A grand challenge in artificial intelligence is creating agents that can adapt to entirely novel situations, a capability that requires moving beyond selecting from a known set of strategies to inventing new ones. This paper details a significant advance in the Fluid Emergent Reasoning Engine (FERE-CRS) project, demonstrating a principled mechanism for such "generative metacognition." We address a fundamental limitation of prior work: while agents could learn to select the right cognitive "tool" from a fixed toolbox, they were helpless when faced with a problem requiring a tool they did not possess. To overcome this, we introduce a new architectural component grounded in Active Inference: the Stance Generation Network (SGN). The SGN is a generative model trained to analyze the abstract features of a novel problem and output a complete, bespoke cognitive stance—a set of motivational parameters (Cognitive Resonance Score weights)—tailored to the unique demands of the situation. We test this "Generative Agent" on a novel class of social-ethical dilemmas requiring a mode of thinking unavailable to previous agents. The results provide powerful, dual confirmations of our hypotheses. First, the SGN demonstrated high accuracy in generating effective stances for unseen problem types (H8: Pearson correlation 0.96). Second, the Generative Agent dramatically outperformed a sophisticated Phase III control agent on the novel task (H9: mean quality score 9.86 vs. 4.99). Crucially, the control agent's failure validates the experimental design, proving it lacked the requisite motivational structure to solve the problem. The Generative Agent succeeded precisely because it could invent this structure on the fly. This work presents the first quantitative evidence of an agent moving from a "cognitive mechanic" to a "cognitive engineer," representing a critical and scientifically validated step toward truly fluid and adaptive artificial intelligence.
Thomas E. Devitt (Wed,) studied this question.