Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Semantic Communication meets System 2 ML: How Abstraction, Compositionality and Emergent Languages Shape Intelligence

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that 6G and AI should stop optimizing bits and start composing meanings.

desk verdict A broad, readable research manifesto for System 2 semantic communication, but the transformative promises are unsupported and the sheaf-composition mechanism is under-defined. read the letter →

arxiv 2505.20964 v1 pith:7FCYEEFR submitted 2025-05-27 cs.LG cs.ITmath.IT

Mehdi Bennis , Salem Lahlou This is my paper
classification cs.LGcs.ITmath.IT
keywords semanticcommunicationSystem2machinelearningworldmodelscompositionalityemergentsheaftheorytopologicalinformation6G
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the next leap in wireless communication and artificial intelligence depends on changing what is transmitted: not raw bits or reconstructed signals, but goal-relevant semantic structure. It proposes a unified research vision built on three pillars, abstraction, compositionality, and emergent communication, inspired by System 2 cognition. The vision is that agents learn world models from sensorimotor data, compose these models algebraically, and invent grounded languages for coordination. If the vision holds, networks would transmit only the semantic information needed for a task, yielding order-of-magnitude gains in bandwidth, communication, and energy efficiency, while AI systems gain reasoning, adaptability, and collaboration. The paper is a programmatic call to merge wireless, machine learning, and robotics under one framework.

What carries the argument

The load-bearing machinery is sheaf-theoretic composition of semantic information spaces. A sheaf assigns algebraic data structures, such as vector spaces, lattices, or topological spaces, to the local observations of each agent or modality, and provides restriction maps for 'glueing' locally consistent pieces into a globally coherent picture. The paper uses this to define communication as the composition of local information spaces for mutual predictability, supported by a separation between a world model and an inference machine. Generative flow networks (GFlowNets), which sample structured objects step by step, serve as the proposed inference mechanism, while bisimulation relations, persistent homology, and epistemic logic provide additional algebraic and logical tools for abstraction and verification.

What would settle it

Set up a cooperative task where two agents with different sensor modalities, such as camera and LiDAR, may exchange only sheaf-composed semantic messages, and compare task success and bandwidth against a deep joint source-channel coding baseline that transmits compressed latent representations. If the composed-semantic system does not match or beat the baseline on both axes, the claimed order-of-magnitude efficiency gain is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that the standard statistical view of communication, Shannon's level A, is insufficient for 6G and for truly intelligent AI systems. It proposes a paradigm shift to System 2-oriented semantic communication, where agents do not merely reconstruct messages but reason about intents, beliefs, and goals. The core discovery is a unified architecture: agents abstract their environment into world models, compose those models using algebraic and topological structure, and communicate through emergent languages that are grounded in interaction. Communication itself is reconceived as agents composing their internal information spaces, or a 'sheaf of world models,' for mutual predictability. The paper asserts that this will enable order-of-magnitude improvements in bandwidth, communication, and energy efficiency, along with agents that can reason, adapt, and collaborate in open-ended environments.

Load-bearing premise

Everything rests on the premise that meaning can be captured as algebraic or topological structure and that gluing these structures across different agents and sensors yields semantic information that is reliable, interpretable, and cheaper than sending raw data.

