Model Context Protocol-based Internet of Experts For Wireless Environment-aware LLM Agents

Hongyang Du; Zongxi Liu

REVIEW 4 major objections 5 minor 1 cited by

Model Context Protocol-based Internet of Experts For Wireless Environment-aware LLM Agents

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that giving LLMs a pool of MCP-accessible expert classifiers lifts wireless attribute classification accuracy from 45%–59% to over 95% without retraining the language model.

desk verdict A clean, modest system paper whose synthetic-only evaluation is the main gap; worth sending to peer review with demands for over-the-air data and an ablation. read the letter →

arxiv 2505.01834 v1 pith:YY5YMN5D submitted 2025-05-03 cs.NI

Zongxi Liu , Hongyang Du This is my paper · ORCID

classification cs.NI

keywords ModelContextProtocolInternetofExpertswirelessenvironmentawarenessLLMagentschannelimpulseresponseclassificationtool-augmentedreasoningline-of-sightdetectionsyntheticsimulation

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

The paper proposes that the right way to give large language models awareness of wireless radio environments is not to retrain them or prompt them with raw numbers, but to let them call a pool of small, task-specific classifiers through the Model Context Protocol at inference time. Each classifier is a lightweight network trained on channel impulse responses to detect one physical attribute, such as line-of-sight propagation, high Doppler shift, or Rayleigh/Rician fading. At query time the LLM decides which experts are relevant, the MCP runtime fetches their confidence scores, and the model answers with those scores in its context. On a 1000-sample synthetic test set, the authors report that this arrangement raises average classification accuracy from 45%–59% for standalone LLMs to over 95% for all eight evaluated models, including weaker ones. The claim matters because it points to a modular way to give general-purpose LLMs reliable, physical perception without per-deployment fine-tuning.

What carries the argument

The load-bearing object is the MCP-based Internet of Experts: MCP provides a uniform JSON-RPC interface that lets an LLM discover, call, and receive structured replies from external tools, and the IoX is the pool of lightweight per-attribute MLP classifiers, each trained with binary cross-entropy on scene-specific positive and negative channel samples. The pipeline has six stages: offline expert registration, LLM-driven expert planning, JSON-RPC invocation, standardized responses with confidence scores, context augmentation, and final reasoning. This machinery decouples physical-layer interpretation from high-level reasoning, so adding a new environment condition means training and registering one new classifier without retouching the LLM or the other experts.

What would settle it

Run the same MCP-IoX pipeline on over-the-air channel impulse response measurements, with ground-truth scene labels from a channel sounder or ray-tracing, and compare the expert and end-to-end accuracy to the synthetic >95% result; a large drop would show the synthetic experts do not transfer.

Watch

Extended reading notes

Core claim

The central claim is that structured wireless perception, not better language reasoning, is what LLM wireless agents lack. The authors build an Internet of Experts in which every wireless attribute is a separate binary classification problem solved by a compact MLP trained on the magnitude of complex channel impulse responses, and they connect the expert pool to the LLM through MCP's JSON-RPC tool layer. The LLM's job is reduced to planning which experts to call and interpreting their outputs; it never sees the raw channel vector as a reasoning target and never changes its weights. The result is that the agent's accuracy is bounded by the experts' own performance rather than by the LLM's ability to read numbers: the reported end-to-end accuracy exceeds 95% for all eight models, and weaker models match or beat stronger standalone models when the experts are available.

Load-bearing premise

The load-bearing assumption is that the synthetic channel simulator used to train the experts and label the 1000-sample test set captures real wireless environments closely enough; if real over-the-air channels differ in Doppler, multipath, or hardware effects, the reported >95% accuracy may not carry over.

