Pith. sign in

REVIEW 3 major objections 4 minor 34 references

The paper claims that a deterministic edge flood-detection system can be coupled with a generative large-language-model reporting layer without losing replayability or audit, provided data flows strictly one way and each stakeholder persona

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

2026-08-01 17:02 UTC pith:BX4IEHKO

load-bearing objection Crisp architectural pattern for LLM reporting over deterministic edge logs, honestly scoped; the report groundedness gap is real and acknowledged, so this is a solid workshop contribution, not a validated system. the 3 major comments →

arxiv 2607.17774 v1 pith:BX4IEHKO submitted 2026-07-20 cs.SE cs.AIcs.HC

Persona-as-Configuration: Generative Stakeholder Reporting for Agricultural Floods

Oliver Aleksander Larsen , Tiziano Santilli , Francesco Daghero , Mahyar T. Moghaddam This is my paper
classification cs.SE cs.AIcs.HC
keywords software architecturelarge language modelscyber-physical systemsedge computingflood detectionpersona patternunidirectional data flowauditability
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.

The paper is trying to establish an architectural pattern for adding LLM-generated stakeholder reports to a safety-relevant edge system whose detection path must stay deterministic and replayable. The pattern rests on two invariants: the generative layer consumes the edge logs read-only and never writes back, and stakeholder adaptation is handled by versioned prompt templates rather than runtime improvisation. The authors instantiate this pattern as a dashboard over JSON decision logs from a standing-water detection system and argue that the boundary admits standard reliability mitigations as configuration or middleware changes. If the paper is right, architects can add generative reporting to deterministic cyber-physical systems without giving up auditability, and can add new stakeholder roles simply by adding a new template.

Core claim

The central claim is that the tension between a non-deterministic LLM and a deterministic edge plane can be resolved architecturally rather than by constraining the model. The safety-critical detection path stays bit-for-bit replayable because the LLM is a strict read-only consumer of the JSON decision-log contract, and the effort of adapting one data stream to six stakeholder viewpoints is moved into a versioned prompt configuration artifact. The paper shows one working instantiation—an agricultural flood dashboard—and analyses how input grounding, field-level attribution, schema-constrained generation, and abstention can be attached at the integration seams without redesigning the system.

What carries the argument

The load-bearing mechanism is the Layer 1 / Layer 2 contract: a per-sensor JSON record that is the only data the generative layer ever sees, together with the persona-as-configuration prompt template (system prompt, data-injection template, focus directive) that renders that same JSON into role-specific reports. The unidirectional invariant—Layer 2 never writes back—turns the integration boundary into the architectural guarantee that keeps the edge plane unaffected by LLM non-determinism.

Load-bearing premise

The load-bearing premise is that LLM reports generated from a fixed JSON block and a persona template are factually accurate and actionable enough for stakeholders to rely on; the paper does not yet test this with any agricultural stakeholder.

What would settle it

Run the six persona templates over a fixed JSON decision-log batch, then check every numeric or loss-related claim in each generated report against the schema fields; the insurer excerpt in the paper already contains an unsupported claim ('expected total-loss for ground-floor contents'), so the test is whether field-level attribution, schema-constrained generation, or abstention eliminates such extrapolations in practice.

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

If this is right

  • Architects can add stakeholder-tailored LLM reporting to a deterministic edge CPS without sacrificing bit-for-bit replayability of the detection path, as long as the generative layer is read-only.
  • New stakeholder roles can be added by a domain expert editing a versioned prompt template, with no change to the data pipeline or dashboard code.
  • The same integration boundary can host standard reliability mitigations—grounded input, cited fields, constrained output, abstention—as configuration or middleware changes, making the pattern transferable to other CPS domains.
  • On-demand, stateless invocation keeps LLM cost per report under a cent and keeps the dashboard responsive, addressing a practical blocker for cloud-based LLM layers.
  • If the pattern is adopted widely, similar systems in predictive maintenance, clinical decision support, and fleet operations can separate deterministic control from advisory generative interpretation.

