AXIOM Cognitive Core v2: Building AI Systems with World Models and Continuous Learning Capabilities

AXIOM Cognitive Core v2 is a modular cognitive architecture project that integrates cutting-edge AI technologies including world models, curiosity-driven learning, continuous learning, abstract reasoning, and self-modeling. It aims to address the lack of deep cognitive capabilities in current large language models, build AI systems with richer cognitive abilities, and represents an important exploration towards Artificial General Intelligence (AGI).

Current large language models excel in language understanding and generation, but lack deep cognitive capabilities—such as forming internal models of the world, curiosity-driven exploration, continuous interactive learning, abstract reasoning, and self-awareness. AXIOM is designed to address these challenges, using a modular architecture to build AI systems with multiple cognitive functions by combining independent modules.

AXIOM contains eight core modules:

1. World Model: Builds dynamically updated internal representations of the external world, supporting prediction, planning, and counterfactual reasoning;
2. Curiosity-Driven Learning: Uses intrinsic reward mechanisms to guide agents to explore the unknown and reduce cognitive uncertainty;
3. Continuous Learning: Prevents catastrophic forgetting through techniques like regularization and memory replay, enabling progressive knowledge integration;
4. Abstract and Conceptual Learning: Extracts abstract concepts using technologies like Beta-VAE to enhance generalization ability;
5. Reasoning and Causal Modeling: Combines symbolic and neural reasoning to support causal analysis and counterfactual reasoning;
6. Active Inference: Unifies perception and action, minimizing free energy to optimize understanding and modification of the world;
7. Attention and Global Workspace: Enables selective information processing and global broadcasting, coordinating information flow between modules;
8. Self-Modeling: Establishes models of one's own capabilities and states, supporting metacognition and adaptive behavior.

AXIOM's technical highlights include:

• Modularity and Composability: Each module is independently developed and tested, allowing flexible configuration and expansion;
• Neuro-Symbolic Hybrid Architecture: Combines the pattern recognition capabilities of neural networks with the reasoning capabilities of symbolic systems;
• Neuroscience-Inspired Components: Uses components like Hopfield networks (associative memory), liquid neural networks (temporal processing), and Beta-VAE (concept discovery), balancing computational power with biological plausibility.

AXIOM is suitable for multiple scenarios:

• Autonomous Robots: Provides world modeling and continuous learning capabilities to support autonomous navigation in unknown environments;
• Interactive AI Assistants: Remembers user preferences and adapts to needs through continuous learning and self-modeling;
• Scientific Discovery: Uses curiosity-driven learning and abstract abilities to identify data patterns and propose hypotheses;
• Educational Applications: Serves as a personalized education partner, stimulating learning interest and adjusting teaching strategies.

AXIOM faces challenges in module integration, coordination, and cognitive ability evaluation. Future directions include:

• Implementing end-to-end joint training of modules;
• Verifying system capabilities in complex real-world environments;
• Combining with large language models to enhance interactive capabilities;
• Deepening the theoretical foundation of cognitive architectures.

AXIOM represents a new approach to building AI systems—pursuing general cognitive capabilities rather than single-task performance, and is an important attempt towards AGI. As an open-source project, it provides an experimental platform for researchers with clear code structure and complete documentation, serving as a good starting point for exploring the intersection of cognitive science and AI.