Editorial extensions

If this is right

Standalone LLM accuracy on wireless attribute prediction is bounded at 45%–59%; adding MCP expert querying raises every tested model above 95%, so the bottleneck is perception, not reasoning.
A frozen, general-purpose LLM can acquire wireless-environment awareness without parameter updates, avoiding costly fine-tuning and preserving language generalization.
The expert pool is modular and extensible: a new attribute only requires new positive samples and a newly trained lightweight classifier, with no retraining of the rest of the system.
Weaker LLMs paired with expert outputs can match or surpass stronger models operating alone, so the architecture shifts the performance ceiling to the expert classifiers.
Because experts output interpretable confidence scores rather than free text, the agent's wireless reasoning is grounded in verifiable physical states, which should reduce hallucinated channel descriptions.

Reading between the lines

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

Real-world transfer is the open risk: the experts are trained on synthetic traces, and the paper lists over-the-air evaluation as future work; a drop in accuracy there would not contradict the synthetic result but would bound its practical value.
Beyond the tested wireless attributes, the same MCP-plus-expert-pool recipe likely transfers to other domains where LLMs must reason over raw high-dimensional signals, such as power-system telemetry, speech features, or sensor arrays.
The reported standalone baseline may understate what LLMs could do with better prompting or multimodal inputs; the headline gain therefore measures raw-number reasoning difficulty as much as MCP's contribution.
A testable next step is cost-aware expert selection: the paper notes latency and cost profiling as future work, and optimal selection would trade the number of expert calls against accuracy and budget.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

This is a straightforward, well-scoped system study. The contribution is not a new learning method or theory; it's a practical orchestration pattern: MCP-mediated access to lightweight wireless attribute classifiers lifts LLM accuracy from roughly 45–59% to over 95% on a synthetic test set. That result is believable and consistent.

What's actually new is the combination of MCP with a pool of independent expert classifiers for wireless physical-layer attributes. The architecture is clearly described, and the comparison across eight LLMs is a nice touch. The experts are simple MLPs, which is fine—the point is that the LLM can call them at inference time without retraining. The paper's internal logic holds up: if the experts are accurate, the LLM wrapper should be accurate, and the experiments confirm it.

The soft spots are real but not fatal. The evaluation is entirely synthetic; the channel simulator is undisclosed, there are no error bars, no code release, and no baseline that uses the expert outputs without the LLM. So we cannot tell whether the test distribution resembles the training distribution, or whether the >95% number survives real over-the-air channels. The authors themselves list over-the-air evaluation as future work, which is honest but confirms the external validity gap. Also, the input representation is just the magnitude of the channel impulse response—real baseband features (phase, complex taps) may be harder to classify. These are limitations, but they don't invalidate the claimed result as scoped to synthetic channels.

The citation pattern looks fine; no red flags. The paper is internally consistent and the authors are upfront about what is not tested.

Bottom line: a serious referee should spend time on this. It's a worthwhile proof-of-concept for tool-augmented LLMs in wireless, and the weaknesses are addressable. I'd ask for over-the-air measurements or a detailed simulator description, error bars, and a non-LLM baseline before publication. But the idea deserves to go through review rather than be desk rejected.

Referee Report

4 major / 5 minor

Summary. The paper proposes an MCP-based Internet of Experts (IoX) framework that equips LLM agents with wireless environment awareness. Lightweight MLP experts are trained separately for each wireless attribute (LoS, high Doppler, Rayleigh fading, Rician fading with K=10) and exposed to the LLM as MCP tools. At inference, the LLM selects and queries the relevant experts and receives structured confidence scores, which are injected into its context for final decision-making. Experiments compare standalone LLMs with MCP-augmented agents on 1000 synthetic channel observations, reporting 45%-59% accuracy without MCP versus over 95% with MCP across eight LLMs. The conclusion frames this as a significant gain from structured wireless perception, with over-the-air validation deferred to future work.

