arxiv: 2512.22579 · v2 · submitted 2025-12-27 · 💻 cs.AI · cs.NI

Recognition: 2 theorem links

· Lean Theorem

SANet: A Semantic-aware Agentic AI Networking Framework for Cross-layer Optimization in 6G

Yong Xiao , Xubo Li , Haoran Zhou , Yingyu Li , Yayu Gao , Guangming Shi , Ping Zhang , Marwan Krunz

Authors on Pith no claims yet

Pith reviewed 2026-05-16 19:18 UTC · model grok-4.3

classification 💻 cs.AI cs.NI

keywords Agentic AI networkingsemantic-aware optimizationmulti-agent multi-objectivePareto-optimal solutions6G wireless networksmodel partition and sharingcross-layer coordinationdecentralized algorithms

0 comments

The pith

SANet coordinates semantic-aware AI agents across network layers to optimize 6G systems by inferring user goals and finding Pareto-optimal solutions despite conflicting objectives.

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

The paper introduces SANet, a semantic-aware AgentNet architecture in which AI agents specialized for different layers of wireless networks first infer the semantic goal from user interactions and then collaborate to fulfill it. Because agents operate decentrally and may hold distinct or opposing objectives, the framework casts coordination as a multi-agent multi-objective problem whose solution is the Pareto front. It supports this with a model partition and sharing scheme that splits large models into shared and agent-specific parts matched to local hardware, plus two decentralized algorithms. Theoretical analysis establishes bounds on a three-way tradeoff among optimization, generalization, and conflicting errors. A hardware prototype spanning radio access and core networks shows the approach delivers up to 14.61 percent performance improvement while consuming only 44.37 percent of the floating-point operations required by prior methods.

Core claim

SANet infers semantic goals from user interactions, assigns agents tied to distinct network layers, and solves the resulting multi-agent multi-objective problem to locate Pareto-optimal operating points; the approach is realized through model partition and sharing together with two decentralized algorithms, and is accompanied by proven bounds on the tradeoff among optimization error, generalization error, and conflicting error.

What carries the argument

The multi-agent multi-objective optimization that locates Pareto-optimal solutions among agents with potentially conflicting layer-specific objectives, realized via the Model Partition and Sharing (MoPS) framework that decomposes large models into jointly trained shared and agent-specific components.

If this is right

Cross-layer wireless management becomes autonomous, supporting self-configuration and real-time adaptation in complex 6G environments.
Large AI models can be deployed across resource-limited agents without duplicating full computation at every node.
New evaluation metrics become available to quantify performance of agentic networking systems.
The hardware prototype demonstrates that the architecture can be implemented on actual RAN and core network equipment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The semantic-inference front end could be evaluated for robustness against noisy or ambiguous user inputs outside the reported prototype scenarios.
The three-way error tradeoff may generalize to other decentralized multi-agent settings such as edge AI or distributed control systems.
Model partitioning offers a template for efficient deployment of foundation models in any resource-heterogeneous agent collective.
Integration with 6G-native features such as joint communication and sensing could be tested by extending the agent objectives accordingly.

Load-bearing premise

Semantic goals can be inferred accurately enough from user interactions, and decentralized agents can reliably converge to useful Pareto-optimal points even when their objectives conflict.

What would settle it

A controlled trial in which semantic inference accuracy falls below the level used in the prototype or agent objectives are deliberately placed in strong conflict, after which measured network performance and compute cost cease to improve over non-agentic baselines.

Figures

Figures reproduced from arXiv: 2512.22579 by Guangming Shi, Haoran Zhou, Marwan Krunz, Ping Zhang, Xubo Li, Yayu Gao, Yingyu Li, Yong Xiao.

**Figure 1.** Figure 1: (a)-(c) Learning trajectories of three agents deployed at the physical layer (pAgent), application layer (aAgent), and network layer (nAgent) of a wireless network (black ✖ denotes the initialization of the model, green ⋆ denotes the end point); (d) SANet’s collaborative learning trajectories of all three agents using static- and dynamic-weighting algorithms where grey line denotes the Pareto front, black … view at source ↗

