Pith. sign in

REVIEW 6 major objections 8 minor 47 references

Compositional Semantic Communication for Physical AI: Category Theory Meets Game Theory

T0 review · 6 major / 8 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read By treating semantic communication as composition rather than compression, and coordinating devices through a Stackelberg game, the paper claims wireless AI systems can cut bandwidth by up to 17% and latency by 53% while keeping 85% inferen

desk verdict A genuinely novel conceptual framework for compositional semantic communication, undone by an uncomputable information measure, game-theoretic results that don't apply to neural-network strategy spaces, and an ADMM algorithm that contradicts the paper's own incomparable-spaces premise. read the letter →

arxiv 2607.18115 v1 pith:FBVONJA2 submitted 2026-07-20 cs.IT cs.GTeess.SPmath.IT

Christo Kurisummoottil Thomas , Walid Saad , Emilio Calvanese Strinati This is my paper
classification cs.ITcs.GTeess.SPmath.IT MSC 94A1591A6518B25
keywords semanticcommunicationcompositionalsemanticscategorytheoryGrothendiecktopologyStackelberggameADMMremoteinferencephysicalAI
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 is trying to establish that the bottleneck in multi-device semantic communication is not compression but composition: representations learned by heterogeneous sensors must be aligned and combined, not just transmitted. Its proposed solution is to formalize compositionality with category-theoretic structures—a semantic information measure, Grothendieck topologies, and fibrational encoders into a shared language—and to treat multi-device coordination as a Stackelberg game solved by ADMM. If the claims hold, distributed physical AI systems could scale to new sensors and novel concept combinations without retraining, while cutting bandwidth by up to 17% and end-to-end latency by 53% at 85% inference accuracy. The paper's central claim is that mathematical structure, rather than more data or deeper end-to-end models, is what makes semantic representations compose.

What carries the argument

The load-bearing object is the compositional semantic information measure S(z), built from the Karoubi envelope of concepts with correlators, which quantifies each device's causal and compositional contribution to inference and splits into task-relevant and generalization terms. Around it, Grothendieck topologies supply consistency axioms—maximality, stability, transitivity—so composed representations remain unambiguously interpretable, and fibrational encoders map each device's latent space into a shared semantic language L via a learned classifier. The Stackelberg game (devices as leaders committing to encoding strategies, the base station as follower composing representations) plus the AD

What would settle it

Run the ADMM protocol on a task where each device's concept extractor is trained independently on disjoint data with no shared semantic vocabulary supplied; if transmitted symbols do not align or the reported bandwidth and latency gains disappear, the framework's load-bearing premise fails.

Watch

Extended reading notes

Core claim

The paper claims that semantic communication can be made compositional: if each device transmits a small representation of its concepts through a fibrational encoder into a shared semantic language, a base station can compose those representations—using category-theoretic consistency—and perform remote inference on concept combinations never seen during training. The paper further claims that the coordination problem is naturally a Stackelberg game whose equilibrium, shown to exist and, under a synergy condition, Pareto optimal, can be computed distributedly by ADMM. In autonomous-driving-style simulations, this yields up to 17% bandwidth reduction and 53% lower end-to-end latency relative t

Load-bearing premise

The framework assumes that concept extraction is a solved problem and that a shared semantic vocabulary with fixed meanings already exists; if either is missing, the encoding and composition pipeline has no place to stand.

Editorial extensions

If this is right

  • If the compositionality claim is right, a device joining an existing fleet only needs to learn its own fibrational encoder into the shared vocabulary; the rest of the system does not retrain.
  • Remote inference can meet tight real-time budgets because devices transmit concepts rather than pixels or point clouds, so end-to-end latency grows slowly with device count.
  • The Stackelberg equilibrium gives each device a principled answer to 'what should I send?': send the representation that is a best response to other devices' choices and to the base station's composition rule.
  • The split of semantic information into task-relevant and off-task parts means a single transmitted representation can support both the current inference and future, unseen tasks.
  • Bandwidth use scales with the number of distinct concepts, not with raw data dimensionality, so adding sensors or concepts is cheaper than in raw or deep joint source-channel coding transmission.