Where Pith is reading between the lines

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

  • The paper's architectural argument is stronger than its empirical evidence; the real test is whether the proposed mitigations (M1–M4) reduce unsupported claims like the insurer's 'expected total-loss' extrapolation to acceptable levels in actual stakeholder use.
  • The unidirectional invariant protects the control plane, but it does not protect the human decision-maker: a confidently wrong read-only report can still trigger a bad action, so the safety boundary shifts to the report-consumption side.
  • Because personas are configuration, prompt templates can be versioned, regression-tested, and A/B compared like any other artifact; this could make persona selection itself an automated, self-adaptive process, as the paper hints with its MAPE-K sketch.
  • The pattern's applicability to other domains depends on whether the edge plane can expose a schema-stable contract; legacy CPS without such a contract would need one introduced before the pattern can attach.

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. The paper proposes a two-layer architecture for agricultural flood monitoring: a deterministic edge detection layer (prior work) producing JSON decision logs, and a human-interface layer that renders maps and generates stakeholder-specific natural-language reports via an LLM. The contributions are two architectural invariants: (C1) unidirectional consumption—the generative layer is a strict read-only consumer of the edge plane and never writes back—and (C2) persona-as-configuration—stakeholder adaptation is encoded as versioned prompt-template artifacts rather than runtime improvisation. The paper instantiates this as a Python dashboard prototype, illustrates six persona outputs from a shared JSON batch, and reports a four-person expert review rated against ISO/IEC 25010-aligned dimensions. The authors explicitly scope the evaluation to architectural properties and state that end-user accuracy and actionability are future work. The central claim is that this integration boundary preserves replayability and auditability of the deterministic plane while admitting generative-reliability mitigations as configuration- or middleware-level extension points.

Significance. If the architectural claims hold, the paper offers a transferable pattern for adding LLM-based stakeholder reporting to deterministic edge CPS without compromising the safety-critical control plane. The paper is strong in making its invariants explicit: C1 is supported by construction, since the described data flow is strictly unidirectional, and C2 is demonstrated by a working prototype and by Table 4, which shows differentiated outputs from one JSON input. The authors are also commendably candid about limitations, including the absence of end-user evaluation and the existence of hallucination/extrapolation risks. Public availability of source code, prompt templates, and decision-log traces is a further strength. However, the significance is tempered by the fact that the paper's usefulness ultimately depends on report groundedness and actionability, and the provided evidence for those properties is limited to a small, explicitly scoped expert review and one unsupported illustrative output. The pattern is architecturally interesting but its practical value remains to be demonstrated.

major comments (3)
  1. [§5.4, §6.3] The only reported evaluation is a four-person expert walk-through explicitly scoped to architectural properties. No agricultural stakeholder assessed report accuracy, actionability, or role-targeting, as the paper itself states in §5.4 and §6.3. Since the abstract and introduction frame the contribution as 'generative stakeholder reporting' whose value depends on stakeholders acting on the reports, this is a load-bearing evidence gap. The word 'safe' in C1 and in the abstract should not be read as covering report content; if the authors intend a broader safety claim, it is unsupported. I recommend either narrowing the claim explicitly to control-plane safety and auditability, or adding at least a small user study.
  2. [Table 4, §6.1] The insurer excerpt contains the assertion 'expected total-loss for ground-floor contents,' which is not present in the JSON schema described in Table 1. The authors acknowledge in §6.1 that this excerpt 'extrapolates beyond the schema.' This is not a mere illustration detail: it contradicts the auditability driver D-Rep, which requires every report to be reconstructable from the decision log. A report containing a material claim absent from the input is not auditable in the sense claimed. The paper should either mark this excerpt explicitly as an example of a failure mode, implement one of the mitigations (e.g., M2 field-level attribution or M3 schema-constrained generation), or revise the auditability claim to exclude ungrounded output.
  3. [§6.2] The reliability-seam analysis lists four mitigations M1–M4, but only M1 (input grounding by construction) is realized in the prototype. M2–M4 are described as configuration- or middleware-level extension points, not as implemented guarantees. This is acceptable as a design analysis, but the paper should avoid suggesting that the architecture currently prevents confidently wrong reports. The future-work discussion (prototyping M3 and M4) correctly treats them as next steps. The distinction between 'the boundary admits mitigations' and 'the system is safe' should be made sharper in the abstract and conclusion.
minor comments (4)
  1. [§4.3 vs. Table 4] The text says the excerpts are generated 'from one sensor record,' while the Table 4 caption says they come from a '25-sensor decision-log batch.' Please reconcile this inconsistency.
  2. [Table captions] Several table captions contain 'T able' instead of 'Table' (e.g., Tables 1–5). This appears to be a formatting artifact and should be corrected.
  3. [§6.1] The claim that a typical report 'remains under one cent at GPT-5-mini pricing' is unsupported by a citation or calculation. Please add a reference or remove the quantitative claim.
  4. [§6.2] The term 'safety surface' is used but not defined. Please define it or replace with a more precise term.

