Semantic Communication meets System 2 ML: How Abstraction, Compositionality and Emergent Languages Shape Intelligence
Mehdi Bennis, Salem Lahlou · 2025
Read Semantic Communication meets System 2 ML: How Abstraction, Compositionality and Emergent Languages Shape Intelligence by Mehdi Bennis, Salem Lahlou (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-20964.txt.
Introduction
Mehdi Bennis and Salem Lahlou argue that future wireless networks and artificial intelligence should be rebuilt around meaning, in a 2025 position paper. They say current 6G plans mostly extend 5G, while today's AI models remain brittle and data-hungry, with weak reasoning. Their proposal draws on System 2 cognition, the slow, deliberate style of thought.
The vision rests on three pillars. Abstraction lets agents learn useful world models from raw sensory data. Compositionality gives tools to combine concepts and subsystems. Emergent communication lets agents invent their own grounded languages for working together. Woven together, the authors claim, these ideas could yield systems that reason, adapt, and collaborate across communications, learning, and robotics.
Catalog record
| CREATOR | Mehdi Bennis, Salem Lahlou |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-20964
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Semantic Communication meets System 2 ML: How Abstraction, Compositionality and Emergent Languages Shape Intelligence Mehdi Bennis, Salem Lahlou Abstract The trajectories of 6G and AI are set for a creative collision. However, current visions for 6G remain largely incremental evolutions of 5G, while progress in AI is hampered by brittle, data-hungry models that lack robust reasoning capabilities. This paper argues for a foundational paradigm shift, moving beyond the purely technical level of communication toward systems capable of semantic understanding and effective, goal-oriented interaction. We propose a unified research vision rooted in the principles of System 2 cognition, built upon three pillars: Abstraction , enabling agents to learn meaningful world models from raw sensorimotor data; Compositionality , providing the algebraic tools to combine learned concepts and subsystems; and Emergent Communication , allowing intelligent agents to create their own adaptive and grounded languages. By integrating these principles, we lay the groundwork for truly intelligent systems that can reason, adapt, and collaborate, unifying advances in wireless communications, machine learning, and robotics under a single coherent framework. I The Wireless Communication World The wireless communication landscape is on the verge of a paradigm shift. As we look towards 6G (the sixth generation of wireless technology) we are seeing ambitious claims about connecting everything and everyone. But beneath these claims lies a more fundamental question: Are we approaching communication from the right perspective? 6G [ 1 ] promises to bridge the digital, physical, and biological worlds truly ushering in an era in which everything is sensed, connected, and intelligent. Yet the current 6G vision(s) of “the more bits/s, more bandwidth, more base stations the better” or “X times 5G requirements” (e.g., in packet-error rate) are inadequate and not sustainable. On the AI/ML front, despite significant advances in the field, current solutions are brittle, energy-hungry, fail to generalize and are curve-fitting at best. Other visions such as O-RAN or statements like “6G is whatever we will have in 2030” are either stemming from a business standpoint or merely lacking ambition. In short, 6G today is nothing but an incremental evolution of 5G . These incremental advances may be rooted in two aspects: (i) the traditional radio engineering approach of the “higher the better, the bigger the better” , which is clearly not sustainable/scalable; (ii) radio engineers tend to cling to their domain knowledge instead of exploring the bigger vision with outside-of-the-box thinking. Instead of clinging to this path, could we flip this paradigm upside down by asking a different set of questions? Among these, can we transmit less data by instead leveraging reasoning and knowledge, thereby relaxing some of the stringent requirements? If 6G is genuinely about human-machine interaction/collaboration and communication for machines, can we learn from human’s two modes of cognition (System 1 and System 2) and imbue this knowledge into machines to augment and assist us? System 1 represents fast, intuitive, and automatic cognitive processes, while System 2 involves slower, deliberate, and analytical reasoning [ 2 ] (see Appendix D for details). By integrating both cognitive approaches into our communication paradigms, we will create systems that are both efficient and capable of deep understanding. These questions are worth delving into to unlock the true 6G revolution. And, as a matter of fact, this is the beginning of the revolution. Fig. 1: Shannon’s three levels of communication [ 3 ] . The answer lies in the creative collision of two technological revolutions: Despite the tremendous progress made in communication, we are still in its infancy. Principles of communication theory can be traced back to the seminal work of Claude Shannon whose underlying communication problem (also referred to as the technical problem ) was to reliably convey information from a sender to a receiver [ 3 ] . Under this definition, communication is tantamount to reproducing at one point either exactly or approximately a message at another point (see Fig. 1 ). Under level A, Shannon information is a statistical/syntactical description of information, concerned with the probability of co-occurrence of messages/symbols. Moreover, decoding a message requires a known mapping and an external observer (decoder) which is completely decoupled from context and (external) symbols are chosen by the external observer. This is clearly in contrast to how the brain and perceptual/biological systems operate. Although explicitly mentioned in the seminal paper (and Shannon himself issuing a word of caution in 1952 about making Shannon information universal in the bandwagon article [ 4 ] ), the semantic and effectiveness problems were irrelevant and put aside. In the age of robotics, LLMs and generative AI, it is now time to dust them off! While communication engineering has primarily focused on level A, the emerging challenges of 6G and human-machine collaboration demand that we now address the semantic and effectiveness problems of communication. Moreover, recent research suggests that standard training methods in AI like reinforcement learning and imitation learning can lead to catastrophic risks through goal misspecification and emergent self-preservation behaviors [ 5 ] . Addressing these risks requires a fundamental rethinking of how we approach both AI and communication systems. This evolution from purely statistical information processing towards semantic and effective communication represents a fundamental shift in how we conceptualize and optimize these increasingly complex wireless systems. In the following section, we explore different notions of information pertaining to these distinct communication levels. II Information: The Semantic Chamaeleon Fig. 2: Various shades of information. One of the key distinctions between Shannon’s three levels of communication pertains to the notion of information. In contrast to Shannon information (a scalar value associated with a probability distribution and measured by entropy), semantic information is concerned with information structures, shapes, spaces or more formally, information categories. Paraphrasing the French Mathematician René Thom, “the word information should be replaced by the word shape ” [ 6 ] , underscoring the fact that information is first and foremost of topological nature [ 7 ] (see Appendix E for an accessible introduction to topological information theory). This perspective transforms how we conceptualize information in communication systems. Depending on the mathematical language and algebraic structure of interest, different notions of information are at stake. One of the earliest theories of information—known as the theory of ambiguity —goes back to Galois theory, which associates a group to an algebraic equation. Galois theory offers a framework for understanding how different solutions to an equation are related, effectively capturing the structure of ambiguity or uncertainty. The associated group acts by permuting the set of solutions, characterizing how solutions may be indistinguishable from one another. As the group’s size increases, so does the degree of indistinguishability among solutions. Galois also examined the effect of conditioning on the group, in which acquiring additional information (represented by new numbers) restricts the group to a subgroup of the original one. In this algebraic setting, random variables X and Y are replaced by geometric spaces—sets on which a group acts, in which the relevant object is the quotient IG(X;Y)=\frac{G(X)\times G(Y)}{G(X,Y)} , formed from the Cartesian product of G(X) and G(Y) divided by the Galois group of the joint extension G(X,Y) . This construction serves as an algebraic analogue to the probabilistic mutual information I(X,Y)=H(X)+H(Y)-H(X,Y). A different notion of information relates to the invariant properties of topological