Reading between the lines

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

  • Editorial: if the shared semantic language L is itself learned rather than given, the alignment problem shifts into the fibrational encoder; a natural extension is a bootstrapping or negotiation protocol that builds L from device interactions.
  • Editorial: the Pareto-efficiency proof assumes device concepts are synergistic; with redundant or substitutable sensors the equilibrium may not be efficient, suggesting a redundancy-aware extension beyond the fixed penalty term.
  • Editorial: because the simulations assume concept extraction is solved, the most convincing next test is an end-to-end deployment on real robots where concept extractors are imperfect—the reported gains may shrink in that setting.
  • Editorial: ablating the category-theoretic measure by replacing it with mutual information inside the same game would reveal whether the measure, rather than the game structure, is what drives the reported gains.
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

6 major / 8 minor

Summary. The paper proposes a compositional semantic communication (CSC) framework for multi-device physical AI. It introduces a category-theoretic measure S(z) of compositional semantic information built on lenses, Karoubi envelopes, Grothendieck topologies, and fibrational encoders, and then formulates multi-device coordination as a Stackelberg game between transmitting devices (leaders) and a base station (follower). A distributed ADMM algorithm is presented to compute equilibrium signaling strategies, and simulations on Gymnasium and CARLA claim up to 17% bandwidth reduction, 53% lower end-to-end latency, and 85% inference accuracy. The paper's central assertion is that the categorical measure, the topological consistency conditions, and the game-theoretic equilibrium jointly produce these gains.

Significance. If established, the framework would be a novel and valuable unification of category theory and game theory for semantic communication in physical AI systems. The two-stage evaluation design (a controlled novel-combination generalization test in Gymnasium plus a realistic CARLA setup) is a reasonable empirical strategy, and the paper is transparent about assuming concept extraction is solved (Footnote 1). However, the manuscript currently does not provide a computable, model-independent definition of S(z); Algorithm 1 is inconsistent with the paper's own incomparable-spaces premise; and the equilibrium existence proofs rely on unverified assumptions. Because these issues directly affect the central theoretical and empirical claims, the result is not established in its present form.