Editorial extensions

If this is right

  • Networks could stop transmitting raw sensor streams and instead send only updates to a shared, composed world model, relaxing rate and energy budgets.
  • ML architectures could be redesigned so that each layer is a control loop and stacking layers is semantic composition, offering a path beyond autoregressive transformers.
  • Agents with different syntax, priors, and beliefs could coordinate through emergent communication protocols that generalize beyond hand-coded signaling.
  • Formal verification through signal temporal logic and sheaf-theoretic consistency could give mission-critical deployments time-bounded safety guarantees.
  • The same framework would unify wireless, ML, and robotics research, replacing fragmented System 1 semantic-communication approaches.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial: a direct quantitative test is still missing; until a multi-agent system demonstrates that sheaf-composed semantic messages beat raw-data baselines in rate-accuracy trade-offs, the efficiency claim should be treated as a hypothesis.
  • Editorial: the framework implies that persistent-homology summaries could serve as rate-distortion-optimal descriptors of semantic content, a consequence the paper sketches but does not prove.
  • Editorial: the strongest risk is that learnable restriction maps will overfit to training environments, so out-of-distribution robustness of composed semantics is the natural benchmark for the whole agenda.
  • Editorial: connecting the proposed compositional communication to causal representation learning would give a concrete way to test whether composed concepts remain invariant across environments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper proposes a unified research agenda for 6G and AI centered on "System 2 Semantic Communication" (System 2 SC), a framework that combines abstraction, algebraic compositionality, and emergent communication. It argues that current approaches remain at Shannon's level A and at Kahneman-style System 1 processing, and it advocates a shift toward agents that learn world models, compose them algebraically (via sheaf theory and category theory), and develop grounded, emergent languages. The manuscript surveys multiple notions of information (Galois-theoretic, topological, epistemic, sensorimotor), connects them to world models, GFlowNets, and topological data analysis, and concludes with four research thrusts. No experiments, simulations, or formal derivations are presented; the contribution is a conceptual synthesis and a research program.

Significance. If the proposed agenda were realized, it could be influential by connecting semantic communication with System-2-style machine learning, world models, and algebraic methods, potentially reshaping how the community thinks about goal-oriented communication and multi-agent reasoning. The paper's strengths are its breadth of synthesis, its clear articulation of three pillars, its accessible appendices on topological information and world models, and its explicit formulation of research questions across multiple disciplines. However, the central quantitative promises, especially the "order-of-magnitude" efficiency claim, and the pivotal sheaf-theoretic composition step are not yet supported by any concrete construction or evidence. The current value is therefore as a vision/position paper rather than as a completed technical proposal.

major comments (4)
  1. [Section V-B / Appendix E (RT1)] The load-bearing enabler of the proposed framework is the claim that agents "compose their internal information spaces (sheaf of world models) for mutual predictability" (Section V-B). For sheaf gluing to apply, the learned local models must be sections of a common sheaf: one needs a base space, restriction maps between agents/modalities, and cocycle compatibility on overlaps. The manuscript supplies none of these; Fig. 8 merely illustrates vector-space sheaves with linear restriction maps. Moreover, the direction is circular as stated: gluing requires local sections to already agree on overlaps, while Section V-C and RT2 say that communication itself is what aligns agents' different syntax, priors, and beliefs. If composition is the prerequisite for communication and communication is the mechanism for achieving compatibility, the initial compatibility is assumed rather than obtained. Please provide a concrete formalization, or a minimal worked example, or explicitly state that this is an open problem whose solution is a goal of RT1 rather than a premise.
  2. [Abstract / Section IV] The abstract and Section IV assert that the vision "promises ... order-of-magnitude improvements in bandwidth-communication-energy efficiency" and will produce "truly intelligent systems that can reason, adapt, and collaborate." No derivation, simulation, benchmark, or even a back-of-the-envelope argument supports the quantitative magnitude. Since the paper contains no experiments or formal claims, these statements should be reframed as hypotheses to be tested, with an explicit discussion of the regimes in which semantic transmission could plausibly outperform raw-bit transmission by an order of magnitude. Without this qualification, the central promise of the paper is an unsupported assertion.
  3. [Appendix D.2] Appendix D.2 states that the mathematical details of System 2 SC are "beyond our scope" and cites only "very early preliminary works" [53], [54]. Since System 2 SC is the paper's proposed paradigm shift, this is a significant gap: the manuscript does not yet contain the central construction it advertises. The paper should either include a precise statement of the intended semantic-information calculus, or clearly label itself as a research manifesto whose formal content is deferred to future work. The current framing claims more than it delivers.
  4. [Section II] Section II defines the Galois-theoretic quantity IG(X;Y) = G(X) x G(Y) / G(X,Y) as an "algebraic analogue" of mutual information. As written, this is not well-defined: G(X,Y) is described as the Galois group of the joint extension, but no embedding of G(X,Y) into G(X) x G(Y) is given, and no group action on the product is specified. If the formula is intended only as an analogy or a schematic, that should be stated explicitly; otherwise the construction must be completed. This matters because Appendix D.2 later claims that semantics has "precise mathematical foundations," and this equation is the only explicit algebraic definition in the main text.
minor comments (6)
  1. [Section V-C] In the bisimulation definition, the notation R_i(s_i,a) = R_j(s_j,a) is inconsistent; reward is a property of states, not of separate functions per state, so it should read R(s_i,a) = R(s_j,a) or the indices should be introduced and explained.
  2. [Section III-A] The symbol Y in the sentence "JEPA predicts an abstract representation of Y" is never defined; the reader must infer that Y is the target observation (e.g., an image or video).
  3. [Section II / Appendix B] The footnote "Betti numbers 1" is incomplete; it should say "Betti numbers" or "the first Betti number" if a specific one is intended.
  4. [Abstract and throughout] The hyphenation of "System 2" and "System-2" is inconsistent; please unify the spelling.
  5. [Reference [51] / Appendix D] Reference [51] is presented as a URL to a conference panel; if the System 1 SC vs. System 2 SC distinction was first proposed in that talk, the citation should include the exact talk title, date, and venue, and the authors should explicitly acknowledge that the paper's central organizational axis originates in the first author's own prior proposal.
  6. [Figures 4 and 7] The captions of Figures 4 and 7 are minimal and do not explain the notation; please state what the arrows, nodes, and labeled blocks represent so the composition operation is intelligible to a reader unfamiliar with the specific diagrams.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the paper is a research vision rather than a derivation; the single self-citation ([51]) provides provenance for the System 1/System 2 SC taxonomy but is not used as proof of any quantitative claim.