Circularity Check

0 steps flagged

No significant circularity: the paper's architectural claims are design invariants, not fitted predictions, and its self-citation is not load-bearing.

full rationale

The paper does not present a derivation chain in which an output is computed from fitted inputs or in which a prediction is equivalent to its calibration by construction. The two central invariants, C1 (unidirectional consumption) and C2 (persona-as-configuration), are defined architectural properties rather than empirically derived results: C1 guarantees the edge plane is unaffected by the generative layer because the boundary is defined as read-only, and C2 is a design pattern for organizing prompt templates. Neither claim is produced by fitting data or by renaming a measured outcome. The only self-citation, [13], supplies the black-box edge detection system used as the Layer 1 producer, and the paper explicitly treats that system as a black-box contract; the value of the present contribution does not depend on the correctness of [13]'s internal algorithms. The expert review is an evaluation of the architecture, not a circular derivation: panel members are not the authors, though the paper acknowledges in §5.4 that panellists aware of the project may introduce positive bias and that notes were coded by two authors. These are validity threats, not circularity. Finally, the paper itself identifies the most significant limitations: lack of end-user evaluation (§5.4, §6.3) and the Table 4 insurance excerpt that 'extrapolates beyond the schema' (§6.1). These are evidence gaps about report groundedness, not instances where an output reduces to its input by construction. Accordingly, no circular step is present and the score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No numeric free parameters are fitted in this paper; thresholds and prompt templates are configuration artifacts inherited or devised, not fitted values. The listed axioms are the domain and evaluation assumptions the claims rest on. No new physical or conceptual entities (forces, particles, dimensions) are introduced; 'persona-as-configuration' is a named architectural artifact, not an entity with independent falsifiable evidence.

axioms (5)
  • domain assumption The six stakeholder types and their focus areas are the correct and complete set for agricultural flood reporting.
    §3.1 (D-Stk) and Table 3 define the personas; no stakeholder validation establishes that these six types and focus areas are sufficient or correct.
  • domain assumption LLM-generated natural-language reports are an appropriate and sufficiently accurate medium for stakeholder action.
    The paper motivates LLM multiplexing in §1 and §4 but explicitly defers accuracy and actionability testing to future work (§5.4, §6.3).
  • domain assumption The Layer 1 / Layer 2 JSON contract (Table 1) is stable and sufficient for report generation.
    §3.2 says every design choice is made against this contract; a changing or incomplete schema would break reconstructability and the persona templates.
  • domain assumption GPT-5-mini availability and pricing as of Q1 2026 are as stated.
    §6.1 asserts a 2,000–3,000-token report is under one cent at GPT-5-mini pricing; this external cost claim is not independently verified in the paper.
  • ad hoc to paper A four-person expert walk-through with possibly project-aware panellists is adequate evidence that the architectural claims hold.
    §5.1 and §5.4 describe the panel and concede positive-bias risk; this is a weak evidential base for a claimed transferable pattern.

pith-pipeline@v1.3.0-alltime-deepseek · 10517 in / 10434 out tokens · 119241 ms · 2026-08-01T17:02:13.502796+00:00 · methodology

0 comments
read the original abstract

Cyber-physical systems built on deterministic edge inference, such as on-vehicle flood detection for agricultural fields, produce structured decision logs that must be interpreted differently by heterogeneous stakeholders. Pairing such systems with large language models (LLMs) to generate stakeholder-specific reports introduces a tension: the generative layer is non-deterministic, while the edge plane must remain replayable and auditable. We propose an architectural pattern resting on two invariants: unidirectional consumption, in which the generative layer is a strict read-only consumer of the deterministic plane and never writes back, and persona-as-configuration, in which stakeholder adaptation is a versioned prompt-template artifact rather than runtime improvisation. We instantiate the pattern as a context-aware dashboard layer over the JSON decision logs of a previously published edge-based standing-water detection system, and analyse how the integration boundary admits standard generative-reliability mitigations as configuration- or middleware-level extension points. A structured expert review rated the pattern favourably across five ISO/IEC 25010-aligned quality dimensions, with strongest agreement on separation of concerns. End-user evaluation with agricultural stakeholders is planned for future work.