**Figure 2.** Figure 2: (a) System model, and (b) operational workflow for [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Model training procedures of (a) static-weighting and (b) dynamic-weighting algorithms, and (c) inference [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: SANet prototype [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: O-error of different agents using MoPS with (a) Scheme 1, (b) Scheme 2 and (c) Scheme 3 trained by the static-weighting algorithm and (d) Scheme 2 and (e) Scheme 3 trained by the dynamic-weighting algorithm. 0 200 400 600 800 1000 Iteration rounds 0.0 0.5 1.0 1.5 2.0 2.5 G-Error aAgent pAgent nAgent (a) 0 200 400 600 800 1000 Coordination rounds 0.0 0.5 1.0 1.5 2.0 2.5 G-Error aAgent pAgent nAgent Joint (b… view at source ↗

**Figure 6.** Figure 6: G-error of different agents using MoPS with (a) Scheme 1, (b) Scheme 2 and (c) Scheme 3 trained by the static-weighting algorithm and (d) Scheme 2 and (e) Scheme 3 trained by the dynamic-weighting algorithm. Scheme 2 (Partial sharing I): Each agent trains a modalityspecific time-series embedding and position encoding module, and the agent controller hosts all the Transformer layers (including 6 encoder/… view at source ↗

**Figure 7.** Figure 7: Prediction of user demand in aAgent based on (a) LSTM, (b) N-BEATS, (c) Transformer, (d) Informer, (e) MoPS, and (f) comparison of different models’ prediction performance based on NMAE. all three agents, e.g., the median NMAE values of MoPS are approximately 0.0109, 0.284, and 0.00267 at the physical, network, and application layers, respectively, corresponding to performance gains of 4.59%, 1.41%, and 14… view at source ↗

**Figure 10.** Figure 10: Convergence performance of different agents based on (a) static- and (b) dynamic-weighting algorithms under Scheme 2. 0 200 400 600 800 1000 Coordination rounds 0.0 0.2 0.4 0.6 0.8 1.0 C-Error (a) 0 200 400 600 800 1000 Coordination rounds 0.0 0.2 0.4 0.6 0.8 1.0 C-Error (b) 0 200 400 600 800 1000 Coordination rounds 0.0 0.2 0.4 0.6 0.8 1.0 C-Error (a) 0 200 400 600 800 1000 Coordination rounds 0.0 0.2 0.… view at source ↗

**Figure 9.** Figure 9: Prediction of the network traffic in nAgent based on (a) LSTM, (b) N-BEATS, (c) Transformer, (d) Informer, (e) MoPS, and (f) comparison of different models’ prediction performance based on NMAE. veals a tradeoff between the local computation load of each agent and overall communication resources for achieving a specific optimization performance. We also present the resource consumption of the inference pha… view at source ↗

**Figure 12.** Figure 12: (a) Tradeoff between G-error and O-error and (b) that between G-error and C-error verified by both theoretical results and experimental results for the static-weighting and dynamicweighting algorithms. We can observe that the experimental results match the trend of the theoretical curves, that is, increasing G-error will cause decreases in both O-error and C-error for both the static-weighting and dynami… view at source ↗

read the original abstract

Agentic AI networking (AgentNet) is a novel AI-native networking paradigm in which a large number of specialized AI agents collaborate to perform autonomous decision-making, dynamic environmental adaptation, and complex missions. It has the potential to facilitate real-time network management and optimization functions, including self-configuration, self-optimization, and self-adaptation across diverse and complex environments. This paper proposes SANet, a novel semantic-aware AgentNet architecture for wireless networks that can infer the semantic goal of the user and automatically assign agents associated with different layers of the network to fulfill the inferred goal. Motivated by the fact that AgentNet is a decentralized framework in which collaborating agents may generally have different and even conflicting objectives, we formulate the decentralized optimization of SANet as a multi-agent multi-objective problem, and focus on finding the Pareto-optimal solution for agents with distinct and potentially conflicting objectives. We propose three novel metrics for evaluating SANet. Furthermore, we develop a model partition and sharing (MoPS) framework in which large models, e.g., deep learning models, of different agents can be partitioned into shared and agent-specific parts that are jointly constructed and deployed according to agents' local computational resources. Two decentralized optimization algorithms are proposed. We derive theoretical bounds and prove that there exists a three-way tradeoff among optimization, generalization, and conflicting errors. We develop an open-source RAN and core network-based hardware prototype that implements agents to interact with three different layers of the network. Experimental results show that the proposed framework achieved performance gains of up to 14.61% while requiring only 44.37% of FLOPs required by state-of-the-art algorithms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SANet gives a workable architecture for semantic-aware agents coordinating across 6G layers with a real prototype, but the 14.61% gain claim at 44% FLOPs lacks checks on inference accuracy and Pareto convergence under conflicts.

