Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Agentic AI can sharpen semantic communications: an agent-built knowledge base for joint source-channel coding lifts image reconstruction PSNR by roughly 9% over a leading JSCC baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 13:40 UTC pith:DKFK24EM

load-bearing objection The framework is a useful synthesis, but the 9% agentic-KB gain in the case study is confounded by a backbone change that no ablation controls for. the 3 major comments →

arxiv 2512.23294 v2 pith:DKFK24EM submitted 2025-12-29 eess.SY cs.SY

Agentic AI-Enhanced Semantic Communications: Foundations, Architecture, and Applications

Haixiao Gao , Mengying Sun , Ruichen Zhang , Yanhan Wang , Xiaodong Xu , Nan Ma , Dusit Niyato , Ping Zhang This is my paper
classification eess.SY cs.SY
keywords semantic communicationsagentic AIjoint source-channel codingknowledge baselarge language modelsreinforcement learning6G networksimage transmission
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that giving semantic communications an agentic layer—agents that perceive, remember, reason, and act—turns a fixed codec into an adaptive system that protects the semantics a task actually needs. It proposes a three-layer architecture spanning the application, semantic, and cloud-edge layers, forming a closed intent-to-evaluation loop. To test the idea, it builds AKB-JSCC, a joint source-channel coding scheme whose source knowledge base is constructed by a vision-language agent and whose channel knowledge base is constructed by a reinforcement-learning agent. On the MSCOCO dataset under AWGN channels, AKB-JSCC reports roughly a 9% PSNR gain over NTSCC at CBR=0.02, with the source-KB-only ablation still beating the baseline. If the attribution holds, agentic knowledge bases offer a concrete route to robust low-bandwidth image transmission.

Core claim

The paper's central discovery is that an "agentic knowledge base" can replace codebook-style priors in deep joint source-channel coding. The LLM/LVM source KB generates a text description of the input image, embeds it, retrieves the closest cross-modal prior from a large embedding library, and feeds that prior to both the encoder and the decoder. The RL channel KB observes the entropy map, SNR map, and previous action, and outputs a rate-control action, trained with PPO. Together they implement variable-length, semantics-aware coding. The reported result is that at CBR=0.02 under AWGN, AKB-JSCC achieves about 9% higher PSNR than NTSCC and maintains higher reconstruction quality across SNR co

What carries the argument

The load-bearing object is the agentic knowledge base (KB). The source KB uses a vision-language agent to turn the input image into a text description, tokenizes and embeds that description, then performs nearest-neighbor retrieval in a multimodal embedding library; the retrieved vector conditions the JSCC encoder and decoder. The channel KB is an actor-critic RL agent whose state concatenates the entropy map, SNR map, and previous action, producing an action for joint rate control; the reward combines reconstruction quality and bandwidth overhead. The work of this machinery is to transfer "memory" (cross-modal source priors) and "reasoning/action" (channel adaptation) directly into the code

Load-bearing premise

The load-bearing premise is that the 9% PSNR gap comes from the agentic knowledge bases themselves, not from the stronger cross-modal Swin Transformer backbone used in AKB-JSCC but not in the NTSCC baseline.

What would settle it

Run AKB-JSCC with both KB components disabled but the same backbone, and compare PSNR against NTSCC at CBR=0.02 over AWGN; if the gap nearly vanishes, the paper's central attribution fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, separating source and channel knowledge into agent-built KBs is a viable design pattern for deep JSCC systems.
  • The source KB alone, without the channel KB, already lifts quality over NTSCC, suggesting cross-modal priors are a strong independent contributor.
  • The RL channel KB enables bitrate selection driven by semantic importance and channel state, offering a path to variable-length coding that avoids the cliff effect of traditional separated schemes.
  • The three-layer closed-loop framework can be transferred to collaborative vehicle perception, multi-robot rescue, and intellicise network operations.
  • Standardized interfaces and unified KPIs between agents and semantic components would be needed before such systems become portable and deployable at scale.