Significance. If the reported results are representative, the architecture is a clean and timely contribution: it decouples physical-layer perception from LLM reasoning, avoids fine-tuning, and is modular. The evaluation across eight LLMs is a strength, as is the explicit comparison against a no-tool baseline. The paper is internally consistent, and there is no equation-level circularity: expert confidences, not ground-truth labels, are fed to the LLM, and final labels are compared to an external ground truth. However, the central quantitative claim rests entirely on one undisclosed synthetic channel simulator, so the paper's broader claim of 'wireless environment-aware' agents remains unvalidated for real over-the-air channels. The proof-of-concept is valuable, but the external-validity gap is the decisive limitation.

major comments (4)

[Section IV-B, Table I] The end-to-end claim rests on a 1000-sample test set generated by a 'randomized channel simulation pipeline' whose parameters, noise model, antenna configuration, and train/test separation are not disclosed. Since the expert classifiers are trained on 'synthetic wireless traces' from the same family of processes, the reported >95% accuracy may reflect an easy, in-distribution test rather than general wireless perception. Please provide the full simulator configuration, specify how the test distribution differs from the training distribution, report error bars or per-run variance, and add an evaluation on measured channels or a standard benchmark (e.g., 3GPP channel models) before claiming general wireless-environment-aware performance.
[Section IV-A, Fig. 6] The expert learning results are presented only as training loss and test accuracy curves, without test-set sizes, final accuracy values, hyperparameters, confidence intervals, or train/test split details. The paper later states that end-to-end accuracy 'approaches the classification accuracy ceiling determined by the experts themselves,' so this expert ceiling must be quantified. Without these numbers, the reader cannot judge how robust the 95% end-to-end ceiling is or whether the binary expert classifiers are well calibrated for the confidence scores the LLM receives.
[Section IV-B, end-to-end evaluation] The comparison is LLM-with-MCP versus LLM-without-MCP, but there is no no-LLM rule-based baseline that thresholds the expert confidence scores (e.g., predict the attribute as present when the expert outputs p > 0.5). Such a baseline is necessary to isolate the marginal contribution of the LLM's planning and reasoning from the contribution of the expert classifiers themselves. If a simple thresholding baseline also achieves >95%, the current experiment does not support the claim that the LLM agent 'interprets' or 'reasons' over expert outputs.
[Section III-B and Section IV-B] Tool-selection and invocation errors are mentioned qualitatively ('minor errors from MCP call formatting inconsistencies or borderline expert predictions') but are never measured. The framework's value depends on the LLM choosing the right experts and correctly reading their outputs. Please report per-LLM tool-call success rates, the number of expert invocations per query, and accuracy conditioned on successful MCP calls. This is needed to distinguish failures of the MCP orchestrator from failures of the underlying experts.

minor comments (5)

[Fig. 6 caption] The caption says 'three representative expert classifiers,' but the figure contains four subplots; please correct the count.
[Abstract and Table I] The abstract's '40%-50% improvements' mixes percentage points and relative improvements; for example, 46.7% to 95.5% is a 48.8 percentage-point increase but a roughly 104% relative increase. Please state the metric unambiguously.
[Section IV-A] The experts are trained for 4000 epochs with no mention of early stopping or regularization; please specify whether overfitting was monitored and report the final hyperparameters.
[Section IV-B] The 1000-sample test set is generated with no seeds or reproducibility details; please provide the code or data release to allow independent replication.
[Section I] The term 'Internet of Experts' is used without a definition or reference; please clarify whether this is a new term or an existing concept.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracy gain is an end-to-end measurement with external ground truth, not a quantity forced by the expert fits.

full rationale

The paper's central quantitative claim (Table I) is an empirical end-to-end measurement, not a derived or predicted quantity. Expert classifiers are trained on synthetic wireless traces (Section IV-A), and the 1000-sample test set is built from a randomized channel simulation pipeline (Section IV-B); ground-truth labels are simulator-defined scene attributes, not the experts' own outputs. The LLM receives only expert confidence scores p(sm|h), must select and call the MCP tools, parse the structured JSON responses, and map confidences to the requested binary attributes, so the resulting >95% accuracy is not equal by construction to the expert training targets. No equation in the paper defines the measured accuracy in terms of the fitted parameters; Eq. (1) and Eq. (2) are standard evaluation objectives, and Eq. (5) merely describes the expert output format. The single self-citation ([7], Du et al.) is background motivation, not a load-bearing argument, and no uniqueness theorem or ansatz is imported from prior work by the same authors. The main weakness is external validity: both expert training and the test set come from the same style of synthetic simulator, and the Conclusion explicitly lists 'evaluate the framework on over-the-air channel measurements' as future work, but that is a generalization limitation, not circularity. The result is self-contained as a simulation benchmark and does not reduce to its own inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the reliability of the trained experts and the synthetic data pipeline. No new physical entities are introduced; the 'Internet of Experts' is an architectural pattern, not a new mechanistic entity. The main unstated premises are the sufficiency of the magnitude-only input and the representativeness of the simulator.

