Pith. sign in

REVIEW 5 major objections 5 minor 40 references

The authors claim that policy-constrained LLM decision support is best framed as an adaptive compute-allocation problem, and that a memory-routed architecture with an expert-verified cache can deliver policy-compliant, actionable advice at

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

2026-08-02 03:16 UTC pith:27QOJJI2

load-bearing objection A genuine systems contribution with a coherent architecture, but the headline numbers rest on an unreleased, likely in-sample simulated benchmark and an unvalidated judge dimension. the 5 major comments →

arxiv 2607.13934 v1 pith:27QOJJI2 submitted 2026-07-15 cs.MA

Pezego-HITL: A policy-grounded large language model architecture for agricultural extension in Ghana

Shunbao Li , Zhipeng Yuan , Amoako Ofori , Benedicta Y. Fosu-Mensah , Yang Li , Manu Kenchappa Junjanna , Qing Xue , Po Yang This is my paper
classification cs.MA
keywords agricultural extensionlarge language modelsretrieval-augmented generationdecision support systemspolicy compliancehuman-in-the-loop AIsmallholder farmingpest management
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 authors try to establish that safe, useful AI advice for smallholder crop protection is not mainly a matter of a better language model; it is a compute-allocation problem. They propose an architecture that answers recurring queries from a cache of expert-verified recommendations, and only runs expensive generation with policy checking for new cases. On 1,240 simulated field queries, they report that this routing lifts policy alignment to 0.94 and agronomic utility to 0.95 while cutting P95 latency from 28.6s to 12.9s, with about 60% of queries served from cache. They also introduce P-EVAL, a protocol that measures safety, usefulness, latency, and expert editing burden together. A sympathetic reader would care because the claimed trade-off—safety without sacrificing speed—is exactly what makes AI decision support viable in low-connectivity rural settings.

Core claim

The central claim is that policy-constrained LLM assessment should be treated as an adaptive test-time compute allocation problem, instantiated by Pezego-HITL. The architecture combines schema-aware retrieval that translates natural-language field queries into SQL constraints over a pesticide registry, an automated constraint-auditing layer that verifies active ingredients and growth-stage compatibility, and a Verified-Case Memory (VCM) that stores expert-approved recommendations. When an incoming query matches a cached case above a structural-semantic threshold (0.83), the system serves the verified recommendation directly, bypassing generation; otherwise it runs a multi-agent pipeline with

What carries the argument

The load-bearing object is the Verified-Case Memory (VCM), a cache of expert-approved recommendation templates indexed by a composite similarity score s = α s_struct + (1−α)s_semantic, with α=0.6. Structural similarity is gated by exact matches on crop, pest, and growth stage; semantic similarity is cosine distance of dense embeddings. Queries with s ≥ τ=0.83 are served from cache, which the paper argues is what converts a safety-verified but slow pipeline (28.6s tail) into a usable one (12.9s) while preserving compliance. The threshold τ is the control knob on the safety-latency frontier: lower thresholds increase reuse but risk policy drift; higher thresholds preserve compliance at the cos

Load-bearing premise

The load-bearing premise is that the 1,240 simulated field queries faithfully represent real farmer questions, and that the LLM judge—calibrated on only 135 expert-reviewed reports—labels the remaining cases correctly; the paper does not describe how the cases were generated or how judge error propagates.

What would settle it

Run the same architecture on an independent set of, say, 300 real or freshly elicited farmer queries that were not drawn from the architecture's own policy templates or VCM entries, and have a panel of agronomists blind to the system label each output. If PAR/AUR drop substantially below 0.94/0.95, or if judge–expert agreement on the new set is much lower than κ=0.77, the headline numbers are partly an artifact of the evaluation setup.

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

If this is right

  • If the results hold, safety compliance and operational latency need not be in opposition: an expert-verified cache can absorb the recurring fraction of queries and reserve expensive generation for genuinely new cases.
  • The framework makes the supervision burden measurable: the paper reports a 14.1% expert modification rate and a mean edit ratio of 38.3% on 135 real reports, and folds these into the P-EVAL objective.
  • The architecture is model-agnostic: the same interventions raise the open-source model's PAR from 0.44 to 0.86 and cut P95 latency by 54.5%, suggesting local deployment is feasible without sacrificing compliance.
  • The VCM threshold τ acts as an explicit Pareto control; the paper claims τ=0.83 is a balanced operating point with 59.6% cache reuse, and that this generalizes across model classes.