Where Pith is reading between the lines

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

  • A same-backbone control is the natural next experiment: disabling both KB components while keeping AKB-JSCC's Swin Transformer encoder would separate the agentic-KB contribution from the backbone upgrade.
  • If the source KB is the active ingredient, a distilled retriever or cached embedding lookup could deliver most of the gain at substantially lower inference cost.
  • The two-KB pattern could be tested on video, point clouds, or joint perception tasks where multimodal priors matter more than in single-image reconstruction.
  • The intent-to-evaluation feedback loop suggests a protocol-level application where QoS scores trigger semantic model switching, extending the model-division multiple access idea already referenced in the paper.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper is a survey/position paper on agentic AI for semantic communications. It proposes a three-layer architecture (application, semantic, and cloud-edge collaboration) with a closed intent-encode-transmit-decode-act-evaluate loop, taxonomizes prior work by agent type (embedded, LLM/LVM, RL), sketches three application scenarios, and presents AKB-JSCC, a JSCC codec whose source KB is built with an LLM/LVM agent and whose channel KB is built with a PPO RL agent. Experiments on MSCOCO 2014 compare AKB-JSCC against NTSCC, two separated source/channel coding baselines, and an ablation without the channel KB. The authors claim about a 9% PSNR gain over NTSCC at CBR=0.02, interpreting this as evidence for the benefit of an agentic KB for SemCom.

Significance. If the case-study result holds, the paper would serve as a useful consolidation: the taxonomy by agent type is practical, the three-layer framework is sufficiently general to organize future research, and the AKB-JSCC case demonstrates a concrete integration of LLM/LVM-built semantic priors with RL-based rate control. Strengths include the use of a public dataset, comparison against standard JSCC and separated baselines, an ablation, and clearly stated future directions. The central quantitative claim, however, is not cleanly established because AKB-JSCC differs from NTSCC in both the agentic KB and the underlying codec backbone; the reported gain cannot be attributed to the agentic KB without an additional control. This is an attribution problem that is addressable with further experiments, not an internal inconsistency.

major comments (3)
  1. [§IV-B.1 and §IV-C, Fig. 4] AKB-JSCC and NTSCC differ in two dimensions: the agentic KB and the codec backbone. Section IV-B.1 states that the JSCC codec 'adopts the cross-modal Swin Transformer in [14] as the backbone,' while NTSCC [15] uses its original nonlinear-transform backbone. The only reported ablation, AKB-JSCC (w/o CKB), removes the RL-based channel KB but retains the source KB and the [14] backbone, so it cannot isolate the contribution of the agentic KB. If the cross-modal Swin Transformer alone accounts for most of the performance gap, the headline 'Agentic KB brings a 9% performance gain' (Fig. 4) is not established. Please add a same-backbone control without the source KB (and ideally a version without both KBs), and report the incremental PSNR for each component.
  2. [§IV-C, Fig. 4] The caption states 'All models are trained once at SNR=10 dB'; no error bars, number of seeds, or statistical significance are reported. At CBR=0.019 the displayed PSNR values are 27.02 (AKB-JSCC) vs 25.62 (NTSCC), a roughly 5.5% relative difference, and the quoted 'about 9%' at CBR=0.02 does not clearly correspond to the displayed operating points. Provide a table with mean±std over multiple independent training runs and the exact PSNR at CBR=0.02, and state the number of seeds. This is necessary to support the quantitative headline claim.
  3. [§IV-C] The case study omits several implementation details needed for reproducibility: the size and construction of the multimodal embedding KB, the retrieval top-k and distance metric used beyond the argmin formula, the prompt template for LLaVA-NeXT-7B, PPO hyperparameters, reward weights in the RL objective, channel model parameters, and training epochs. Without these details, an independent group cannot replicate the comparison or assess whether the reported gains are robust to implementation choices.
minor comments (4)
  1. [§III-B] Typo: 'UA Vs' should be 'UAVs'.
  2. [§II-B.6] The phrase 'cross-modal feature semantic feature alignment' repeats 'feature'; please rephrase.
  3. [Fig. 4 caption] The caption says CBR is fixed at 0.03 for all schemes except JPEG+LDPC, which uses 0.035 'to ensure a fair comparison.' Please explain why JPEG2000+LDPC can use 0.03 but JPEG+LDPC cannot, and how the fairness criterion is defined.
  4. [Introduction] The claim of being 'the first systematic study' is difficult to verify and may invite unnecessary criticism; consider softening to 'one of the first' or adding a comparison with the most recent surveys.