full rationale

The manuscript is a position/research-agenda paper. It contains no fitted parameters, no quantitative predictions derived from equations, and no claim that a derived result is forced by prior work. The main pillars—abstraction, compositionality, emergent communication—are supported by external literature (Kahneman, GFlowNets, sheaf theory, TDA), and the 'order-of-magnitude' language is an aspirational motivation, not an output of a calculation. The only self-referential element is Appendix D's attribution of the System 1 SC vs System 2 SC distinction to the first author's 2021 talk [51]. That citation is a provenance note for a taxonomic framing, not a load-bearing proof: the paper does not use it to forbid alternatives or to justify a uniqueness theorem. The sheaf-theoretic gluing proposal is programmatic; Appendix D explicitly says 'the details are beyond our scope' and cites only preliminary works [40], [53], [54]. This is an open-construction gap, not a circular reduction of conclusions to premises. Therefore no circular step meeting the required standard is present.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The paper introduces no free parameters because it makes no quantitative model. Its central claims rest on a series of domain assumptions about the nature of semantics, the feasibility of System 2 ML, and the composability of information via sheaf theory. The only invented 'entity' is the System 2 SC label itself, which organizes the agenda but has no independent empirical content.

assumptions (4)
  • domain assumption Semantic information is fundamentally topological and is best captured by algebraic structures such as shapes, spaces, and categories.
    Invoked throughout Section II and Appendix B, this is a philosophical premise stated via René Thom, not a proven fact. The paper does not justify why topological structure is the right formalization of meaning.
  • domain assumption System 2 reasoning can be implemented as a combination of a world model and an amortized inference machine (for example GFlowNets).
    Assumed in Section III-A and Appendix C. The paper points to literature but provides no demonstration that this combination achieves compositional generalization or sample efficiency in communication tasks.
  • domain assumption Sheaf theory provides an effective and principled way to glue heterogeneous semantic representations across agents and modalities.
    Central to RT1 and Section V-B. The paper references sheaf theory [39] and a preliminary work [40], but no evidence is given in this paper that sheaf-based composition scales or preserves task-relevant semantics.
  • domain assumption Agents can learn robust and grounded emergent languages through cooperative POMDP training with bisimulation abstractions.
    Section V-C proposes this as the basis for protocol learning. This is a substantial empirical assumption that is not tested or supported by simulation in the paper.
invented entities (1)
  • System 2 Semantic Communication (System 2 SC)
    purpose: A label for the proposed research paradigm that contrasts statistical, reconstruction-based communication (System 1 SC) with reasoning-driven, goal-oriented communication (System 2 SC).
    The paper defines this category based on the authors' prior proposal [51]. It is a conceptual framing, not a new physical or mathematical entity, and it does not carry a falsifiable prediction of its own.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic Communication meets System 2 ML: How Abstraction, Compositionality and Emergent Languages Shape Intelligence." pith.science (2026). https://pith.science/paper/7FCYEEFR

@misc{pith2026250520964,
  author       = {Pith},
  title        = {Pith review of: Semantic Communication meets System 2 ML: How Abstraction, Compositionality and Emergent Languages Shape Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FCYEEFR}},
  note         = {Machine review of arXiv:2505.20964}
}
read the original 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.