Figures

Figures reproduced from arXiv: 2607.17774 by Francesco Daghero, Mahyar T. Moghaddam, Oliver Aleksander Larsen, Tiziano Santilli.

Figure 1
Figure 1. Figure 1: Two-layer architecture. Layer 1 (edge detection, described in [13]) writes JSON decision logs that fix the Layer 1 / Layer 2 contract. Layer 2 (this paper’s contribution) consumes them for visualization and on-demand stakeholder-tailored LLM reporting. Data flows strictly downward. The unidirectional invariant prevents the generative layer from writing back to the safety-critical detection plane [PITH_FUL… view at source ↗
Figure 2
Figure 2. Figure 2: Dashboard prototype showing the Fyn Island monitoring area. Colour-coded map pins indicate per-sensor flood classifications (red = Flooded, orange = Suspicious, green = Normal). Summary cards report aggregate counts. The Generate AI Report button triggers stakeholder-tailored report generation via GPT-5-mini. phrasing variability. Substitutability (D-Sub) confines provider choice to a narrow interface (sen… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 5 canonical work pages

  1. [1]

    Patterns7(1), 101429 (2026)

    An, T., Zhou, Y., Zou, H., Yang, J.: IoT-LLM: A framework for enhancing large lan- guage model reasoning from real-world sensor data. Patterns7(1), 101429 (2026). https://doi.org/10.1016/j.patter.2025.101429

    arXiv 2026

  2. [2]

    Addison- Wesley Professional, Boston, MA, 4 edn

    Bass, L., Clements, P., Kazman, R.: Software Architecture in Practice. Addison- Wesley Professional, Boston, MA, 4 edn. (2021)

    2021

  3. [3]

    In: Proc

    Bucaioni, A., Weyssow, M., He, J., Lyu, Y., Lo, D.: A functional software refer- ence architecture for LLM-integrated systems. In: Proc. IEEE ICSA 2025 Compan- ion (New and Emerging Ideas). pp. 1-5 (2025).https://doi.org/10.1109/ICSA- C65153.2025.00006

    arXiv 2025

  4. [4]

    Sensors24(16), 5320 (2024).https://doi.org/10.3390/s24165320

    Dauda, A., Flauzac, O., Nolot, F.: A survey on IoT application architectures. Sensors24(16), 5320 (2024).https://doi.org/10.3390/s24165320

    work page doi:10.3390/s24165320 2024

  5. [5]

    Future Generation Computer Systems171, 107829 (2025).https://doi.org/10.1016/j.future.2025.107829

    De Vito, G., Palomba, F., Ferrucci, F.: The role of large language models in address- ing IoT challenges: A systematic literature review. Future Generation Computer Systems171, 107829 (2025).https://doi.org/10.1016/j.future.2025.107829

    arXiv 2025

  6. [6]

    arXiv preprint arXiv:2312.10997 (2023)

    Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, M., Wang, H.: Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 (2023)

    Pith/arXiv arXiv 2023

  7. [7]

    AI7(2), 51 (2026).https://doi.org/10.3390/ ai7020051

    Gonz´ alez-Potes, A., Mart ´ ınez-Castro, D., Paredes, C.M., Ochoa-Brust, A., Mena, L.J., Mart ´ ınez-Pel´ aez, R., F´ elix, V.G., F´ elix, R.A.: Hybrid AI and LLM-enabled agent-based real-time decision support architecture for industrial batch processes: A clean-in-place case study. AI7(2), 51 (2026).https://doi.org/10.3390/ ai7020051

    2026

  8. [8]

    RFC 9556 (Apr 2024)

    Hong, J., Hong, Y.G., de Foy, X., Kovatsch, M., Schooler, E., Kutscher, D.: Internet of Things (IoT) edge challenges and functions. RFC 9556 (Apr 2024)

    2024

  9. [9]

    White Paper G507-2065-00, IBM Corporation, Hawthorne, NY, USA (2006)

    IBM Corporation: An architectural blueprint for autonomic computing. White Paper G507-2065-00, IBM Corporation, Hawthorne, NY, USA (2006)