read the letter

SANet sets up a system where agents infer a user's semantic goal from interactions and then assign themselves to different network layers to meet it. The paper adds MoPS to split big models into shared and agent-specific pieces based on local compute, three new evaluation metrics, and a proof that optimization, generalization, and conflicting errors trade off in a three-way bound. Two decentralized algorithms handle the multi-agent multi-objective setup when layer goals clash. They also built and ran a hardware prototype on RAN and core network layers with open-source code released.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes SANet, a semantic-aware agentic AI networking framework for 6G that infers users' semantic goals from interactions and assigns specialized agents across network layers for cross-layer optimization. It formulates the decentralized setting as a multi-agent multi-objective problem to locate Pareto-optimal solutions under potentially conflicting layer objectives, introduces a Model Partition and Sharing (MoPS) framework for efficient deployment of large models, proposes two decentralized optimization algorithms, defines three novel evaluation metrics, derives theoretical bounds proving a three-way tradeoff among optimization, generalization, and conflicting errors, and reports hardware-prototype results showing up to 14.61% performance gains while using only 44.37% of the FLOPs required by state-of-the-art algorithms.

Significance. If the experimental claims and theoretical bounds hold under rigorous validation, the work would advance AI-native networking by demonstrating how semantic inference and decentralized Pareto optimization can be combined to reduce computational cost while handling cross-layer conflicts. The open-source prototype and explicit tradeoff bound constitute concrete contributions that could inform future 6G agentic systems.

major comments (3)

[Abstract] Abstract: the central claim of 14.61% gains at 44.37% FLOPs is presented without error bars, number of trials, dataset characteristics, or explicit baseline definitions, preventing assessment of whether the gains are attributable to SANet mechanisms rather than experimental conditions.
[Abstract] Abstract: no quantitative results are supplied on semantic goal inference accuracy or on the quality of Pareto-optimal solutions reached by the two decentralized algorithms under conflicting objectives, leaving the weakest assumptions unverified and the attribution of performance gains unsupported.
[Abstract] Abstract: the existence of a three-way tradeoff bound among optimization, generalization, and conflicting errors is asserted, yet the abstract supplies neither the defining equations nor any numerical validation of the bound against the prototype experiments.

minor comments (2)

The three novel metrics are introduced without names or brief definitions in the abstract, which would improve immediate readability.
Clarify how the MoPS partitioning interacts with the two proposed decentralized algorithms during joint training and inference on the hardware prototype.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We agree that the abstract requires additional details to support the claims and will revise it in the next version to include experimental specifics, quantitative results on key components, and a reference to the theoretical bound.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 14.61% gains at 44.37% FLOPs is presented without error bars, number of trials, dataset characteristics, or explicit baseline definitions, preventing assessment of whether the gains are attributable to SANet mechanisms rather than experimental conditions.

Authors: We agree that the abstract should provide more context on the experimental conditions. In the revised version we will add the number of trials, mention of error bars or variability, dataset characteristics, and explicit baseline definitions. These details appear in the experimental evaluation section but summarizing them in the abstract will improve verifiability. revision: yes
Referee: [Abstract] Abstract: no quantitative results are supplied on semantic goal inference accuracy or on the quality of Pareto-optimal solutions reached by the two decentralized algorithms under conflicting objectives, leaving the weakest assumptions unverified and the attribution of performance gains unsupported.

Authors: We acknowledge this limitation in the current abstract. The revised abstract will report quantitative accuracy figures for semantic goal inference and metrics assessing the quality of the Pareto-optimal solutions produced by the decentralized algorithms. These results are already computed and presented in the main body; we will extract the key numbers for the abstract. revision: yes
Referee: [Abstract] Abstract: the existence of a three-way tradeoff bound among optimization, generalization, and conflicting errors is asserted, yet the abstract supplies neither the defining equations nor any numerical validation of the bound against the prototype experiments.

