Reasoning Agent (MRA) to update its CRS weights based on performance. The results provide strong support for our primary hypotheses. Agents trained on a single curriculum developed highly specialized "cognitive stances," becoming either a "cautious logician" (prioritizing rule-adherence and efficiency) or an "exploratory creative" (prioritizing novelty). Furthermore, an agent trained on a mixed curriculum demonstrated a nuanced adaptation, developing a skewed strategy that favored the more reliably rewarded logical task. This provides a powerful, quantitative demonstration of how an agent's emergent cognitive strategy is a direct function of its experience and the reward landscape of its environment. This work represents a crucial step in moving from brittle, task-specific systems to more general and adaptive artificial intelligence.
Thomas E. Devitt (Sat,) studied this question.