Where Pith is reading between the lines

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

  • Editorial inference: if the simulated 1,240-case benchmark were partly generated from the same policy templates or VCM entries that the architecture uses, the reported PAR/AUR and reuse ratio would be optimistic; an independent field sample or hold-out set of queries would be needed to confirm.
  • Editorial inference: the cache-reuse mechanism suggests a testable general principle for other high-stakes advisory domains (e.g., clinical or legal triage): route by verified precedent first, generate only on novelty, and use expert edits to grow the precedent base.
  • Editorial inference: the paper leaves the generation of the simulated workload opaque; varying those cases to measure whether VCM reuse genuinely reflects recurring farmer questions could distinguish a caching effect from an in-sample artifact.

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

5 major / 5 minor

Summary. The paper proposes Pezego-HITL, a human-in-the-loop LLM decision-support architecture for agricultural extension in Ghana. It combines schema-aware SQL retrieval against registered pesticide/policy databases, an automated constraint auditing layer, critique agents, and an expert-verified case memory (VCM) cache that serves previously approved templates when a composite structural/semantic similarity exceeds a threshold. The paper also introduces P-EVAL, a multi-objective evaluation protocol measuring Policy Alignment Rate (PAR), Agronomic Utility Rate (AUR), P95 latency, and supervision overhead. The main empirical claims are based on a simulated 1,240-query workload: PAR 0.94, AUR 0.95, P95 latency 12.9 s versus 28.6 s for a critique-only baseline, and a 59.6% cache reuse ratio, with a parallel open-source model evaluation (PAR 0.86). The socio-technical section reports survey responses from 30 Ghanaian ESOs and 36 smallholder farmers.

Significance. The architectural idea of routing recurring, structurally similar queries to an expert-verified cache is practical and relevant for low-bandwidth agricultural settings, and the attempt to formalize safety, utility, latency, and supervision in one evaluation framework is useful. The two-cohort survey adds credible qualitative evidence about trust and usability. However, the paper's central quantitative claims currently rest on an unreleased, vaguely described simulated benchmark and on an LLM judge that was calibrated on only 135 reports with 19 positive cases. If the benchmark and judge validation were made rigorous and transparent, the contribution could be a valuable template for policy-constrained deployment; as submitted, the headline numbers are not yet established.