Figures

Figures reproduced from arXiv: 2505.20964 by the authors.

Figure 1
Figure 1. Shannon’s three levels of communication [3]. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Various shades of information. One of the key distinctions between Shan￾non’s three levels of communication pertains to the notion of information. In contrast to Shannon information (a scalar value asso￾ciated 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 Math￾em… view at source ↗
Figure 3
Figure 3. Daniel Kahneman (System 1 vs. System 2) [2], Michael S.A. Graziano (world model and metacognition), Marvin Minsky [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Confluence of abstraction, algebraic compositionality and semantic communication. Here, the semantics of active inference [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Three Key Pillars underly￾ing the proposed vision. Abstraction (Pillar 1) • How can agents learn abstractions (equivalence classes, relations, partitions, programs, etc.) from their multimodal sensorimotor signals through interactive experience? • How can agents plan v…
Figure 6
Figure 6. Figure 6: Abstractions along the statistical, logical, dynamical and topological continuum. This section outlines the essential requirements and design principles for emergent reasoning-driven compositional systems. Building on the concepts introduced in previous sections, we el…
Figure 7
Figure 7. Figure 7: Algebraic composition of the semantics of [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Sheaf-theoretic approach to communication: [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Simplicial complexes, persistent homological [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Theory of Goal-Oriented Medium Access: Protocol Design and Distributed Bandit Learning

    cs.NI 2025-08 conditional novelty 7.0 of 10

    For a shared channel where each node knows the value of its own data, the optimal distributed access rule is a threshold, and the authors provide algorithms that reach it.

Reference graph

Works this paper leans on

75 extracted references · 55 canonical work pages · cited by 1 Pith paper

  1. [53]

    Semantics-native communication via contextual reasoning,

    H. Seo, J. Park, M. Bennis, and M. Debbah, “Semantics-native communication via contextual reasoning,”IEEETransactions on Cognitive Communications and Networking, pp. 1–1, 2023

    work page 2023

  2. [54]

    Bayesian inverse contextual reasoning for heterogeneous semantics- native communication,

    H. Seo, Y. Kang, M. Bennis, and W. Choi, “Bayesian inverse contextual reasoning for heterogeneous semantics- native communication,” IEEE Transactions on Communications, vol. 72, no. 2, pp. 830–844, 2024

    work page 2024

  3. [1]

    A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,

    W. Saad, M. Bennis, and M. Chen, “A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,” IEEE network, vol. 34, no. 3, pp. 134–142, 2019

    2019

  4. [2]

    Kahneman, Thinking, Fast and Slow

    D. Kahneman, Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011

    work page 2011

  5. [3]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,”The Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  6. [4]

    The bandwagon (edtl.),

    C. Shannon, “The bandwagon (edtl.),”IRE Transactions on Information Theory, vol. 2, no. 1, pp. 3–3, 1956

    work page 1956

  7. [5]

    Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path?,

    Y. Bengio, M. Cohen, D. Fornasiere, J. Ghosn, P. Greiner, M. MacDermott, S. Mindermann, A. Oberman, J. Richardson, O. Richardson, M.-A. Rondeau, P.-L. St-Charles, and D. Williams-King, “Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path?,”arXiv preprint arXiv: 2502.15657, 2025

    arXiv 2025

  8. [6]

    Renè thom: forms, catastrophes and complexity,

    A. Rossi, “Renè thom: forms, catastrophes and complexity,” 2011

    work page 2011

Show all 75 references
  1. [7]

