Introduction
Current trajectories for 6G and Artificial Intelligence (AI) are approaching fundamental limits. While 6G visions often focus on incremental improvements in bandwidth and latency, AI progress is largely driven by "System 1" models—fast, intuitive pattern-matchers that remain brittle and data-hungry. This paper proposes a foundational shift toward "System 2" semantic communication, integrating abstraction, algebraic compositionality, and emergent languages to enable intelligent agents to reason and collaborate through grounded interaction.
The Problem: Beyond Shannon's Level A
The core challenge lies in our historical adherence to Claude Shannon’s "Level A" communication: the technical problem of accurately reproducing symbols. While the industry has mastered bit transmission, it has largely ignored "Level B" (semantics: how symbols convey meaning) and "Level C" (effectiveness: how received meaning affects conduct). In an era of autonomous robotics and human-machine collaboration, systems must understand why they communicate, moving beyond the mere exchange of raw bits.
Methodology: A System 2 Research Vision
The authors advocate for a unified research vision rooted in System 2 Cognition—the slow, deliberate, and analytical mode of human thought. This framework is built upon three pillars:
- Abstraction (World Models): Decoupling knowledge of causal relationships (the "World Model") from the search for solutions (the "Inference Machine"). This allows for compressed, meaningful representations beyond raw sensor data.
- Algebraic Compositionality: Utilizing Sheaf Theory and Category Theory to provide formal tools for combining learned concepts. This "gluing" of local knowledge into a coherent global view is essential for multi-agent coordination.
- Emergent Communication: Moving away from hard-coded protocols toward adaptive, grounded languages that evolve through interaction, ensuring resilience in unforeseen environments.
Information as Structure: The Semantic Chameleon
A key insight of this work is that semantic information should be treated as a structure or category rather than a simple scalar value like entropy. The "Semantic Chameleon" concept illustrates how information changes based on the mathematical language employed—whether it be Galois theory, Topology, or Modal Logic.
By separating the world model from the inference process (e.g., using Generative Flow Networks), AI systems can become more interpretable and less prone to the hallucinations common in monolithic Large Language Models.
Conclusion and Future Roadmap
The transition from statistical "curve-fitting" to formal algebraic reasoning presents significant mathematical and computational challenges. Scaling these structures to real-time 6G networks and ensuring energy efficiency for edge devices remain open research questions. However, by embracing System 2 semantic communication, we can move toward a new paradigm where communication is viewed as the composition of internal information spaces, driving the next generation of truly intelligent, collaborative systems.
Summary updated