major comments (5)
  1. [§5.2, Table 1] The simulated 1,240-query benchmark is described only as 'representative' of typical crop cycles in Eastern and Ashanti regions. No sampling procedure, source query log, crop/pest distribution, or independence from the VCM template database is given. If the queries were generated from the same policy/template database that populates the VCM, the 59.6% reuse ratio and the resulting PAR/AUR are partly predetermined. Please release the dataset or provide a detailed generation protocol, and evaluate on a held-out set constructed independently of the VCM templates.
  2. [§5.1, Table 1] The LLM judge calibration is both inconsistent and insufficiently validated. The confusion matrix in §5.1 (TP=16, FP=2, TN=114, with 3 FN implied) gives accuracy (16+114)/135 = 96.3%, not the reported 91.9%. Additionally, calibration covers only the binary compliance decision, whereas AUR in Eq. (4) is a multi-criteria score, so AUR = 0.95 has no human calibration. Report the corrected accuracy, per-criterion agreement for AUR, and confidence intervals, ideally with a larger human-labeled sample.
  3. [§3.1, Eq. (1), Fig. 4] The routing parameters alpha = 0.6 and tau = 0.83 are selected or ablated on the same 1,240-case benchmark that produces the headline results. Figure 4 identifies tau = 0.83 as a 'Pareto-optimal' point on this benchmark, so the reported reuse ratio and PAR/latency numbers are in-sample selection results, not unbiased out-of-sample estimates. Please provide a nested or cross-validated evaluation, or report the variance of the metrics across resamples of the benchmark.
  4. [§5.2, Table 1] The reported PAR/AUR for Pezego-HITL conflate two different processes: 59.6% of queries are served from the VCM cache and inherit prior expert approval, while only 40.4% go through the generation pipeline. Comparing the aggregate 0.94/0.95 to B4's 0.89/0.90 therefore overstates the architecture's generation quality. Report PAR/AUR separately for cache-routed and generation-routed cases, and compare B4 against Pezego-HITL on the cache-miss subset alone. The latency advantage of cache reuse is real but should be presented as a caching benefit, not as a generation-quality improvement.
  5. [§5.2, Table 1] No confidence intervals, standard errors, or significance tests are reported for any headline metric (PAR, AUR, P95 latency, reuse ratio). With N=1,240, bootstrapped intervals would be straightforward and are necessary to assess whether differences such as 0.94 vs 0.89 are meaningful, especially given the judge's calibration error rate.
minor comments (5)
  1. [§5.1] Correct the reported accuracy from 91.9% to 96.3% based on the given confusion matrix, or explain the discrepancy if a different denominator was used.
  2. [§5.5] The phrase 'high student-to-teacher ratio' appears to be a typo; in context it should be 'officer-to-farmer ratio'.
  3. [Fig. 4] The shaded operating region tau in [0.80, 0.85] is described qualitatively. Adding the actual PAR and latency values at the endpoints and at tau = 0.83 would make the sensitivity analysis more informative.
  4. [§4.1, Ref. [33]] The open-source model 'Qwen3.5-9B-DeepSeek-V4-Flash' is cited only via a HuggingFace URL with no version or release date; verify the model name and provide a stable citation.
  5. [§1] The introduction contains two overlapping paragraphs beginning 'To address these limitations...' and 'To address these issues...' that cover nearly identical ground; consolidate to reduce redundancy.

Circularity Check

0 steps flagged

No significant circularity: headline PAR/AUR/latency are empirical telemetry, not identities with inputs.

full rationale

The derivation chain defines Pezego-HITL operations (VCM similarity Eq. 1, retrieval/auditing, expert feedback) and P-EVAL metrics (PAR/AUR/P95, Eqs. 2-6), then measures those metrics on 1,240 simulated cases after calibrating the judge on 135 ESO-verified reports. No equation of the paper reduces to another by construction. High PAR/AUR on cache hits is expected because the cache stores expert-approved templates, but the paper explicitly claims this reuse mechanism; that is an empirical mechanism, not a tautological derivation. The routing threshold tau=0.83 is selected by ablation on the same simulated benchmark (Section 5.3, Fig. 4), so headline numbers are in-sample rather than out-of-sample; this is a generalization/experimental-design weakness, not definitional circularity. The simulated benchmark construction is not described (Section 5.2), so possible template leakage cannot be ruled out, but it is not shown or stated in the paper. The LLM judge is calibrated only on binary compliance, leaving AUR uncalibrated, and the stated confusion matrix arithmetic appears inconsistent (TP=16, FP=2, TN=114 implies 96.3% accuracy, not 91.9%); these are validation/reporting concerns, not circularity. Self-citations (PEZEGO [1], TARAG [21]) are contextual, not load-bearing. Therefore no significant circularity is found.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central results rest on a simulated workload whose generation is undisclosed, an LLM judge extrapolated from 135 calibration reports, an expert-approved cache whose reuse is governed by hand-set similarity weights, and self-reported field surveys. The ledger lists the main hand-chosen parameters and domain assumptions; no new physical entities are postulated.