    Topological information data analysis,

    P. Baudot, M. Tapia, D. Bennequin, and J.-M. Goaillard, “Topological information data analysis,”Entropy, vol. 21, no. 9, p. 869, 2019

    2019

  2. [8]

    The homological nature of entropy,

    P. Baudot and D. Bennequin, “The homological nature of entropy,”Entropy, vol. 17, no. 5, pp. 3253–3318, 2015

    2015

  3. [9]

    Semantic information,

    Y. Bar-Hillel and R. Carnap, “Semantic information,”The British Journal for the Philosophy of Science, vol. 4, no. 14, pp. 147–157, 1953

    1953

  4. [10]

    Modal logic,

    P. Blackburn, M. de Rijke, and Y. Venema, “Modal logic,” 2001

    2001

  5. [11]

    Noë, Action in perception

    A. Noë, Action in perception. MIT press, 2004

    2004

  6. [12]

    Deep learning: A critical appraisal,

    G. Marcus, “Deep learning: A critical appraisal,”arXiv preprint arXiv: 1801.00631, 2018

    2018 arXiv

  7. [13]

    A complexity-based theory of compositionality,

    E. Elmoznino, T. Jiralerspong, Y. Bengio, and G. Lajoie, “A complexity-based theory of compositionality,” 2024

    2024

  8. [14]

    Geometric signatures of compositionality across a language model’s lifetime,

    J. H. Lee, T. Jiralerspong, L. Yu, Y. Bengio, and E. Cheng, “Geometric signatures of compositionality across a language model’s lifetime,” 2024

    2024

  9. [15]

    Green ai,

    R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green ai,”arXiv preprint arXiv: 1907.10597, 2019

    1907 arXiv

  10. [16]

    C. J. Buckner, From deep learning to rational machines: What the history of philosophy can teach us about the future of artificial intelligence. Oxford University Press, 2024

    2024

  11. [17]

    Neurosymbolic ai: The 3rd wave,

    A. Garcez and L. Lamb, “Neurosymbolic ai: The 3rd wave,”Artificial Intelligence Review, 2020

    2020

  12. [18]

    Minsky,The Society of Mind

    M. Minsky,The Society of Mind. Simon and Schuster, 1986

    1986

  13. [19]

    Hawkins, A Thousand Brains: A New Theory of Intelligence

    J. Hawkins, A Thousand Brains: A New Theory of Intelligence. Basic Books, 2021. 13

    2021

  14. [20]

    A free energy principle for the brain,

    K. Friston, J. Kilner, and L. Harrison, “A free energy principle for the brain,”Journal of Physiology-Paris, vol. 100, no. 1, pp. 70–87, 2006. Theoretical and Computational Neuroscience: Understanding Brain Functions

    2006

  15. [21]

    On the principles of parsimony and self-consistency for the emergence of intelligence,

    Y. Ma, D. Tsao, and H.-Y. Shum, “On the principles of parsimony and self-consistency for the emergence of intelligence,” 2022

    2022

  16. [22]

    Human-level concept learning through probabilistic program induction,

    B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum, “Human-level concept learning through probabilistic program induction,” Science, vol. 350, no. 6266, pp. 1332–1338, 2015

    2015

  17. [23]

    Computational models of analogy,

    D. Gentner and K. D. Forbus, “Computational models of analogy,”Wiley interdisciplinary reviews: cognitive science, vol. 2, no. 3, pp. 266–276, 2011

    2011

  18. [24]

    Building machines that learn and think like people,

    B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman, “Building machines that learn and think like people,” Behavioral and brain sciences, vol. 40, p. e253, 2017

    2017

  19. [25]

    The consciousness prior,

    Y. Bengio, “The consciousness prior,”arXiv preprint arXiv:1709.08568, 2017

    2017 arXiv

  20. [26]

    World models,

    D. Ha and J. Schmidhuber, “World models,”CoRR, vol. abs/1803.10122, 2018

    2018 arXiv

  21. [27]

    A path towards autonomous machine intelligence,

    Y. LeCun, “A path towards autonomous machine intelligence,” 06 2022

    2022

  22. [28]