major comments (6)
  1. [Section IV-E, Algorithm 1; Section II-A; Section III-D] The consensus update \bar{z} = (1/K)\sum_k z_k and the penalty \|z_k - \bar{z}\|^2 in Eqs. (19)-(21) require a common Euclidean vector space and an arithmetic mean over devices. This contradicts the paper's premise of incomparable, device-specific representation spaces E_1,...,E_K (Section II-A) and the fiber-product composition over semantic symbols L (Eq. (7)). Moreover, the penalty actively drives all z_k toward the same average vector, erasing the distinct concepts the fiber product is meant to compose. The transmitted quantities in Section III-D are symbols \ell_k \in L, while the algorithm averages continuous z_k; the consensus constraint is therefore not over the communicated symbols. Consequently, the 'equilibrium' computed by Algorithm 1 is not a Stackelberg equilibrium of the game in Section IV-B, and the simulated bandwidth/latency improvements cannot be attributed to the prop
  2. [Section III-B, Eqs. (3)-(5); Footnote 2] S(z) and S(y;z) are defined in terms of P(c|z), P(c'|c,π), and p(B|c,c',π), which are said to be approximated by learned energy functions E_ψ and E_ϕ (footnote 2). No training procedure, architecture, or correctness guarantee is provided, so S is not a well-defined measure independent of a learned model; it is whatever the learned energy network outputs. This circularity undermines the claim that S captures causal/compositional information beyond mutual information, and it breaks the continuity and differentiability assumptions invoked in Theorem 2, Proposition 2, and the Dinkelbach derivation.
  3. [Section IV-C, Theorem 2 and Lemma 3] Theorem 2 assumes Θ_k is non-empty, compact, and convex, but Definition 10 sets Θ_k = {(θ_k,w_k): θ_k ∈ R^{|θ|}, w_k ∈ Δ^{D-1}}, so θ_k is unbounded and Θ_k is not compact. If θ_k are neural-network parameters, compactness and convexity are not automatic. Continuity and quasi-concavity of U_k in Eq. (8) require properties of S(y;z_k), H(z_k), and D_r that are not proved; the energy function in Eq. (11) is not shown to make U_k quasi-concave. Lemma 3 invokes Berge's theorem without verifying that the constraint correspondence defined by Eq. (13) is nonempty-valued, compact-valued, and continuous. Theorem 3 also refers to 'Lemma 4', which is never stated.
  4. [Section V-B vs. Section IV-B] The CARLA evaluation describes the proposed strategy as 'the BS is the leader optimizing channel allocation, and each vehicle is a follower selecting semantic concepts' (Section V-B), but Definition 10 defines devices as leaders and the BS as follower. The simulated protocol is therefore a concept-selection/channel-allocation scheme, not the ADMM solution of the Stackelberg game from Section IV. There is no ablation separating the contribution of the category-theoretic measure, the Grothendieck constraints, or the game equilibrium, so the reported latency and bandwidth gains cannot be attributed to the paper's theoretical model. In addition, the abstract's '53% lower latency' is inconsistent with Fig. 5a: 70.8 ms vs. 79.6 ms for cooperative MA is about 11%.
  5. [Appendix A, Theorem 1] The proof of Theorem 1 does not establish the theorem. It asserts that the Maximality Axiom 'ensures completeness,' that the Stability Axiom 'preserves semantic alignment,' and that the Transitivity Axiom 'enables multi-hop semantic reasoning.' These are restatements of the axioms as their purported consequences, not proofs. Since Theorem 1 is the basis for the paper's semantic-consistency claims, this is a logical gap rather than a presentation issue.
  6. [Footnote 1; Section III-D] The framework assumes that concept extraction is a solved problem and that a shared communication language L with fixed semantics exists, but no construction of L is given. The fibrational encoder p#_k is implemented as a learned classifier p(ℓ|z_k) in Eq. (6), yet no method guarantees that the same symbol ℓ has identical semantics across heterogeneous devices or that L can be learned without centralized access to all device data. Because the fiber-product alignment in Lemma 2 is defined over L, the entire composition argument is conditional on an unsupported assumption.
minor comments (8)
  1. [Section II-C, Eq. (1)] C(z_k) is defined using R(c_{k,d}), but c_{k,d} is never defined; the relationship between the index d, the number of concepts D, and the simplex dimension D in Eq. (12) is unclear.
  2. [Section III-D, Definition 8 and Eq. (6)] The fibrational encoder p#_k is introduced as a functor, but Eq. (6) defines it as the arg max of a learned classifier. The relation between the two is not specified, and functoriality of the classifier is not verified.
  3. [Section IV-D, Eq. (16)] Equation (16) is displayed but not referenced in the text; Proposition 1 refers to a first-order condition that is not clearly connected to this garbled equation.
  4. [Theorem 3 and Appendix D] The statement and proof refer to 'Lemma 4', which does not exist; the intended reference is presumably Lemma 3.
  5. [References] Reference [41] is cited for both the Gymnasium intersection-v1 environment (Section V-A) and the COCO dataset (Section V-B); only the COCO citation matches the reference list entry.
  6. [Abstract and Section V-B] The abstract claims 'up to 17% bandwidth reduction,' while Section V-B and Fig. 6a report up to 99% bandwidth reduction; this discrepancy should be explained in the text.
  7. [Algorithm 1, line 2] The dual variable is initialized as λ_k = 0, but λ_k already denotes a utility weight in Eq. (8); reusing the symbol for the ADMM multiplier is confusing and should be changed.
  8. [Footnote 2] The footnote says the probabilities are approximated by learned energy functions 'see [34]', but [34] is a category-theory paper, not an energy-based-model reference; the citation appears mismatched.

Circularity Check

2 steps flagged · score 6.0 of 10

Two core theoretical pillars reduce to their own inputs: Theorem 1 restates the Grothendieck axioms, and the semantic information measure is defined through learned energy functions then optimized as a design objective.

  1. self definitional [Section III-C, Definition 7 and Theorem 1; Appendix A]
    "Theorem 1. If transmitted representations {z_k}_{k in K} satisfy the Grothendieck topology consistency conditions of Definition 7 on representation space C_repr, then: 1) Consistency: The composed representation z = L_k z_k is uniquely interpretable at the BS (no semantic ambiguity). 2) Robustness to device dynamics: If a device k leaves the system, the sieve S_y shrinks but remains well-defined (Stability Axiom)... Proof: See Appendix A. ... The Maximality Axiom ensures completeness ... The Stability Axiom ensures ... The Transitivity Axiom ensures ... Together, these three axioms guarantee t"

    The theorem's conclusions are exactly the axioms of Definition 7 restated with new labels. Appendix A does not derive consistency, robustness, or multi-hop relevance; it recites Maximality, Stability, and Transitivity and declares that together they guarantee the conclusions. Consistency and compositional closure are loaded into the Grothendieck topology definition, so Theorem 1 is a renaming of the axioms as a 'formal guarantee' rather than an independent derivation. This is load-bearing because the paper uses Theorem 1 to assert that the categorical framework ensures consistency.