free parameters (4)
  • alpha (VCM similarity mixing weight) = 0.6
    Eq. 1 balances structural vs semantic similarity; chosen by hand with no sensitivity analysis.
  • tau (VCM routing threshold) = 0.83
    Selected from ablation over [0.70, 0.95] as 'Pareto-optimal'; directly determines the 59.6% cache reuse ratio, hence P95 latency and PAR composition.
  • PAR/AUR rubric (constraint set M and utility criteria K) = unspecified
    Equations 3-4 define PAR/AUR over M binary constraints and K utility criteria, but the actual lists used by the judge are not enumerated; the reported 0.94/0.95 depend on this unstated rubric.
  • Supervision overhead weights lambda_1, lambda_2 = not assigned
    Eq. 5 includes lambda_1 and lambda_2, but they are never set or optimized; the paper instead reports modification rate and edit ratio, so the formal H(theta) objective is not actually evaluated.
axioms (6)
  • domain assumption The 1,240 simulated queries are representative of real smallholder field queries and are not generated from the VCM templates or policy database under test.
    Section 5.2 calls the dataset 'representative' but gives no generation protocol, source data, or independence guarantee; this underpins all PAR/AUR/latency headline numbers.
  • domain assumption The LLM-as-judge (gpt-5.5-pro) generalizes from 135 calibration reports to correctly label policy compliance and utility on the 1,240 simulated cases.
    Section 4.3/5.1: calibration kappa=0.77 with 16/19 recall and 2 false positives; extrapolation to the unlabeled simulated set is assumed, not tested.
  • domain assumption The 135 ESO verification decisions constitute gold-standard ground truth for policy compliance.
    Section 5.4 uses ESO edits as ground truth; no independent inter-rater reliability or adjudication by a second expert panel is reported.
  • domain assumption The Ghana EPA/pesticide registry used by the Constraint Auditing Layer is complete and current.
    Section 3.1 and Appendix C assume the registry contains all registered active ingredients, banned chemicals, and growth-stage rules; its provenance is not cited or released.
  • domain assumption An expert-approved VCM template remains safe to serve for any query whose structural/semantic similarity score exceeds tau=0.83.
    Section 3.1/Eq. 1 formalizes reuse by cosine similarity plus boolean crop-pest-stage match; the paper provides no analysis of false-positive cache hits causing safety drift.
  • domain assumption Self-reported Likert ratings and perceived yield savings are adequate proxies for real usability and crop-protection outcomes.
    Sections 5.5-5.6 and Fig. E.12 rely on perception data; no yield measurements, randomized control, or longitudinal adoption data are reported.

pith-pipeline@v1.3.0-alltime-deepseek · 19772 in / 16417 out tokens · 153987 ms · 2026-08-02T03:16:21.265353+00:00 · methodology

0 comments
read the original abstract

Large language models are increasingly deployed in agricultural decision-support settings, yet high-stakes crop protection in smallholder agriculture requires more than output-quality benchmarks. Over a two-year design and evaluation programme, we formalise policy-constrained large language model assessment as an adaptive compute allocation problem that jointly captures safety compliance, helpfulness, operational latency, and expert supervision workload. We introduce P-EVAL (Policy-grounded Expert-calibrated VALidation protocol), a unified evaluation framework for policy-grounded decision support, evaluating the architecture on a simulated field query database consisting of 1,240 cases. The protocol is instantiated on the Pezego advisory architecture (Pezego-HITL) and evaluated in Ghana. Following offline judge calibration against gold-standard human expert decisions ($\kappa = 0.77$), we evaluate the architectural performance under simulated query workloads. Under P-EVAL, our memory-routed architecture improves the Policy Alignment Rate (PAR) to 0.94 and the Agronomic Utility Rate (AUR) to 0.95, while reducing P95 latency by 55% (from 28.6s to 12.9s) through a 59.6% cache reuse ratio. We also demonstrate generalisability using the open-source \texttt{Qwen3.5-9B-DeepSeek-V4-Flash} model, achieving a PAR of 0.86 and a 54.5% latency reduction (to 10.2s). To evaluate practical utility and socio-technical integration, we administer detailed questionnaires to Ghanaian Extension Services Officers ($N=30$) and smallholder farmers ($N=36$). Taken together, this work demonstrates how policy-grounded structured retrieval-augmented generation with validated-memory routing makes safety-utility-latency trade-offs explicit, offering a scalable template for trustworthy AI-driven extension in smallholder farming systems.

Figures

Figures reproduced from arXiv: 2607.13934 by Amoako Ofori, Benedicta Y. Fosu-Mensah, Manu Kenchappa Junjanna, Po Yang, Qing Xue, Shunbao Li, Yang Li, Zhipeng Yuan.