    Gflownet foundations,

    Y. Bengio, S. Lahlou, T. Deleu, E. J. Hu, M. Tiwari, and E. Bengio, “Gflownet foundations,”J. Mach. Learn. Res., vol. 24, Mar. 2024

    2024

  23. [29]

    Scaling in the service of reasoning & model-based ml,

    Y. Bengio and E. J. Hu, “Scaling in the service of reasoning & model-based ml,” 2023. Unpublished blog post

    2023

  24. [30]

    Biological sequence design with gflownets,

    M. Jain, Y. Bengio, A. Hernandez-Garcia, J. Rector-Brooks, B. F. P. Dossou, C. A. Ekbote, J. Zhang, N. Malkin, D. Zhang, L. Simine,et al., “Biological sequence design with gflownets,”arXiv preprint arXiv:2203.04115, 2022

    2022 arXiv

  25. [31]

    Can large language models reason and plan?,

    S. Kambhampati, “Can large language models reason and plan?,”Annals of the New York Academy of Sciences, vol. 1534, p. 15–18, Mar. 2024

    2024

  26. [32]

    Inverse scaling: When bigger isn’t better,

    I. R. McKenzie, A. Lyzhov, M. M. Pieler, A. Parrish, A. Mueller, A. Prabhu, E. McLean, X. Shen, J. Cavanagh, A. G. Gritsevskiy, et al., “Inverse scaling: When bigger isn’t better,”Transactions on Machine Learning Research

  27. [33]

    Measuring compositionality in representation learning,

    J. Andreas, “Measuring compositionality in representation learning,” inInternational Conference on Learning Representa- tions, 2019

    2019

  28. [34]

    Emergent multi-agent communication in the deep learning era,

    A. Lazaridou and M. Baroni, “Emergent multi-agent communication in the deep learning era,” arXiv preprint arXiv:2006.02419, 2020

    2006 arXiv

  29. [35]

    Resilient-by-design: A resiliency framework for future wireless networks,

    N. H. Mahmood, S. Samarakoon, P. Porambage, M. Bennis, and M. Latva-aho, “Resilient-by-design: A resiliency framework for future wireless networks,” 2024

    2024

  30. [36]

    An STL-based approach to resilient control for cyber-physical systems,

    H. Chen, S. A. Smolka, N. Paoletti, and S. Lin, “An STL-based approach to resilient control for cyber-physical systems,” in Proceedings of the 26th ACM International Conference on Hybrid Systems: Computation and Control, pp. 1–12, 2023

    2023

  31. [37]

    Semanticandlogicalcommunication-controlcodesignforcorrelateddynamical systems,

    A.M.Girgis,H.Seo,J.Park,andM.Bennis,“Semanticandlogicalcommunication-controlcodesignforcorrelateddynamical systems,” IEEE Internet of Things Journal, vol. 11, no. 7, pp. 12631–12648, 2024

    2024

  32. [38]

    Learning latent wireless dynamics from channel state information,

    C. B. Chaaya, A. M. Girgis, and M. Bennis, “Learning latent wireless dynamics from channel state information,”IEEE Wireless Communications Letters, pp. 1–1, 2024

    2024

  33. [39]

    G. E. Bredon,Sheaf Theory, vol. 170 ofGraduate Texts in Mathematics. New York: Springer-Verlag, second ed., 1997

    1997

  34. [40]

    Tackling feature and sample heterogeneity in decentralized multi-task learning: A sheaf-theoretic approach,

    C. B. Issaid, P. Vepakomma, and M. Bennis, “Tackling feature and sample heterogeneity in decentralized multi-task learning: A sheaf-theoretic approach,” 2025

    2025

  35. [41]

    A mathematical characterization of minimally sufficient robot brains,

    B. Sakcak, K. G. Timperi, V. Weinstein, and S. M. LaValle, “A mathematical characterization of minimally sufficient robot brains,” 2023

    2023

  36. [42]

    An internal model principle for robots,

    V. K. Weinstein, T. Alshammari, K. G. Timperi, M. Bennis, and S. M. LaValle, “An internal model principle for robots,” CoRR, vol. abs/2406.11237, 2024

    2024 arXiv

  37. [43]