  2. fitted input called prediction [Section III-B, Definition 4, Eqs. (3)-(4), Footnote 2; Section IV-A Eq. (8)]
    "Definition 4. The compositional semantic information conveyed by a representation z is S(z)= sum_{B in P(C)} sum_{c in C_bar} P(c|z)S(c;B)1(p(c)=B). ... S(c;B)= sum_{pi in Pi} P(pi) sum_{c' in C\c} P(c'|c,pi)[-ln p(B|c,c',pi)] p(B|c,c',pi). 2In practice, P(c|z) and p(B|c,c',pi) are approximated by learned energy functions E_psi and E_phi; see [34]."

    The proposed 'compositional semantic information' S is a weighted conditional self-information computed from P(c|z), P(c'|c,pi), and p(B|c,c',pi), which the paper explicitly says are approximated by learned energy functions. The TX utility in Eq. (8) then maximizes alpha_k times S(y;z_k), so the quantity being optimized is the learned model's own conditional distribution. The equilibrium's high S is therefore a property of the fitted energy model, not an independent information-theoretic measure; the claimed 'beyond mutual information' content is not tested independently of the fitted inputs.

full rationale

The paper's headline empirical claims (up to 17% bandwidth reduction, 53% lower latency, 85% accuracy) come from simulations, and those simulations are not themselves derived from the circular definitions, so this is not a full reduction of the paper to its inputs. However, two central theoretical pillars are partial reductions. Theorem 1 is a restatement of the Grothendieck topology axioms, and the compositional semantic information measure is defined through learned energy functions and then used as the optimization objective, making high S at equilibrium true by construction. These are the main circularity findings. Several additional load-bearing gaps are correctness/implementation issues rather than circularity and therefore do not appear as steps: Algorithm 1 averages device representations z_k in a common Euclidean space and penalizes the squared norm between z_k and the average z_bar, which is undefined if the spaces E_1,...,E_K are truly incomparable as stated in Section II-A; no convergence proof to the Stackelberg equilibrium is given; and the CARLA implementation describes the BS as leader while Section IV defines devices as leaders. No load-bearing self-citation chain was found: the cited prior works by the present authors are background, and the external proof cited in Lemma 1 is not a self-citation. Overall, the theoretical 'guarantees' reduce partly to definitions and fitted models, but substantial independent algorithmic and experimental content remains, so the circularity score is 6 rather than higher.

Assumptions & free parameters 4 free parameters · 7 assumptions · 3 invented entities

The central claim rests on many unproven assumptions. The 'novel measure' is defined in terms of learned functions, making it circular. The game-theoretic results rely on compactness/convexity that neural network parameter spaces do not satisfy.

free parameters (4)
  • α_k, β_k, γ_k, λ_k
    Trade-off weights in utility function (8); chosen by system designer, no values given.
  • C_min
    Minimum compositionality constraint in (8) and (12); hand-chosen threshold.
  • ρ
    ADMM penalty parameter in Algorithm 1; hand-chosen.
  • Energy function parameters E_ψ, E_ϕ
    Used to approximate probabilities in (3)-(4) and (10)-(11); learned from data but not specified.
assumptions (7)
  • standard math Standard category theory definitions (lens, Grothendieck topology, fibration)
    Used throughout Section III; assumed background without proof.
  • standard math Rosen's theorem for NE existence
    Invoked in Appendix C; requires compact convex strategies and quasi-concave utilities.
  • standard math Berge's Maximum Theorem
    Used in Lemma 3; requires continuity of objective and constraint correspondence.
  • standard math Dinkelbach's method for fractional programming
    Used in Stage 2 solution (IV-D1).
  • domain assumption Concept extraction is a solved problem
    Footnote 1 states this assumption explicitly; the entire framework depends on it.
  • domain assumption A shared communication language L exists
    Assumed in Section III-D; no construction or learning procedure is given beyond arg max classification.
  • domain assumption Probabilities P(c|z), p(B|c,c',π) are known/learnable
    Definition of S(z) in (3)-(4) requires these; footnote 2 says they are approximated by energy functions.
invented entities (3)
  • Communication language L
    purpose: Shared vocabulary of semantic symbols that all devices map to
    No falsifiable handle; only defined abstractly in Section III-D.
  • Correlator η
    purpose: Composition rule defining how concepts combine in the Karoubi envelope
    Introduced in Definition 3; no empirical content specified.
  • Fibrational encoder p#_k
    purpose: Maps device-specific representations to the shared language L
    Defined in Definition 8; implementation is a learned classifier (6) with no guarantees.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compositional Semantic Communication for Physical AI: Category Theory Meets Game Theory." pith.science (2026). https://pith.science/paper/FBVONJA2