Authors: The three-way tradeoff is formally derived and proven in the theoretical analysis section, with numerical validation against the hardware prototype results. We will revise the abstract to include a concise statement of the bound and note its empirical validation, thereby making the theoretical contribution more self-contained in the summary. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation chain is self-contained and independent of fitted inputs or self-citations

full rationale

The paper introduces SANet as a new semantic-aware AgentNet architecture, formulates decentralized optimization as a multi-agent multi-objective problem, defines three novel evaluation metrics, develops the MoPS model partitioning framework, proposes two decentralized algorithms, and derives a three-way tradeoff bound among optimization, generalization, and conflicting errors. The abstract and described structure contain no equations that reduce claimed performance gains (e.g., 14.61% improvement or 44.37% FLOPs) to quantities defined by construction from the same experimental data or fitted parameters. No load-bearing steps rely on self-citations for uniqueness theorems or ansatzes; the theoretical bounds and hardware prototype results are presented as independently derived from the proposed framework. This is the normal case of a self-contained proposal without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 3 invented entities

The central claims rest on several newly introduced concepts whose independent support is not shown in the abstract.

axioms (2)

domain assumption Decentralized agents may have distinct and conflicting objectives
Used to motivate the multi-agent multi-objective formulation
domain assumption Semantic goals of users can be reliably inferred
Core premise enabling agent assignment in SANet

invented entities (3)

SANet no independent evidence
purpose: Semantic-aware AgentNet architecture for cross-layer optimization
Newly proposed framework
MoPS no independent evidence
purpose: Model partition and sharing for large models across agents
Developed in this paper
Three novel metrics no independent evidence
purpose: Evaluating SANet performance
Proposed for this work

pith-pipeline@v0.9.0 · 5622 in / 1392 out tokens · 36204 ms · 2026-05-16T19:18:38.048752+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We formulate the decentralized optimization of SANet as a multi-agent multi-objective problem... derive theoretical bounds... three-way tradeoff among optimization, generalization, and conflicting errors (Theorems 1-4, Table 1).
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Pareto stationary solution... min_γ ∑ γ_i ∇L_i = 0; dynamic-weighting γ update (11).

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

SANEmerg: An Emergent Communication Framework for Semantic-aware Agentic AI Networking
cs.AI 2026-05 unverdicted novelty 5.0

SANEmerg enables emergent communication among bounded-intelligence AI agents for semantic-aware task fulfillment in AgentNet systems via a bandwidth-adaptable importance filter and MDL-based complexity regularizer.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · cited by 1 Pith paper · 3 internal anchors

[1]

Sannet: A semantic-aware agentic ai networking framework for multi-agent cross-layer coordination,

Y. Xiao, H. Zhou, X. Li, Y. Gao, G. Shi, and P . Zhang, “Sannet: A semantic-aware agentic ai networking framework for multi-agent cross-layer coordination,” inProceedings of IEEE GLOBECOM, Taipei, Taiwan, December 2025

work page 2025
[2]

Enhancing immersion and presence in the meta- verse with over-the-air brain-computer interface,

N. Q. Hieu, D. T. Hoang, D. N. Nguyen, V .-D. Nguyen, Y. Xiao, and E. Dutkiewicz, “Enhancing immersion and presence in the meta- verse with over-the-air brain-computer interface,”IEEE Transac- tions on Wireless Communications, vol. 23, no. 12, pp. 18 532–18 548, December 2024

work page 2024
[3]

From semantic communication to semantic-aware networking: Model, architecture, and open problems,

G. Shi, Y. Xiao, Y. Li, and X. Xie, “From semantic communication to semantic-aware networking: Model, architecture, and open problems,”IEEE Communications Magazine, vol. 59, no. 8, pp. 44– 50, August 2021

work page 2021
[4]

Joint source–channel coding: Fundamentals and recent progress in practical designs,

D. Gunduzet al., “Joint source–channel coding: Fundamentals and recent progress in practical designs,”Proceedings of the IEEE, pp. 1–32, 2024

work page 2024
[5]

Illuminating the dark spaces of healthcare with ambient intelligence,

A. Haque, A. Milstein, and L. Fei-Fei, “Illuminating the dark spaces of healthcare with ambient intelligence,”Nature, vol. 585, no. 7824, pp. 193–202, September 2020

work page 2020
[6]

Towards agentic AI networking in 6G: A generative foundation model-as-agent approach,

Y. Xiao, G. Shi, and P . Zhang, “Towards agentic AI networking in 6G: A generative foundation model-as-agent approach,”IEEE Communications Magazine, vol. 63, no. 9, September 2025

work page 2025
[7]

Toward self- learning edge intelligence in 6G,

Y. Xiao, G. Shi, Y. Li, W. Saad, and H. V . Poor, “Toward self- learning edge intelligence in 6G,”IEEE Communications Magazine, vol. 58, no. 12, pp. 34–40, December 2020

work page 2020
[8]