free parameters (2)

Expert MLP weights (W1,b1,W2,b2,W3,b3 in Eq. 3) = Not reported
The trained parameters of each expert are not disclosed. The reported end-to-end accuracy depends directly on these fitted weights, so the central quantitative claim is not independently checkable from the paper.
Synthetic channel generation parameters (CIR length, Doppler threshold, Rician K-factor, mixture ratios) = Not reported (Rician K=10 is named)
The paper does not specify the simulator used to create both training and test data. The definition of each attribute ('high Doppler', 'Rician fading with K=10') is a hand-chosen boundary that determines the achievable expert accuracy.

assumptions (3)

domain assumption The magnitude of the complex channel impulse response retains sufficient information to classify LoS, Doppler, and fading conditions.
Section II-A defines the wireless observation as a complex CIR h in C^N, but all experts take a real-valued magnitude vector as input (Section III-A, IV-A). The paper does not justify discarding phase, which carries Doppler information in many channel models.
domain assumption The synthetic simulator using a mixture of Rayleigh and Rician components with time-varying Doppler is representative enough for the experts to be accurate in real deployment.
Section IV-B builds the test set from this synthetic pipeline. The paper's conclusion explicitly defers over-the-air validation, so this representativeness is assumed, not demonstrated.
domain assumption MCP (Model Context Protocol) is a reliable, standardized interface that the evaluated LLMs can use to invoke tools.
The design relies on MCP's JSON-RPC tool layer (Section III-B, Fig. 3-5). The paper cites Anthropic's release and a survey, but does not detail failure modes; the empirical results suggest it works in this testbed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Model Context Protocol-based Internet of Experts For Wireless Environment-aware LLM Agents." pith.science (2026). https://pith.science/paper/YY5YMN5D

@misc{pith2026250501834,
  author       = {Pith},
  title        = {Pith review of: Model Context Protocol-based Internet of Experts For Wireless Environment-aware LLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YY5YMN5D}},
  note         = {Machine review of arXiv:2505.01834}
}

read the original abstract

Large Language Models (LLMs) exhibit strong general-purpose reasoning abilities but lack access to wireless environment information due to the absence of native sensory input and domain-specific priors. Previous attempts to apply LLMs in wireless systems either depend on retraining with network-specific data, which compromises language generalization, or rely on manually scripted interfaces, which hinder scalability. To overcome these limitations, we propose a Model Context Protocol (MCP)-based Internet of Experts (IoX) framework that equips LLMs with wireless environment-aware reasoning capabilities. The framework incorporates a set of lightweight expert models, each trained to solve a specific deterministic task in wireless communications, such as detecting a specific wireless attribute, e.g., line-of-sight propagation, Doppler effects, or fading conditions. Through MCP, the LLM can selectively query and interpret expert outputs at inference time, without modifying its own parameters. This architecture enables modular, extensible, and interpretable reasoning over wireless contexts. Evaluated across multiple mainstream LLMs, the proposed wireless environment-aware LLM agents achieve 40%-50% improvements in classification tasks over LLM-only baselines. More broadly, the MCP-based design offers a viable paradigm for future LLMs to inherit structured wireless network management capabilities.

Figures

Figures reproduced from arXiv: 2505.01834 by the authors.