    Neuro-symbolic artificial intelligence: The state of the art,

    P. Hitzler and M. K. Sarker, “Neuro-symbolic artificial intelligence: The state of the art,” 2022

    2022

  38. [44]

    Neurosymbolic ai: The 3 rd wave,

    A. d. Garcez and L. C. Lamb, “Neurosymbolic ai: The 3 rd wave,”Artificial Intelligence Review, vol. 56, no. 11, pp. 12387– 12406, 2023

    2023

  39. [45]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou,et al., “Chain-of-thought prompting elicits reasoning in large language models,”Advances in neural information processing systems, vol. 35, pp. 24824–24837, 2022

    2022

  40. [46]

    Port:Preferenceoptimizationonreasoningtraces,

    S.Lahlou,A.Abubaker,andH.Hacid,“Port:Preferenceoptimizationonreasoningtraces,” arXivpreprintarXiv:2406.16061 , 2024

    2024 arXiv

  41. [47]

    From raw data to structural semantics: Trade-offs among distortion, rate, and inference accuracy,

    C. Asirimath, C. Weeraddana, S. Samarakoon, J. Ratnayake, and M. Bennis, “From raw data to structural semantics: Trade-offs among distortion, rate, and inference accuracy,” 2024

    2024

  42. [48]

    Dehaene, Consciousness and the brain: Deciphering how the brain codes our thoughts

    S. Dehaene, Consciousness and the brain: Deciphering how the brain codes our thoughts. Penguin, 2014

    2014

  43. [49]

    Discrete, compositional, and symbolic representations through attractor dynamics,

    A. Nam, E. Elmoznino, N. Malkin, J. McClelland, Y. Bengio, and G. Lajoie, “Discrete, compositional, and symbolic representations through attractor dynamics,”arXiv preprint arXiv: 2310.01807, 2023

    2023 arXiv

  44. [50]

    Bayesian structure learning with generative flow networks,

    T. Deleu, A. G’ois, C. C. Emezue, M. Rankawat, S. Lacoste-Julien, S. Bauer, and Y. Bengio, “Bayesian structure learning with generative flow networks,”Conference on Uncertainty in Artificial Intelligence, 2022

    2022

  45. [51]

    M. bennis, 6g: The next horizon – from connected people and things to connected intelligence, tuesday 30 march 2021, ieee wcnc, nanjing, china

    “M. bennis, 6g: The next horizon – from connected people and things to connected intelligence, tuesday 30 march 2021, ieee wcnc, nanjing, china.” https://wcnc2021.ieee-wcnc.org/panels.html

    2021

  46. [52]

    W. Saad, C. Chaccour, C. K. Thomas, and M. Debbah,Foundations of Semantic Communication Networks. WILEY, 2025

    2025

  47. [55]

    Relational inductive biases, deep learning, and graph networks,

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Malinowski, A. Tacchetti, D. Ra- poso, A. Santoro, R. Faulkner,et al., “Relational inductive biases, deep learning, and graph networks,”arXiv preprint arXiv:1806.01261, 2018. 14

    2018 arXiv

  48. [56]

    Recurrent world models facilitate policy evolution,

    D. Ha and J. Schmidhuber, “Recurrent world models facilitate policy evolution,”Advances in neural information processing systems, vol. 31, 2018

    2018

  49. [57]

    A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel,et al., “A general reinforcement learning algorithm that masters chess, shogi, and go through self-play,”Science, vol. 362, no. 6419, pp. 1140–1144, 2018

    2018

  50. [58]

    Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks,

    B. Lake and M. Baroni, “Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks,” inInternational conference on machine learning, pp. 2873–2882, PMLR, 2018

    2018

  51. [59]

    Invariant risk minimization,

    M. Arjovsky, L. Bottou, I. Gulrajani, and D. Lopez-Paz, “Invariant risk minimization,”arXiv preprint arXiv:1907.02893, 2019

    1907 arXiv

  52. [60]

    Pearl, Causality

    J. Pearl, Causality. Cambridge university press, 2009

    2009

  53. [61]

    Toward causal representation learning,

    B. Schölkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y. Bengio, “Toward causal representation learning,” Proceedings of the IEEE, vol. 109, no. 5, pp. 612–634, 2021. Acknowledgments The authors express their gratitude to Yoshua Bengio for his valuable...