@misc{pith2026260718115,
  author       = {Pith},
  title        = {Pith review of: Compositional Semantic Communication for Physical AI: Category Theory Meets Game Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FBVONJA2}},
  note         = {Machine review of arXiv:2607.18115}
}
read the original abstract

Physical artificial intelligence (AI) systems involve distributed sensing agents with embedded AI models that must coordinate to perceive, reason, and act in networked environments. Transmitting raw sensor data incurs significant communication overhead, latency, and redundancy. While semantic communication (SC) mitigates these challenges by transmitting task-relevant information, existing deep learning-based joint source-channel coding approaches exhibit limited adaptability, poor out-of-distribution generalization, and scalability challenges. To address these limitations, this paper proposes a framework for compositional semantic communication (CSC), enabling heterogeneous physical AI sources to transmit semantic representations (SRs) that compose meaningfully at a base station (BS) or edge server for remote inference. First, a category-theoretic measure of compositional semantics is developed to quantify each device's contribution to inference tasks beyond mutual information. Second, Grothendieck topologies and presheaves formalize semantic composition across devices, ensuring consistency and task relevance. Building on these foundations, multi-device coordination is formulated as a Stackelberg game in which devices commit to encoding strategies and the BS optimally composes received SRs. An ADMM-based algorithm computes equilibrium signaling strategies. Equilibrium existence is established under mild conditions and is Pareto optimal when compositional information yields increasing collective benefit. Simulation results demonstrate that the proposed approach achieves up to 17% bandwidth reduction and 53% lower end-to-end latency than cooperative multi-agent, distributed gradient descent, and uniform-selection CSC baselines while maintaining 85% inference accuracy across diverse autonomous driving scenarios.

Figures

Figures reproduced from arXiv: 2607.18115 by the authors.

Figure 1
Figure 1. Overview of the proposed CSC system, with applications across physical AI and networking. A. Lenses: Modeling Bidirectional Semantic Transformations In SC, two operations are required. The first operation, g, extracts semantics from data, which corresponds to encoding. The second operation, p, updates or generates data based on semantic extraction, which corresponds to decoding or reconstruction. These operations mu… view at source ↗
Figure 2
Figure 2. Hierarchical semantics via Fibrational semantic language. 3) Multi-hop task relevance: If zk → zcomposed → y (zk contributes through intermediate composition), then zk is task-relevant (Transitivity Axiom). Proof: See Appendix A. The Grothendieck topology at the representation level is induced by Karoubi correlators at the concept level. Specif￾ically, consistency morphisms g : Zj → Zk exist if and only if correlato… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a) Number of bits communicated vs number of sensors, showing the scalability of proposed CSC compared to baselines. (b) Task accuracy vs compression ratio, which is the ratio of number of concepts communicated vs the total extracted semantic concepts. 7.88 MB/s to a c…
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Latency vs compositionality tradeoff. stem from the game-theoretic incentive structure that explic￾itly rewards agents for transmitting compositionally valuable semantic concepts. Unlike distributed GD which optimizes local objectives independently, or cooperative MA a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 6 linked inside Pith

  1. [1]

    Temporal windows of integration for multisensory wireless systems as enablers of physical ai,

    A. Mishra, J. H. I. de Souza, and P. Popovski, “Temporal windows of integration for multisensory wireless systems as enablers of physical ai,” arXiv preprint arXiv:2512.09589, 2025

    arXiv 2025

  2. [2]

    A joint learning and communications framework for federated learning over wireless networks,

    M. Chen, Z. Yang, W. Saad, C. Yin, H. V . Poor, and S. Cui, “A joint learning and communications framework for federated learning over wireless networks,”IEEE Transactions on Wireless Communications, vol. 20, no. 1, pp. 269–283, 2020

    2020

  3. [3]

    A review of applications in federated learning,

    L. Li, Y . Fan, M. Tse, and K. Y . Lin, “A review of applications in federated learning,”Computers & Industrial Engineering, vol. 149, 2020

    2020

  4. [4]