**Figure 2.** The training process of IoX. Raw channel observations are organized into wireless attribute-specific datasets, where positive and negative samples [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Runtime pipeline of an MCP-based wireless environment-aware [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Expert invocation procedure. Given a preprocessed input [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Training and validation performance of three representative expert classifiers. Each subplot shows the evolution of training loss and test accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

From Large AI Models to Agentic AI: A Tutorial on Future Intelligent Communications
cs.AI 2025-05 conditional novelty 2.0 of 10

This paper is a broad tutorial on applying LAMs and agentic AI to 6G, largely restating existing research rather than introducing new results.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages · cited by 1 Pith paper

[1]

Large language model (LLM) for telecommu- nications: A comprehensive survey on principles, key techniques, and opportunities,

H. Zhou, C. Hu, Y . Yuan, Y . Cui, Y . Jin, C. Chen, H. Wu, D. Yuan, L. Jiang, D. Wu et al., “Large language model (LLM) for telecommu- nications: A comprehensive survey on principles, key techniques, and opportunities,” IEEE Commun. Surv. Tutor., to appear, 2025

work page 2025
[2]

Deepseek LLM: Scaling open-source language models with longtermism,

X. Bi, D. Chen, G. Chen, S. Chen, D. Dai, C. Deng, H. Ding, K. Dong, Q. Du, Z. Fu et al. , “Deepseek LLM: Scaling open-source language models with longtermism,” arXiv preprint arXiv:2401.02954, 2024

arXiv 2024
[3]

A survey on large language model- based autonomous agents,

L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y . Lin et al. , “A survey on large language model- based autonomous agents,” Front. Comput. Sci. , vol. 18, no. 6, p. 186345, June 2024

work page 2024
[4]

Large language models for wireless networks: An overview from the prompt engineering perspective,

H. Zhou, C. Hu, D. Yuan, Y . Yuan, D. Wu, X. Chen, H. Tabassum, and X. Liu, “Large language models for wireless networks: An overview from the prompt engineering perspective,” IEEE Wireless Commun. , 2025

work page 2025
[5]

LLM4WM: Adapting LLM for wireless multi-tasking,

X. Liu, S. Gao, B. Liu, X. Cheng, and L. Yang, “LLM4WM: Adapting LLM for wireless multi-tasking,” arXiv preprint arXiv:2501.12983, 2025

arXiv 2025
[6]

Adaptive resource allocation optimization using large language models in dynamic wireless envi- ronments,

H. Noh, B. Shim, and H. J. Yang, “Adaptive resource allocation optimization using large language models in dynamic wireless envi- ronments,” arXiv preprint arXiv:2502.02287, 2025

arXiv 2025
[7]

Rein- forcement learning with large language models (LLMs) interaction for network services,

H. Du, R. Zhang, D. Niyato, J. Kang, Z. Xiong, and D. I. Kim, “Rein- forcement learning with large language models (LLMs) interaction for network services,” in Proc. International Conference on Computing, Networking and Communications (ICNC) . IEEE, 2024, pp. 799–803

work page 2024
[8]

Model context protocol (MCP): Landscape, security threats, and future research directions,

X. Hou, Y . Zhao, S. Wang, and H. Wang, “Model context protocol (MCP): Landscape, security threats, and future research directions,” arXiv preprint arXiv:2503.23278, 2025

arXiv 2025

Show all 14 references

[9]

Model context protocol,

Anthropic, “Model context protocol,” https://www.anthropic.com/ news/model-context-protocol, 2025, accessed: 27 Apr. 2025

2025
[10]

Initial access in millimeter wave cellular systems,

C. N. Barati, S. A. Hosseini, M. Mezzavilla, T. Korakis, S. S. Panwar, S. Rangan, and M. Zorzi, “Initial access in millimeter wave cellular systems,” IEEE Trans. Wireless Commun. , vol. 15, no. 12, pp. 7926– 7940, Dec. 2016

2016
[11]

Model-aided wireless artificial intelligence: Embedding expert knowl- edge in deep neural networks for wireless system optimization,

