T11 – Reasoning-Driven Semantic Communication for AI-Native Networks

Presenter: Christo K Thomas, Worcester Polytechnic Institute, USA

Abstract: Artificial General Intelligence (AGI) and its practical realizations such as physical AI systems necessitate AI models to perceive, reason and act in the real world by understanding the physics behind real world operations. Such advancements in artificial intelligence (AI) will not emerge from isolated models alone, but from networks of intelligent agents that must communicate, align, and coordinate their internal world models in real time. As AI systems increasingly operate in distributed, embodied, and multi-agent environments such as autonomous networks, digital twins, and the Internet of Intelligence, the fundamental bottleneck shifts from local perception or learning to the ability of machines to exchange meaning, intent, and knowledge over communication networks. In this context, communication is no longer a problem of reliably transmitting bits, but of enabling aligned and interoperable intelligence. Classical Shannon-based communication theory, which treats messages as syntactic symbols and neglects their semantic effect on the receiver, is therefore insufficient for supporting such intelligence-native systems. Semantic communication (SC) offers a principled alternative by enabling systems to transmit task-relevant meaning rather than raw data, with the potential to dramatically improve efficiency, latency, reliability, and resilience. However, despite a surge of recent works, existing SC approaches either reduce semantics to narrow data-driven representations or rely on rigid, application-specific solutions, leaving fundamental questions about semantic representation, causal structure, reasoning, generalization, and interoperability largely unanswered.

In this tutorial, we present a rigorous and holistic framework for end-to-end SC founded on generalizable artificial intelligence, integrating causal reasoning, neuro-symbolic AI, information theory, game theory, and category theory. We show that scalable semantic communication requires a transition from data-driven AI-augmented networks to reasoning-driven AI-native networks, in which agents construct and exchange causal semantic world models that explicitly represent cause–effect relationships in the environment. Crucially, as SC systems scale to multi-agent and multi-vendor ecosystems, interoperability and AI alignment emerge as fundamental bottlenecks: heterogeneous agents trained on different data, models, and semantic languages may fail to interpret each other’s messages, leading to semantically defective cooperation. We introduce principled alignment mechanisms based on category-theoretic mappings, relative representations, and semantic language transformations that enable interoperable SC across heterogeneous agents while preserving intent, goals, and meaning. Beyond alignment, we establish compositional SC as a key enabler of scalable collective intelligence, allowing semantic representations from distributed agents to be formally combined to generate novel meanings and support generalization to unseen tasks and sensor configurations. Using categorical constructs such as lenses, fibrations, and Grothendieck topologies, we formalize semantic composition and introduce semantic information metrics that go beyond Shannon’s notion of uncertainty. To operationalize reasoning in semantic communication systems, we further introduce causal representation learning and neuro-symbolic AI as principled mechanisms for bridging data-driven perception with symbolic, causal, and logical world models.

We show how reasoning-driven SC enables agents to discover causal structure, learn semantic representations and signaling strategies from limited data, generalize to out-of-distribution scenarios, and perform semantic inference and content generation through structured reasoning rather than statistical correlation alone, ultimately allowing agents to do more with less. Additionally, we will discuss the impact of SC across distinct layers in AI-RAN, with specific examples including semantic importance aware adaptive constellation design, computing and communication resource allocation.

Finally, we discuss how these foundations enable practical SC architectures for emerging applications such as digital twins, connected intelligence, and AI-native wireless networks, and how semantic aware resource allocation and cross-layer design can be achieved in multi-user systems. Overall, this tutorial provides a unified theoretical and systems-level foundation for designing interoperable, compositional, and causality-aware reasoning-driven semantic communication networks, positioning SC as a critical infrastructure for future AI-native wireless systems and distributed intelligent agents.

Presenter’s Bio:

Christo K Thomas:

Christo K Thomas received his BS in Electronics and Communication Engineering from National Institute of Technology, Calicut, India in year 2010, his MS in Telecommunication Engineering from Indian Institute of Science, Bangalore, India in year 2012, and his PhD from EURECOM, France in year 2020. He is currently an assistant professor at the Electrical and Computer Engineering Department at WPI. Previously, he was a postdoctoral associate at the Electrical and Computer Engineering Department at Virginia Tech. His research interests include semantic communications, statistical signal processing, and artificial general intelligence (AGI)-native wireless systems. From 2012 to 2014, he was a staff design engineer on 4G LTE with Broadcom communications, Bangalore, and from 2014 to 2017, he was a design engineer with Intel corporation, Bangalore. During November 2020 till June 2022, he was a staff engineer on 5G modems with wireless research and development division of Qualcomm Inc., Espoo, Finland. He was a recipient of the best student paper award at IEEE SPAWC 2018, Kalamata, Greece, and received third prize for his team titled “Learned Chester” ML5G-PHY channel estimation challenge, as part of the ITU AI/ML in 5G challenge, conducted at NCSU, US, 2020. He had presented multiple tutorials on approximate Bayesian inference techniques at several IEEE conferences such as ICASSP and EUSIPCO, and a tutorial on AGI-native networks at IEEE GLOBECOM. He has also edited a Wiley-IEEEPress book on semantic communications and is co-author of a visionary paper on AGI-native networks in Proceedings of IEEE. He has received AI-RAN Alliance’s “Innovation Award” for his proposal on emergent semantic communication, during November 2025.