    6g networks: Beyond shannon towards semantic and goal-oriented communications,

    E. C. Strinati and S. Barbarossa, “6g networks: Beyond shannon towards semantic and goal-oriented communications,”Computer Networks, vol. 190, 2021

    2021

  5. [5]

    Semantics-empowered communication for networked intelligent systems,

    M. Kountouris and N. Pappas, “Semantics-empowered communication for networked intelligent systems,”IEEE Communications Magazine, vol. 59, no. 6, pp. 96–102, Jun. 2021

    2021

  6. [6]

    Less Data, More Knowledge: Building Next Generation Semantic Communication Networks,

    C. Chaccour, W. Saad, M. Debbah, Z. Han, and H. V . Poor, “Less Data, More Knowledge: Building Next Generation Semantic Communication Networks,”IEEE Communications Surveys & Tutorials, vol. 27, no. 1, Nov. 2024

    2024

  7. [7]

    Rethinking modern communication from semantic coding to semantic communication,

    K. Lu, Q. Zhou, R. Li, Z. Zhao, X. Chen, J. Wu, and H. Zhang, “Rethinking modern communication from semantic coding to semantic communication,”IEEE Wireless Communications, vol. 30, no. 1, pp. 158–164, 2022

    2022

  8. [8]

    Latent Space Align- ment for Semantic Channel Equalization,

    T. Huttebraucker, M. Sana, and E. C. Strinati, “Latent Space Align- ment for Semantic Channel Equalization,” inProceedings of IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN), Stockholm, Sweden, May 2024

    2024

Show all 47 references
  1. [9]

    Frame- based zero-shot semantic channel equalization for ai-native communi- cations,

    S. Fiorellino, C. Battiloro, E. C. Strinati, and P. D. Lorenzo, “Frame- based zero-shot semantic channel equalization for ai-native communi- cations,”arXiv preprint arXiv:2507.17835, 2025

    2025 arXiv

  2. [10]

    Semantic channel equalizer: Modelling language mismatch in multi-user semantic communications,

    M. Sana and E. C. Strinati, “Semantic channel equalizer: Modelling language mismatch in multi-user semantic communications,” inIEEE Global Communications Conference. IEEE, 2023, pp. 2221–2226

    2023

  3. [11]

    Semantic channel equalization strategies for deep joint source-channel coding,

    L. Pannacci, S. Fiorellino, M. E. Pandolfo, E. C. Strinati, and P. D. Lorenzo, “Semantic channel equalization strategies for deep joint source-channel coding,”arXiv preprint arXiv:2510.04674, 2025

    2025

  4. [12]

    Soft partitioning of latent space for semantic channel equalization,

    T. H ¨uttebr¨aucker, M. Sana, and E. C. Strinati, “Soft partitioning of latent space for semantic channel equalization,” inProceedings of 19th International Symposium on Wireless Communication Systems (ISWCS), 2024

    2024

  5. [13]

    Learning network sheaves for ai-native semantic communi- cation,

    E. Grimaldi, M. E. Pandolfo, G. D’Acunto, S. Barbarossa, and P. Di Lorenzo, “Learning network sheaves for ai-native semantic communi- cation,”arXiv preprint arXiv:2512.03248, 2025

    2025

  6. [14]

    A Theory of Semantic Commu- nication,

    Y . Shao, Q. Cao, and D. Gunduz, “A Theory of Semantic Commu- nication,”IEEE Transactions on Mobile Computing, vol. 23, no. 12, 2024

    2024

  7. [15]

    Resource allocation for text semantic communications,

    L. Yan, Z. Qin, R. Zhang, Y . Li, and G. Y . Li, “Resource allocation for text semantic communications,”IEEE Wireless Communications Letters, vol. 11, no. 07, pp. 1394–1398, 2022

    2022

  8. [16]

    Performance optimization for semantic communications: An attention- based reinforcement learning approach,

    Y . Wang, M. Chen, T. Luo, W. Saad, D. Niyato, H. V . Poor, and S. Cui, “Performance optimization for semantic communications: An attention- based reinforcement learning approach,”IEEE Journal on Selected Areas in Communications, vol. 40, no. 09, pp. 2598–2613, 2022

    2022

  9. [17]

    Task-Oriented Multi- User Semantic Communications ,

    H. Xie, Z. Qin, X. Tao, and K. B. Letaief, “Task-Oriented Multi- User Semantic Communications ,”IEEE Journal on Selected Areas in Communications., Jul. 2022

