NAGI: Neuro-Adaptive General Intelligence —— A Theoretical Framework Distinct from Neuroevolutionary Approaches

Current artificial intelligence systems based on gradient optimization and external reward signals face fundamental limitations: lack of autonomous decisionmaking ability, non-interpretable reasoning processes, reliance on large-scale labeled data, and frequent hallucination problems. This paper proposes NAGI (Neuro-Adaptive General Intelligence) -a biomimetic architecture grounded in neuroscience and cognitive science, aiming to provide an alternative theoretical pathway for artificial general intelligence. The core components of NAGI include : (1) Flexible-Transfer-Memory (FTM) neurons, achieving multi-timescale adaptation through a three-parameter model ; (2) Global workspace cognitive integration mechanism, realizing information broadcast through competitive coalition formation; (3) Authentic Tree-of-Thought (A-ToT), mapping each reasoning step to a traceable global workspace ignition event ; (4) Continuous Creative Dynamics (CCD), modeling the goal breakthrough process using neural ordinary differential equations ; (5) De Novo goal constructor, achieving autonomous goal generation through strategies such as counterfactual composition and value-guided interpolation. This paper derives the theoretical properties of each module from the free energy principle, including Lyapunov stability analysis of the CCD energy landscape and statistical guarantees of goal diversity from De Novo construction. An accompanying executable prototype validates the computational consistency of the theoretical framework. This work belongs to the position paper category, aiming to provide a testable theoretical blueprint for AGI research.