A. Zappone, M. Di Renzo, M. Debbah, T. T. Lam, and X. Qian, “Model-aided wireless artificial intelligence: Embedding expert knowl- edge in deep neural networks for wireless system optimization,” IEEE Veh. Technol. Mag., vol. 14, no. 3, pp. 60–69, Mar. 2019

2019
[12]

Panic, M

S. Panic, M. Stefanovic, J. Anastasov, and P. Spalevic, Fading and interference mitigation in wireless communications. CRC press, 2013

2013
[13]

Binary cross entropy with deep learning technique for image classification,

U. Ruby, V . Yendapalliet al., “Binary cross entropy with deep learning technique for image classification,” Int. J. Adv. Trends Comput. Sci. Eng, vol. 9, no. 10, 2020

2020
[14]

ToolLLM: Facilitating large language models to master 16000+ real-world APIs,

Y . Qin, S. Liang, Y . Ye, K. Zhu, L. Yan, Y . Lu, Y . Lin, X. Cong, X. Tang, B. Qian et al. , “ToolLLM: Facilitating large language models to master 16000+ real-world APIs,” in Proc. Int. Conf. Learn. Represent., 2024

2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Large language model (LLM) for telecommu- nications: A comprehensive survey on principles, key techniques, and opportunities,

H. Zhou, C. Hu, Y . Yuan, Y . Cui, Y . Jin, C. Chen, H. Wu, D. Yuan, L. Jiang, D. Wu et al., “Large language model (LLM) for telecommu- nications: A comprehensive survey on principles, key techniques, and opportunities,” IEEE Commun. Surv. Tutor., to appear, 2025

work page 2025

[2] [2]

Deepseek LLM: Scaling open-source language models with longtermism,

X. Bi, D. Chen, G. Chen, S. Chen, D. Dai, C. Deng, H. Ding, K. Dong, Q. Du, Z. Fu et al. , “Deepseek LLM: Scaling open-source language models with longtermism,” arXiv preprint arXiv:2401.02954, 2024

arXiv 2024

[3] [3]

A survey on large language model- based autonomous agents,

L. Wang, C. Ma, X. Feng, Z. Zhang, H. Yang, J. Zhang, Z. Chen, J. Tang, X. Chen, Y . Lin et al. , “A survey on large language model- based autonomous agents,” Front. Comput. Sci. , vol. 18, no. 6, p. 186345, June 2024

work page 2024

[4] [4]

Large language models for wireless networks: An overview from the prompt engineering perspective,

H. Zhou, C. Hu, D. Yuan, Y . Yuan, D. Wu, X. Chen, H. Tabassum, and X. Liu, “Large language models for wireless networks: An overview from the prompt engineering perspective,” IEEE Wireless Commun. , 2025

work page 2025

[5] [5]

LLM4WM: Adapting LLM for wireless multi-tasking,

X. Liu, S. Gao, B. Liu, X. Cheng, and L. Yang, “LLM4WM: Adapting LLM for wireless multi-tasking,” arXiv preprint arXiv:2501.12983, 2025

arXiv 2025

[6] [6]

Adaptive resource allocation optimization using large language models in dynamic wireless envi- ronments,

H. Noh, B. Shim, and H. J. Yang, “Adaptive resource allocation optimization using large language models in dynamic wireless envi- ronments,” arXiv preprint arXiv:2502.02287, 2025

arXiv 2025

[7] [7]

Rein- forcement learning with large language models (LLMs) interaction for network services,

H. Du, R. Zhang, D. Niyato, J. Kang, Z. Xiong, and D. I. Kim, “Rein- forcement learning with large language models (LLMs) interaction for network services,” in Proc. International Conference on Computing, Networking and Communications (ICNC) . IEEE, 2024, pp. 799–803

work page 2024

[8] [8]

Model context protocol (MCP): Landscape, security threats, and future research directions,

X. Hou, Y . Zhao, S. Wang, and H. Wang, “Model context protocol (MCP): Landscape, security threats, and future research directions,” arXiv preprint arXiv:2503.23278, 2025

arXiv 2025