Logos is a cognitive architecture that enables systems to learn not only from outcomes, but from the structure of their own reasoning failures. While large language models can reason but do not learn from real-world interaction, and reinforcement learning systems optimize actions without improving reasoning processes, Logos unifies reasoning, prediction, action, and learning within a single iterative loop. This work introduces a learning mechanism based on prediction–reality discrepancy. Each decision is formulated as a hypothesis with explicit rationale and assumptions, evaluated through both predicted and actual outcomes. The discrepancy between expectation and reality is transformed into dual learning signals: a numerical signal that improves predictive models, and a semantic signal that refines reasoning itself. Through this mechanism, Logos continuously improves both its world model and its reasoning strategy. Learning is not driven solely by outcomes, but by understanding how and why predictions fail. This work extends the original Logos architecture by formalizing its learning process, defining structured experience representations that include reasoning components, and establishing evaluation as a first-class process that bridges reasoning and learning.
Seongyun Ko (Sat,) studied this question.