6G network AI architecture for everyone-centric customized services,

Y. Yanget al., “6G network AI architecture for everyone-centric customized services,”IEEE Network, vol. 37, no. 5, September 2023

work page 2023
[9]

Position: Levels of AGI for operationalizing progress on the path to AGI,

M. R. Morriset al., “Position: Levels of AGI for operationalizing progress on the path to AGI,” inICML, Vienna, Austria, July 2024

work page 2024
[10]

Practices for governing agentic ai systems,

Y. Shavit, S. Agarwal, M. Brundage, S. Adler, C. O’Keefe, R. Camp- bell, T. Lee, P . Mishkin, T. Eloundou, A. Hickeyet al., “Practices for governing agentic ai systems,”Research Paper, OpenAI, 2023

work page 2023
[11]

Cross-layer optimization for statistical qos provision in C-RAN with finite-length coding,

C. Wu, H. Lu, Y. Chen, and L. Qin, “Cross-layer optimization for statistical qos provision in C-RAN with finite-length coding,”IEEE Transactions on Communications, vol. 72, no. 6, pp. 3393–3407, Jun. 2024

work page 2024
[12]

Cross- layer optimization: Joint user scheduling and beamforming design with qos support in joint transmission networks,

S. He, Z. An, J. Zhu, M. Zhang, Y. Huang, and Y. Zhang, “Cross- layer optimization: Joint user scheduling and beamforming design with qos support in joint transmission networks,”IEEE Transac- tions on Communications, vol. 71, no. 2, pp. 792–807, Feb. 2023

work page 2023
[13]

Proximal policy optimization for cross-layer joint design of miso beamforming and ris phase con- figuration,

L.-W. Chien and S.-M. Tseng, “Proximal policy optimization for cross-layer joint design of miso beamforming and ris phase con- figuration,”IEEE Open Journal of the Communications Society, pp. 1–1, Jul. 2025

work page 2025
[14]

On-demand optimization method for cross-layer topology in multi-task vleo and mega-leo heterogeneous satellite networks,

K. Han, M. Siew, B. Xu, S. Guo, W. Gong, T. Q. S. Quek, and Q. Ren, “On-demand optimization method for cross-layer topology in multi-task vleo and mega-leo heterogeneous satellite networks,” IEEE Transactions on Wireless Communications, pp. 1–1, Jun. 2025

work page 2025
[15]

AI/ML for beyond 5G systems: Concepts, technology enablers & solutions,

T. Taleb, C. Benzaid, R. A. Addad, and K. Samdanis, “AI/ML for beyond 5G systems: Concepts, technology enablers & solutions,” Computer Networks, vol. 237, p. 110044, 2023

work page 2023
[16]

A review of the current usage of ai/ml for radio access network (ran),

A. AlZailaa, J. Corona, R. Teixeira, H. R. Chi, M. Antunes, A. Rad- wan, and R. L. Aguiar, “A review of the current usage of ai/ml for radio access network (ran),”IEEE Access, 2025

work page 2025
[17]

SANSee: A physical-layer semantic-aware network- ing framework for distributed wireless sensing,

H. Zhuet al., “SANSee: A physical-layer semantic-aware network- ing framework for distributed wireless sensing,”IEEE Transactions on Mobile Computing, vol. 24, no. 3, pp. 1636–1653, March 2025

work page 2025
[18]

Optimizing through- put performance in distributed MIMO Wi-Fi networks using deep reinforcement learning,

N. Nurani Krishnan, E. Torkildson, N. B. Mandayam, D. Ray- chaudhuri, E.-H. Rantala, and K. Doppler, “Optimizing through- put performance in distributed MIMO Wi-Fi networks using deep reinforcement learning,”IEEE Transactions on Cognitive Communi- cations and Networking, vol. 6, no. 1, pp. 135–150, 2020

work page 2020
[19]

Toward decentral- ized task offloading and resource allocation in user-centric mec,

L. Qin, H. Lu, Y. Chen, B. Chong, and F. Wu, “Toward decentral- ized task offloading and resource allocation in user-centric mec,” IEEE Transactions on Mobile Computing, vol. 23, no. 12, pp. 11 807– 11 823, 2024

work page 2024
[20]

Towards an AI/ML-driven SMO framework in O-RAN: Scenar- ios, solutions, and challenges,