    2022

  10. [18]

    Non-orthogonal multiple access enhanced multi-user semantic communication,

    W. Li, H. Liang, C. Dong, X. Xu, P. Zhang, and K. Liu, “Non-orthogonal multiple access enhanced multi-user semantic communication,”IEEE Transactions on Cognitive Communications and Networking, vol. 9, no. 6, pp. 1438–1453, 2023

    2023

  11. [19]

    Toward semantic communication protocols: A probabilistic logic perspective,

    S. Seo, J. Park, S. W. Ko, J. Choi, M. Bennis, and S. L. Kim, “Toward semantic communication protocols: A probabilistic logic perspective,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 8, pp. 2670–2686, 2023

    2023

  12. [20]

    Swin transformer-based dynamic semantic communication for multi-user with different computing capacity,

    L. X. Nguyen, Y . L. Tun, Y . K. Tun, M. N. Nguyen, C. Zhang, Z. Han, and C. S. Hong, “Swin transformer-based dynamic semantic communication for multi-user with different computing capacity,”IEEE Transactions on Vehicular Technology, vol. 73, no. 6, pp. 8957–8972, 2024. 13

    2024

  13. [21]

    Neuromorphic wireless cognition: Event-driven semantic communications for remote inference,

    J. Chen, N. Skatchkovsky, and O. Simeone, “Neuromorphic wireless cognition: Event-driven semantic communications for remote inference,” IEEE Transactions on Cognitive Communications and Networking, vol. 9, no. 2, pp. 252–265, 2023

    2023

  14. [22]

    Collaborative semantic communication for edge inference ,

    W. F. Lo, N. Mital, H. Wu, and D. G ¨und¨uz, “Collaborative semantic communication for edge inference ,”IEEE Wireless Communications Letters, vol. 12, no. 07, pp. 1125–1129, 2023

    2023

  15. [23]

    Neuro-Symbolic Causal Reasoning Meets Signaling Game for Emergent Semantic Communications,

    C. K. Thomas and W. Saad, “Neuro-Symbolic Causal Reasoning Meets Signaling Game for Emergent Semantic Communications,”IEEE Transactions on Wireless Communications, Oct. 2023

    2023

  16. [24]

    Causal Semantic Communication for Digital Twins: A Generalizable Imitation Learning Approach,

    C. K. Thomas, W. Saad, and Y . Xiao, “Causal Semantic Communication for Digital Twins: A Generalizable Imitation Learning Approach,” Journal on Selected Areas in Information Theory, vol. 4, Nov. 2023

    2023

  17. [25]

    Causal abstraction learning based on the semantic embedding principle,

    G. D’Acunto, F.M. Zennaro, Y . Felekis, and P. Di Lorenzo, “Causal abstraction learning based on the semantic embedding principle,”arXiv preprint arXiv:2502.00407, 2025

    2025 arXiv

  18. [26]

    Learning consistent causal abstraction networks,

    g. D’Acunto, p. D. Lorenzo, and s. Barbarossa, “Learning consistent causal abstraction networks,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2026

    2026

  19. [27]

    A Survey on Compositional Learning of AI Models: Theoretical and Experimental Practices,

    S. Sinha, T. Premsri, and P. Kordjamshidi, “A Survey on Compositional Learning of AI Models: Theoretical and Experimental Practices,”arXiv preprint arXiv:2406.08787, 2024

    2024 arXiv

  20. [28]

    Predictive learning enables compositional representations,

    G. Boeshertz and C. Clopath, “Predictive learning enables compositional representations,”bioRxiv, 2025

    2025

  21. [29]

    The role of fibration symmetries in geometric deep learning,

    O. Velarde, L. Parra, P. Boldi, and H. Makse, “The role of fibration symmetries in geometric deep learning,”Proceedings of the National Academy of Sciences, vol. 123, no. 4, 2026

    2026

  22. [30]

    Th ´eorie des topos et Cohomologie Etale des Sch ´emas ,

    M. Artin, A. Grothendieck, and J. L. Verdier, “Th ´eorie des topos et Cohomologie Etale des Sch ´emas ,”Lecture Notes in Math, Springer- Verlag, 1972

    1972

  23. [31]

    Topos and stacks of deep neural networks,

    J. C. Belfiore and D. Bennequin, “Topos and stacks of deep neural networks,”arXiv preprint arXiv:2106.14587, 2021

    2021 arXiv

  24. [32]