Circularity Check

0 steps flagged

No significant circularity: the AKB-JSCC case study is an empirical comparison against external baselines, not a definitional reduction.

full rationale

The central quantitative claim (Fig. 4: 'At CBR=0.02, AKB-JSCC delivers about a 9% PSNR gain over NTSCC, confirming the benefit of an agentic KB for SemCom') is an experimental result comparing AKB-JSCC against a separately published NTSCC baseline [15] and traditional JPEG2000+LDPC / JPEG+LDPC schemes; the gain is not derived from the model equations or from a fitted parameter renamed as a prediction. The paper relies on several prior works by the same research group ([2], [4], [12], [13], [14], [15]), but none is invoked as an unverified uniqueness theorem or as a premise that logically forces the reported PSNR difference. The source-KB retrieval rule 'argmin_{c in C} ||r-c||_2^2' is a design choice, not a definition of 'benefit of agentic KB'. The internal ablation AKB-JSCC (w/o CKB) is a genuine control, even though it does not hold the backbone fixed relative to NTSCC; the absence of a same-backbone no-KB control is an attribution/experimental-design limitation, which I flag as a correctness risk rather than circularity. No equation in the paper equates an input to an output by construction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central claims rest on several unspecified or under-specified choices. Free parameters: RL reward weights, the entropy-to-rate preset map, and the training SNR directly shape the reported quality-rate curves and are not specified. Axioms: the paper assumes the SemCom benefits cited in [1]-[4], that PSNR on MSCOCO represents semantic reconstruction quality, that the [14] backbone is neutral in the comparison against NTSCC [15], and that retrieved LLM/LVM embeddings monotonically enhance JSCC. Invented entity: the agentic KB is the paper's central new construction and has no external falsifiable handle — its only support is the in-paper experiment, which is confounded and not reproducible from the text alone.

free parameters (3)
  • Reward weights in the RL channel-KB objective
    The reward r_t is defined as a weighted combination of reconstruction quality and channel bandwidth overhead (Section IV-B.3); the weights are never specified, and they directly shape the adaptive rate policy whose output is later reported as a quality gain.
  • Predefined discrete rate set for the rate preset map
    The entropy map is 'mapped to a predefined discrete rate set' (Section IV-B.1, Fig. 3); the rate levels are hand-chosen and unspecified, and they determine the semantics-aware rate allocation behavior that is presented as a result.
  • Training SNR and CBR operating points = 10 dB
    All models are trained once at SNR=10 dB and tested across SNRs (Section IV-C, Fig. 4 caption); this training-SNR choice affects the robustness curves, and the JPEG+LDPC baseline is assigned CBR=0.035 vs 0.03 for the other schemes to 'ensure a fair comparison'.
axioms (5)
  • domain assumption SemCom provides the robustness and efficiency benefits its foundations cite (low-SNR robustness, cliff-effect mitigation, bandwidth reduction).
    The motivation (Section I) and the framework rest on accepting [1]-[4]'s characterization of SemCom; the paper gives no independent verification.
  • domain assumption PSNR on MSCOCO 2014 is an adequate proxy for 'information reconstruction quality' and task utility.
    Section IV-C reports PSNR and CBR as the sole quantitative metrics; PSNR is a pixel-fidelity metric that may not capture the semantic/task-level fidelity SemCom claims to optimize.
  • ad hoc to paper The cross-modal Swin Transformer backbone [14] is a neutral component whose contribution relative to the NTSCC backbone [15] is negligible in the comparison.
    Section IV-B.1 adopts [14] as the backbone and Section IV-C compares against NTSCC without a matched-backbone control; the ablation logic presumes remaining differences are due to the KB agents.
  • domain assumption LLM/LVM-generated caption embeddings retrieved from the knowledge base monotonically improve JSCC encoding and decoding.
    Section IV-B.2 assumes the retrieved cross-modal embedding 'enhances' the encoder and decoder; no analysis is given for cases where retrieval introduces noise or semantic misalignment.
  • standard math Standard PPO training on the AKB-JSCC environment converges to a policy that generalizes over the tested SNR/CBR range.
    Section IV-B.3 relies on standard PPO machinery; convergence and generalization are assumed, not demonstrated with learning curves or seed studies.
invented entities (1)
  • Agentic knowledge base (AKB): source KB built by LLM/LVM agents + channel KB built by RL agents no independent evidence
    purpose: Central architectural construction of the case study: the source KB supplies cross-modal semantic priors to the JSCC codec; the channel KB performs entropy/SNR-driven adaptive rate control.
    The dual-KB agentic design is introduced in Section IV-B; its only supporting evidence is the in-paper experiment (Fig. 4), which lacks a matched-backbone control, error bars, and released artifacts. No falsifiable prediction outside the paper is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 12020 in / 18191 out tokens · 163289 ms · 2026-08-03T13:40:17.298920+00:00 · methodology