M. A. Habibi, B. Han, M. Saimler, I. L. Pav ´on, and H. D. Schotten, “Towards an AI/ML-driven SMO framework in O-RAN: Scenar- ios, solutions, and challenges,” inIEEE Future Networks World Forum, Dubai, UAE, October 2024, pp. 7–14. 17

work page 2024
[21]

5G evolution: A view on 5G cellular technology beyond 3GPP release 15,

A. Ghosh, A. Maeder, M. Baker, and D. Chandramouli, “5G evolution: A view on 5G cellular technology beyond 3GPP release 15,”IEEE Access, vol. 7, pp. 127 639–127 651, 2019

work page 2019
[22]

An overview of 5G advanced evolution in 3GPP release 18,

X. Lin, “An overview of 5G advanced evolution in 3GPP release 18,”IEEE Communications Standards Magazine, vol. 6, no. 3, pp. 77–83, 2022

work page 2022
[23]

The bridge toward 6G: 5G-advanced evolution in 3GPP release 19,

——, “The bridge toward 6G: 5G-advanced evolution in 3GPP release 19,”IEEE Communications Standards Magazine, vol. 9, no. 1, pp. 28–35, 2025

work page 2025
[24]

Requirements of semantic-aware networking for future networks,

G. Shi and Y. Xiao, “Requirements of semantic-aware networking for future networks,” ITU-T TR.Reqts-SAN, Nov. 2023

work page 2023
[25]

An introduction to semantic communication and semantic- aware networking standardization for 6G,

——, “An introduction to semantic communication and semantic- aware networking standardization for 6G,”GetMobile: Mobile Comp. and Comm., vol. 28, no. 3, p. 14–19, Oct. 2024

work page 2024
[26]

The roadmap to 6G: AI empowered wireless networks,

K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y.-J. A. Zhang, “The roadmap to 6G: AI empowered wireless networks,”IEEE Communications Magazine, vol. 57, no. 8, pp. 84–90, August 2019

work page 2019
[27]

AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges,

R. Sapkota, K. I. Roumeliotis, and M. Karkee, “AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges,” Information Fusion, vol. 126, p. 103599, 2026

work page 2026
[28]

Agent AI: Surveying the horizons of multimodal interaction,

Z. Duranteet al., “Agent AI: Surveying the horizons of multimodal interaction,”arXiv preprint arXiv:2401.03568, 2024

work page arXiv 2024
[29]

Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration,

J. Wang, H. Xu, H. Jia, X. Zhang, M. Yan, W. Shen, J. Zhang, F. Huang, and J. Sang, “Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration,” inProceedings of NeurIPS Conference, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37. Curran Associates, Inc., Van...

work page arXiv 2024
[30]

Rethinking the bounds of LLM reasoning: Are multi-agent discussions the key?

Q. Wang, Z. Wang, Y. Su, H. Tong, and Y. Song, “Rethinking the bounds of LLM reasoning: Are multi-agent discussions the key?” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L.-W. Ku, A. Martins, and V . Srikumar, Eds. Bangkok, Thailand: Association for Computational Linguistics, Aug. 2024...

work page 2024
[31]

Self-refine: iterative refinement with self-feedback,

A. Madaan, N. Tandon, P . Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, S. Gupta, B. P . Ma- jumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P . Clark, “Self-refine: iterative refinement with self-feedback,” inProceed- ings of NeurIPS Conference, New Orleans, LA, USA, December 2023

work page 2023
[32]

Rethinking Mixture-of-Agents: Is mixing different large language models beneficial?

W. Li, Y. Lin, M. Xia, and C. Jin, “Rethinking Mixture-of-Agents: Is mixing different large language models beneficial?”arXiv preprint arXiv:2502.00674, Feburary 2025

work page arXiv 2025
[33]

MoA is all you need: Building LLM research team using mixture of agents,

S. Chen, L. Zeng, A. Raghunathan, F. Huang, and T. C. Kim, “MoA is all you need: Building LLM research team using mixture of agents,”arXiv preprint arXiv:2409.07487, September 2024

work page arXiv 2024
[34]

Mixture-of-Agents Enhances Large Language Model Capabilities

J. Wang, J. Wang, B. Athiwaratkun, C. Zhang, and J. Zou, “Mixture-of-Agents enhances large language model capabilities,” arXiv preprint arXiv:2406.04692, June 2024

work page internal anchor Pith review arXiv 2024
[35]