    2021

  54. [62]

    Thinking, Fast and Slow

    Dual-Process Theory of Cognition The dual-process theory, popularized by psychologist Daniel Kahneman in his book “Thinking, Fast and Slow” [2], distinguishes between two modes of thought: • System 1operates automatically, quickly, with little or no effort, and no sense of vol...

  55. [63]

    The systems are pattern-based, statistical, and trained on large datasets to respond quickly to familiar inputs

    Relevance to Communication Systems and AI Current communication systems and AI predominantly employ what could be considered System 1-like processing. The systems are pattern-based, statistical, and trained on large datasets to respond quickly to familiar inputs. However, they...

  56. [64]

    Implementation Challenges Implementing true System 2-like capabilities in machines remains challenging. Current approaches include: • Neuro-symbolic AI: Combining neural networks (System 1-like pattern recognition) with symbolic rea- soning (System 2-like logical processing) [...

  57. [65]

    However, this approach discards the structural relationships within data, which are the very elements that often carry meaning

    From Numbers to Shapes Shannon’s information theory quantifies information as a number (bits) derived from probability distribu- tions. However, this approach discards the structural relationships within data, which are the very elements that often carry meaning. Topological i...

  58. [66]

    • Homology: A method to assign algebraic structures (like groups) to topological spaces, capturing essential features like holes, voids, and connectivity

    Key Concepts • Topology: The mathematical study of shapes and spaces that are preserved under continuous deformations (stretching, bending, but not tearing). • Homology: A method to assign algebraic structures (like groups) to topological spaces, capturing essential features l...

  59. [67]

    • Multi-scale Analysis: Capturing structures at different scales allows communication systems to adapt to different levels of detail

    Applications to Communication Topological approaches to information offer several advantages for semantic communication systems: • Robust Feature Extraction: Topological features are invariant to many transformations, making them robust descriptors of data. • Multi-scale Analy...

  60. [68]

    Example: Message Understanding Consider two communication systems: • A Shannon-based systemmight accurately transmit all the words in a message but miss the context that gives them meaning. • A Topology-based systemmight identify key structural relationships in the message (wh...

  61. [69]

    the cup is on the table

    Core Components and Principles Our proposed framework integrates two essential elements: world models that represent knowledge about the environment, and inference machinery that reasons with this knowledge. This separation is inspired by cognitive science theories like the Gl...

  62. [70]

    Generative Flow Networks for Compositional Reasoning Generative Flow Networks (GFlowNets) [28] are a relatively recent framework for learning to sample from complex probability distributions. Unlike discriminative models that map inputs to outputs, or generative models that pr...

  63. [71]

    cup,” “table,

    Learning Meaningful Representations The framework learns to represent concepts as low-dimensional attractors in a dynamical system [49]. These attractors function like symbols in a compositional language: • Discrete concepts:Each attractor corresponds to a fundamental concept ...

  64. [72]

    Key Advantages This approach offers several advantages over traditional deep learning methods: • Combinatorial generalization: The ability to combine learned concepts in novel ways to solve new problems • Sample efficiency:Structured reasoning reduces the need for extensive tr...

  65. [73]

    to causal reasoning [50]. D. System 1 SC vs. System 2 SC Despite the large body of articles, the topic of semantic communication remains very fragmented in terms of what the word semantic means (beyond its reduction to Greek etymology), what semantic communica- tion means or e...

    2021

  66. [74]

    This boils down to applying blackbox ML to any communication problem across different OSI layers (e.g., physical and MAC layers)

    System 1 SC: Current Approaches System 1 SC encapsulates all the current progress found in the literature (a recent book on the topic can be found in [52]). This boils down to applying blackbox ML to any communication problem across different OSI layers (e.g., physical and MAC...

    1952

  67. [75]

    three cars in left lane, one pedestrian crossing

    System 2 SC: Future Directions In contrast, System 2 SC goes beyond System 1 SC in many ways. First, in terms of targeted use cases (e.g., human-machine collaboration), then in terms of its cognitive capabilities rooted in logical reasoning, planning, abstraction and analogy-m...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.