0 comments
read the original abstract

Semantic communications (SemCom), as one of the key technologies for 6G, is shifting networks from bit transmission to semantic information exchange. On this basis, introducing agentic artificial intelligence (AI) with perception, memory, reasoning, and action capabilities provides a practicable path to intelligent communications. This paper provides a systematic exposition of how agentic AI empowers SemCom from the perspectives of research foundations, system architecture, and application scenarios. We first provide a comprehensive review of existing studies by agent types, covering embedded agents, large language model (LLM)/large vision model (LVM) agents, and reinforcement learning (RL) agents. Additionally, we propose a unified agentic AI-enhanced SemCom framework covering the application layer, the semantic layer, and the cloud-edge collaboration layer, forming a closed loop from intent to encoding to transmission to decoding to action to evaluation. We also present several typical scenarios, including multi-vehicle collaborative perception, multi-robot cooperative rescue, and agentic operations for intellicise (intelligent and concise) networks. Furthermore, we introduce an agentic knowledge base (KB)-based joint source-channel coding case study, AKB-JSCC, where the source KB and channel KB are built by LLM/LVM agents and RL agents, respectively. Experimental results show that AKB-JSCC achieves higher information reconstruction quality under different channel conditions. Finally, we discuss future evolution and research directions, providing a reference for portable, verifiable, and controllable research and deployment of agentic SemCom.

Figures

Figures reproduced from arXiv: 2512.23294 by Dusit Niyato, Haixiao Gao, Mengying Sun, Nan Ma, Ping Zhang, Ruichen Zhang, Xiaodong Xu, Yanhan Wang.

Figure 1
Figure 1. Figure 1: Illustration of the agentic AI-enhanced SemCom architecture. It comprises multiple layers, where (A) denotes the application layer, (B) denotes the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of typical application scenarios of agentic AI-enhanced SemCom, covering multi-vehicle collaborative perception, multi-robot cooperative [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the AKB-JSCC framework. The three parts denote (1) the overall framework of AKB-JSCC, (2) the LLM/LVM agent-based source [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Simulation results under the AWGN channel. Left: visual reconstructions at varying CBR and SNR, together with entropy map and rate preset map. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. Secure Intellicise Wireless Network: Agentic AI for Coverless Semantic Steganography Communication

    cs.CR 2026-01 unverdicted novelty 6.0

    Agentic AI enables coverless semantic steganography without private keys or cover images, delivering higher capacity and security than prior schemes in semantic communication.

  2. Generative Channel Knowledge Base With Environmental Information for Joint Source-Channel Coding in Semantic Communications

    cs.IT 2026-04 unverdicted novelty 5.0

    A Transformer-based generative model builds an environment-aware channel knowledge base that is injected into JSCC encoders and decoders, achieving 10^{-3} level channel estimation error and outperforming benchmarks i...

Reference graph

Works this paper leans on

15 extracted references · 1 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Semantic communications for future internet: Fundamentals, applications, and challenges,

    W. Yang, H. Du, Z. Q. Liew, W. Y . B. Lim, Z. Xiong, D. Niyato, X. Chi, X. Shen, and C. Miao, “Semantic communications for future internet: Fundamentals, applications, and challenges,”IEEE Commun. Surveys Tuts., vol. 25, no. 1, pp. 213–250, 2022

    2022

  2. [2]

    ComAI: The convergence of communication and artificial intelligence,

    P. Zhang, K. Niu, X. Wang, Y . Liu, Z. Liang, C. Dong, J. Dai, X. Xu, W. Xu, Z. Zhanget al., “ComAI: The convergence of communication and artificial intelligence,”IEEE Commun. Surveys Tuts., 2025

    2025

  3. [3]

    Toward edge general intelligence with agentic AI and agentification: Concepts, technologies, and future directions,

    R. Zhang, G. Liu, Y . Liu, C. Zhao, J. Wang, Y . Xu, D. Niyato, J. Kang, Y . Li, S. Maoet al., “Toward edge general intelligence with agentic AI and agentification: Concepts, technologies, and future directions,”arXiv preprint arXiv:2508.18725, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Intellicise wireless networks from semantic communications: A survey, research issues, and challenges,