Small Language Models are the Future of Agentic AI

P . Belcak, G. Heinrich, S. Diao, Y. Fu, X. Dong, S. Muralidharan, Y. C. Lin, and P . Molchanov, “Small language models are the future of agentic ai,”arXiv preprint arXiv:2506.02153, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[36]

Agentic AI: Au- tonomous intelligence for complex goals—a comprehensive sur- vey,

D. B. Acharya, K. Kuppan, and B. Divya, “Agentic AI: Au- tonomous intelligence for complex goals—a comprehensive sur- vey,”IEEE Access, vol. 13, January 2025

work page 2025
[37]

Mitigating gradient bias in multi-objective learning: A provably convergent approach,

H. D. Fernando, H. Shen, M. Liu, S. Chaudhury, K. Murugesan, and T. Chen, “Mitigating gradient bias in multi-objective learning: A provably convergent approach,” inProceedings of ICLR Confer- ence, Kigali, Rwanda, May 2023

work page 2023
[38]

Three-way trade-off in multi-objective learn- ing: Optimization, generalization and conflict-avoidance,

L. Chenet al., “Three-way trade-off in multi-objective learn- ing: Optimization, generalization and conflict-avoidance,”NIPS, vol. 36, pp. 70 045–70 093, New Orleans, LA, December 2023

work page 2023
[39]

Rethinking bias-variance trade-off for generalization of neural networks,

Z. Yang, Y. Yu, C. You, J. Steinhardt, and Y. Ma, “Rethinking bias-variance trade-off for generalization of neural networks,” inProceedings of ICML Conference, Virtual, July 2020, pp. 10 767– 10 777

work page 2020
[40]

Your own mobile network,

srsRAN, “Your own mobile network,”https://www.srslte.com/, srsRAN, 2019

work page 2019
[41]

Qwen Technical Report

J. Bai, S. Bai, Y. Chuet al., “Qwen technical report,”arXiv preprint arXiv:2309.16609, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[42]

OpenManus: An open-source framework for building general AI agents,

X. Liang, J. Xiang, Z. Yuet al., “OpenManus: An open-source framework for building general AI agents,” 2025

work page 2025
[43]

Stein,Inadmissibility of the usual estimator for the mean of a multivariate normal distribution

C. Stein,Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. University of California Press, 1956, vol. 3

work page 1956
[44]

Time series forecasting using LSTM and ARIMA,

K. Albeladi, B. Zafar, and A. Mueen, “Time series forecasting using LSTM and ARIMA,”International Journal of Advanced Computer Science and Applications, vol. 14, no. 1, pp. 313–320, 2023

work page 2023
[45]

N- beats: Neural basis expansion analysis for interpretable time series forecasting,

B. N. Oreshkin, D. Carpov, N. Chapados, and Y. Bengio, “N- beats: Neural basis expansion analysis for interpretable time series forecasting,” inProceedings of ICLR Conference, New Orleans, LA, May 2019

work page 2019
[46]

A transformer-based framework for multivariate time series representation learning,

G. Zerveas, S. Jayaraman, D. Patel, A. Bhamidipaty, and C. Eick- hoff, “A transformer-based framework for multivariate time series representation learning,” inProceedings of SIGKDD Conference, Virtual, August 2021, pp. 2114–2124

work page 2021
[47]

Informer: Beyond efficient transformer for long se- quence time-series forecasting,

H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long se- quence time-series forecasting,” inProceedings of AAAI Conference, vol. 35, no. 12, Virtual, February 2021, pp. 11 106–11 115

work page 2021
[48]

Stability and generalization of stochastic optimization with nonconvex and nonsmooth problems,

Y. Lei, “Stability and generalization of stochastic optimization with nonconvex and nonsmooth problems,” inAnnual Conference on Learning Theory, Bangalore, India, July 2023, pp. 191–227. Yong Xiao(Senior Member, IEEE) received his B.S. degree in electrical engineering from China University of Geosciences, Wuhan, China in 2002, M.Sc. degree in telecommunic...

work page 2023
[49]

Ping Zhang(Fellow, IEEE) is a professor in the School of Information and Communication Engineering at the Beijing University of Posts and Telecommunications

His research interests include Artificial Intelligence, Semantic Communications, and Human-Computer Interaction. Ping Zhang(Fellow, IEEE) is a professor in the School of Information and Communication Engineering at the Beijing University of Posts and Telecommunications. He is a member of the National Academy of Engineering of China. He is currently the di...

work page 2020