    Unsupervised learning of compositional energy concepts,

    Y . Du, S. Li, Y . Sharma, J. Tenenbaum, and I. Mordatch, “Unsupervised learning of compositional energy concepts,” inAdvances in Neural Information Processing Systems, 2021, vol. 34, pp. 15608–15620

    2021

  25. [33]

    Lenses in functional programming,

    A. Steckermeier, “Lenses in functional programming,” Preprint, avail- able at https://sinusoid.es/misc/lager/lenses.pdf, 2015

    2015

  26. [34]

    Categories of Semantic Concepts ,

    J. Hefford, V . Wang, and M. Wilson, “Categories of Semantic Concepts ,”arxiv preprint arXiv:2004.10741, 2020

    2004 arXiv

  27. [35]

    Stackelberg game for utility-based cooperative cognitive radio networks,

    J. Zhang and Q. Zhang, “Stackelberg game for utility-based cooperative cognitive radio networks,” inProceedings of the Tenth ACM Interna- tional Symposium on Mobile Ad Hoc Networking and Computing, May 2009, pp. 23–32

    2009

  28. [36]

    Managing price uncertainty in prosumer-centric energy trading: A prospect-theoretic stackelberg game approach,

    G. El Rahi, S. R. Etesami, W. Saad, N. B. Mandayam, and H. V . Poor, “Managing price uncertainty in prosumer-centric energy trading: A prospect-theoretic stackelberg game approach,”IEEE Transactions on Smart Grid, vol. 10, no. 1, pp. 702–713, Jan. 2019

    2019

  29. [37]

    A generalization of the maximum theorem,

    Mark Walker, “A generalization of the maximum theorem,”International Economic Review, vol. 20, no. 2, pp. 267–272, 1979

    1979

  30. [38]

    On nonlinear fractional programming,

    W. Dinkelbach, “On nonlinear fractional programming,”Management Science, vol. 13, no. 7, pp. 492–498, 1967

    1967

  31. [39]

    Gymnasium: A standard interface for reinforcement learning environ- ments,

    M. Towers, A. Kwiatkowski, J. Terry, J. U. Balis, G. De Cola, T. Deleu, M. Goul ˜ao, A. Kallinteris, M. Krimmel, A. KG, and R. Perez-Vicente, “Gymnasium: A standard interface for reinforcement learning environ- ments,”arXiv preprint arXiv:2407.17032, 2024

    2024 arXiv

  32. [40]

    Carla: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “Carla: An open urban driving simulator,” inProceedings of the Conference on Robot Learning. October 2017, pp. 1–16, PMLR

    2017

  33. [41]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inEuropean Conference on Computer Vision (ECCV), Cham, 2014, Springer, pp. 740–755

    2014

  34. [42]

    Existence and uniqueness of equilibrium points for concave n-person games,

    J. B. Rosen, “Existence and uniqueness of equilibrium points for concave n-person games,”Econometrica, vol. 33, no. 3, pp. 520–534, 1965. APPENDIXA PROOF OFTHEOREM1 In our multi-device SC system, an object is a semantic concept or inference targety. The Maximality Axiom en- su...

    1965

  35. [43]

    For composed objectsz 1 ⊕z ′ 1 ∈ Z1 and Z2 ⊕z ′ 2 ∈ Z2, the fibrational structure is preserved when p# 1 (z1 ⊕z ′

    =ℓ ′ (possibly different fromℓ). For composed objectsz 1 ⊕z ′ 1 ∈ Z1 and Z2 ⊕z ′ 2 ∈ Z2, the fibrational structure is preserved when p# 1 (z1 ⊕z ′

  36. [44]

    =p # 1 (z1)⊗p # 1 (z′

  37. [45]

    =ℓ⊗ℓ ′ andp # 2 (z2 ⊕z ′

  38. [46]

    = p# 2 (z2)⊗p # 2 (z′

  39. [47]

    Since both map toℓ⊗ℓ ′, the composed pair remains in the fiber product overℓ⊗ℓ ′, preserving the compositional structure

    =ℓ⊗ℓ ′, where⊗denotes composition inL. Since both map toℓ⊗ℓ ′, the composed pair remains in the fiber product overℓ⊗ℓ ′, preserving the compositional structure. ForKdevices, proof follows by induction. APPENDIXC PROOF OFTHEOREM2 Follows from Rosen’s theorem on concave games [4...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.