    P. Zhang, W. Xu, Y . Liu, X. Qin, K. Niu, S. Cui, G. Shi, Z. Qin, X. Xu, F. Wanget al., “Intellicise wireless networks from semantic communications: A survey, research issues, and challenges,”IEEE Commun. Surveys Tuts., vol. 27, no. 3, pp. 2051–2084, 2025

    2051

  5. [5]

    Agent-driven generative semantic communication with cross-modality and prediction,

    W. Yang, Z. Xiong, Y . Yuan, W. Jiang, T. Q. Quek, and M. Debbah, “Agent-driven generative semantic communication with cross-modality and prediction,”IEEE Trans. Wireless Commun., vol. 24, no. 3, pp. 2233–2248, 2025

    2025

  6. [6]

    On the uses of large language models to design end-to-end learning semantic communication,

    Y . Wang, Z. Sun, J. Fan, and H. Ma, “On the uses of large language models to design end-to-end learning semantic communication,” in2024 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2024, pp. 1–6

    2024

  7. [7]

    LaMoSC: Large language model-driven semantic communication system for visual trans- mission,

    Y . Zhao, Y . Yue, S. Hou, B. Cheng, and Y . Huang, “LaMoSC: Large language model-driven semantic communication system for visual trans- mission,”IEEE Trans. Cogn. Commun. Netw., vol. 10, no. 6, pp. 2005– 2018, 2024

    2005

  8. [8]

    Decision-Making Large Language Model for Wireless Communication: A Comprehensive Sur- vey on Key Techniques,

    N. Yang, M. Fan, W. Wang, and H. Zhang, “Decision-Making Large Language Model for Wireless Communication: A Comprehensive Sur- vey on Key Techniques,”IEEE Commun. Surveys Tuts., 2025

    2025

  9. [9]

    Adaptive resource allocation for semantic communication networks,

    L. Wang, W. Wu, F. Zhou, Z. Yang, Z. Qin, and Q. Wu, “Adaptive resource allocation for semantic communication networks,”IEEE Trans. Commun., vol. 72, no. 11, pp. 6900–6916, 2024

    2024

  10. [10]

    Task-oriented semantic communi- cation in large multimodal models-based vehicle networks,

    B. Du, H. Du, D. Niyato, and R. Li, “Task-oriented semantic communi- cation in large multimodal models-based vehicle networks,”IEEE Trans. Mobile Comput., vol. 24, no. 10, pp. 9822–9836, 2025

    2025

  11. [11]

    Resource allocation driven by large models in future semantic-aware networks,

    H. Zhang, J. Ni, Z. Wu, X. Liu, and V . Leung, “Resource allocation driven by large models in future semantic-aware networks,”IEEE Wireless Commun., vol. 32, no. 4, pp. 116–122, 2025

    2025

  12. [12]

    Model division multiple access for semantic communications,

    P. Zhang, X. Xu, C. Dong, K. Niu, H. Liang, Z. Liang, X. Qin, M. Sun, H. Chen, N. Maet al., “Model division multiple access for semantic communications,”Frontiers of Information Technology & Electronic Engineering, vol. 24, no. 6, pp. 801–812, 2023

    2023

  13. [13]

    Codebook- enabled generative end-to-end semantic communication powered by transformer,

    P. Ye, Y . Sun, S. Yao, H. Chen, X. Xu, and S. Cui, “Codebook- enabled generative end-to-end semantic communication powered by transformer,” inIEEE INFOCOM 2024-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), 2024, pp. 1–6

    2024

  14. [14]

    Adaptive cross-modal super-resolution semantic communication for mobile AI- generated panoramic video,

    H. Gao, M. Sun, X. Xu, X. Cheng, S. Han, and P. Zhang, “Adaptive cross-modal super-resolution semantic communication for mobile AI- generated panoramic video,”IEEE Trans. Cogn. Commun. Netw., 2025

    2025

  15. [15]

    Nonlinear transform source-channel coding for semantic communications,

    J. Dai, S. Wang, K. Tan, Z. Si, X. Qin, K. Niu, and P. Zhang, “Nonlinear transform source-channel coding for semantic communications,”IEEE J. Sel. Areas Commun., vol. 40, no. 8, pp. 2300–2316, 2022

    2022