    2065

  10. [10]

    Standard, ISO, Geneva, Switzerland (2023), edition 2 16 O

    International Organization for Standardization: ISO/IEC 25010:2023, systems and software engineering, systems and software quality requirements and evaluation (SQuaRE), product quality model. Standard, ISO, Geneva, Switzerland (2023), edition 2 16 O. A. Larsen et al

    2023

  11. [11]

    Computer36(1), 41-50 (2003).https://doi.org/10.1109/MC.2003.1160055

    Kephart, J.O., Chess, D.M.: The vision of autonomic computing. Computer36(1), 41-50 (2003).https://doi.org/10.1109/MC.2003.1160055

    Pith/arXiv arXiv 2003

  12. [12]

    Pervasive and Mobile Computing 17, 184-206 (2015).https://doi.org/10.1016/j.pmcj.2014.09.009

    Krupitzer, C., Roth, F.M., VanSyckel, S., Schiele, G., Becker, C.: A survey on engineering approaches for self-adaptive systems. Pervasive and Mobile Computing 17, 184-206 (2015).https://doi.org/10.1016/j.pmcj.2014.09.009

    work page doi:10.1016/j.pmcj.2014.09.009 2015

  13. [13]

    In: Proc

    Larsen, O.A., Moghaddam, M.T.: Edge-based standing-water detection via FSM- guided tiering and multi-model consensus. In: Proc. IEEE Int. Conf. Software Ar- chitecture (ICSA), In Practice Track (2026),https://arxiv.org/abs/2604.03308, arXiv:2604.03308

    Pith/arXiv arXiv 2026

  14. [14]

    Sensors25(12), 3583 (2025).https://doi.org/10.3390/s25123583

    Miller, T., Mikiciuk, G., Durlik, I., Mikiciuk, M., Lobodzi´ nska, A.,´Snieg, M.: The IoT and AI in agriculture: The time is now, a systematic review of smart sensing technologies. Sensors25(12), 3583 (2025).https://doi.org/10.3390/s25123583

    work page doi:10.3390/s25123583 2025

  15. [15]

    Naikwade, S.S., Nannaware, Y.S., Narkhede, A.A., Navale, P.S., Nikampatil, A.S., Bhaskar, T.: IoT-based flood level monitoring system with real-time alert and dashboard visualization. J. Sensor and Cloud Computing2(2) (2025)

    2025

  16. [16]

    OpenAI: GPT-5 mini model card.https://developers.openai.com/api/docs/ models/gpt-5-mini(2025)

    2025

  17. [17]

    Plotly Technologies Inc.: Dash: Analytical web apps for Python.https:// dash.plotly.com/(2024)

    2024

  18. [18]

    Cambridge University Press, Cambridge, UK (2000)

    Reiter, E., Dale, R.: Building Natural Language Generation Systems. Cambridge University Press, Cambridge, UK (2000)

    2000

  19. [19]

    https://python-visualization.github.io/folium/(2024)

    Rob Story and the Folium contributors: Folium: Python data, Leaflet.js maps. https://python-visualization.github.io/folium/(2024)

    2024

  20. [20]

    Frontiers in Water4, 786040 (2022).https://doi.org/10.3389/ frwa.2022.786040

    Samadi, S.: The convergence of AI, IoT, and big data for advancing flood ana- lytics research. Frontiers in Water4, 786040 (2022).https://doi.org/10.3389/ frwa.2022.786040

    arXiv 2022

  21. [21]

    In: Proc

    Schreiber, W., White, J., Schmidt, D.C.: Toward a pattern language for persona- based interactions with LLMs. In: Proc. 31st Conf. on Pattern Languages of Pro- grams (PLoP). ACM (2024).https://doi.org/10.64346/PLoP2024p27

    work page doi:10.64346/plop2024p27 2024

  22. [22]

    arXiv preprint arXiv:2406.06608 (2024)

    Schulhoff, S., Ilie, M., Balepur, N., Kahadze, K., Liu, A., Si, C., Li, Y., Gupta, A., Han, H., et al.: The prompt report: A systematic survey of prompt engineering techniques. arXiv preprint arXiv:2406.06608 (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    In: Proc

    Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaud- hary, V., Young, M., Crespo, J.F., Dennison, D.: Hidden technical debt in machine learning systems. In: Proc. NIPS. pp. 2503-2511 (2015)

    2015

  24. [24]

    In: Proc

    Shamsujjoha, M., Lu, Q., Zhao, D., Zhu, L.: Swiss cheese model for AI safety: A taxonomy and reference architecture for multi-layered guardrails of founda- tion model based agents. In: Proc. IEEE ICSA 2025. pp. 37-48 (2025),https: //ieeexplore.ieee.org/document/10978931

    arXiv 2025

  25. [25]

    Sensors24(21), 7090 (2024).https://doi.org/10.3390/s24217090

    Tao, Y., Tian, B., Adhikari, B.R., Zuo, Q., Luo, X., Di, B.: A review of cutting- edge sensor technologies for improved flood monitoring and damage assessment. Sensors24(21), 7090 (2024).https://doi.org/10.3390/s24217090

    work page doi:10.3390/s24217090 2024

  26. [26]

    In: Findings of the Association for Computational Linguistics: EMNLP

    Tseng, Y.M., Huang, Y.C., Hsiao, T.Y., Chen, W.L., Huang, C.W., Meng, Y., Chen, Y.N.: Two tales of persona in LLMs: A survey of role-playing and personal- ization. In: Findings of the Association for Computational Linguistics: EMNLP

  27. [27]

    arXiv preprint arXiv:2302.11382 (2023) Persona-as-Configuration for Agricultural Flood Reporting 17

    White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert, H., Elnashar, A., Spencer-Smith, J., Schmidt, D.C.: A prompt pattern catalog to enhance prompt engineering with ChatGPT. arXiv preprint arXiv:2302.11382 (2023) Persona-as-Configuration for Agricultural Flood Reporting 17

    Pith/arXiv arXiv 2023

  28. [28]

    arXiv preprint arXiv:2307.09702 (2023)

    Willard, B.T., Louf, R.: Efficient guided generation for large language models. arXiv preprint arXiv:2307.09702 (2023)

    Pith/arXiv arXiv 2023

  29. [29]

    Xu, F., Ma, J., Li, N., Cheng, J.C.P.: Large language model applications in disaster management: An interdisciplinary review. Int. J. Disaster Risk Reduction127, 105642 (2025).https://doi.org/10.1016/j.ijdrr.2025.105642

    arXiv 2025

  30. [30]

    In: Proc

    Xu, W., Liu, M., Sokolsky, O., Lee, I., Kong, F.: LLM-enabled cyber-physical sys- tems: Survey, research opportunities, and challenges. In: Proc. IEEE Int. Workshop on Foundation Models for Cyber-Physical Systems and IoT (FMSys), at CPS-IoT Week (2024),https://ieeexplore.ieee.org/document/10590269

    arXiv 2024

  31. [31]

    Frontiers in Plant Science16, 1668545 (2025).https://doi.org/ 10.3389/fpls.2025.1668545

    Yu, P., Teng, F., Zhu, W., Shen, C., Chen, Z., Song, J.: Cloud-edge-device col- laborative computing in smart agriculture: Architectures, applications, and future perspectives. Frontiers in Plant Science16, 1668545 (2025).https://doi.org/ 10.3389/fpls.2025.1668545

    arXiv 2025

  32. [32]

    IEEE Access8, 141748-141761 (2020).https://doi.org/10.1109/ACCESS.2020.3013005

    Zhang, X., Cao, Z., Dong, W.: Overview of edge computing in the agricultural Internet of Things: Key technologies, applications, challenges. IEEE Access8, 141748-141761 (2020).https://doi.org/10.1109/ACCESS.2020.3013005

    arXiv 2020

  33. [33]

    a helpful assistant

    Zheng, M., Pei, J., Logeswaran, L., Lee, M., Jurgens, D.: When “a helpful assistant” is not really helpful: Personas in system prompts do not improve performances of large language models. In: Findings of the Association for Computational Lin- guistics: EMNLP 2024. pp. 15126-15154 (2024).https://doi.org/10.18653/v1/ 2024.findings-emnlp.888

    doi:10.18653/v1/ 2024

  34. [2024]

    16612-16631 (2024).https://doi.org/10.18653/v1/2024.findings- emnlp.969

    pp. 16612-16631 (2024).https://doi.org/10.18653/v1/2024.findings- emnlp.969

    doi:10.18653/v1/2024.findings- 2024