Figure 1
Figure 1. Figure 1: 3.1. Pezego-HITL: Decoupled Policy-Aligned Architecture The proposed architecture decomposes query resolution into a modular, multi-agent generative pipeline and an expert-verified cache routing layer (VCM), configured as follows: 1. Policy-Grounded Structured Retrieval and Auditing: When a query requires raw generation, it is routed to our schema-aware Retrieval-Augmented Generation (SRAG) pipeline. This … view at source ↗
Figure 1
Figure 1. Figure 1: Decoupled policy-constrained decision-support architecture (Pezego-HITL). The proposed framework separates [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Integrated evaluation protocol workflow. The protocol comprises two primary streams: (A) Online Telemetry (green [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: End-to-end latency distribution comparing B4 and Pezego-HITL. The left panel (CDF) shows a consistent left shift [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation of memory-routing threshold τ in the VCM mechanism. Lower thresholds increase reuse ratio but may risk quality drift (reduced compliance), while higher thresholds preserve strict matching at the cost of reduced reuse and higher latency. The shaded interval (τ ∈ [0.80, 0.85]) marks a practical operating region where policy compliance remains stable, VCM reuse ratio is high, and latency is minimised… view at source ↗
Figure 5
Figure 5. Figure 5: Results of the empirical validation on 135 crop protection reports reviewed by Ghanaian ESOs. Panel (a) reports the [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] 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

40 extracted references · 4 linked inside Pith

  1. [1]

    Z. Yuan, K. Liu, S. Li, R. Peng, D. Leybourne, N. Musa, P. Yang, PEZEGO: A Precision Agriculture System Based on Large Language Models and Internet of Things for Pest Management, IEEE Internet of Things Journal, 2025

    2025

  2. [2]

    Lewis, E

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, D. Kiela, Retrieval-augmented generation for knowledge-intensive NLP tasks, Advances in Neural Information Processing Systems 33 (2020) 9459-9474

    2020

  3. [3]

    Schick, J

    T. Schick, J. Dwivedi-Yu, R. Dessì, R. Flynn, L. Dunlap, P. Yih, D. Metzler, L. Zettlemoyer, T. Riedel, Toolformer: Language models can teach themselves to use tools, Advances in Neural Information Processing Systems 36 (2024)

    2024

  4. [4]

    Madaan, N

    A. Madaan, N. Tandon, P. Gupta, P. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, S. Gupta, B. P. Lim, S. Le, E. Hovy, D. Pratt-Hartmann, Self-refine: Iterative refinement with self-feedback, Advances in Neural Information Processing Systems 36 (2023)

    2023

  5. [5]

    S. Li, Z. Yuan, R. Peng, D. Leybourne, Q. Xue, Y. Li, P. Yang, An effective farmer-centred mobile intelligence solution using lightweight deep learning for integrated wheat pest management, Journal of Industrial Information Integration 42 (2024) 100705

    2024

  6. [6]

    M. Gao, Y. Liu, X. Hu, X. Wan, J. Bragg, A. Cohan, Re-evaluating Automatic LLM System Ranking for Alignment with Human Preference, Findings of the Association for Computational Linguistics: NAACL 2025. 22

    2025

  7. [7]

    Y. Liu, P. Liu, A. Cohan, On Evaluating LLM Alignment by Evaluating LLMs as Judges, Advances in Neural Information Processing Systems, 2025

    2025

  8. [8]

    Ashkinaze, H

    J. Ashkinaze, H. Shen, S. Avula, E. Gilbert, C. Budak, Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences, Advances in Neural Information Processing Systems, 2025

    2025

  9. [9]

    D. Li, R. Sun, Y. Huang, M. Zhong, B. Jiang, J. Han, X. Zhang, W. Wang, H. Liu, Preference Leakage: A Contamination Problem in LLM-as-a-Judge, International Conference on Learning Representations, 2025

    2025

  10. [10]

    Chang, Z

    C.-Y. Chang, Z. Jiang, V. Rakesh, M. Pan, C.-C. M. Yeh, G. Wang, M. Hu, Z. Xu, Y. Zheng, H. Ji, MAIN-RAG: Multi-Agent Filtering Retrieval-Augmented Generation, arXiv preprint arXiv:2501.00332, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    W. Liu, J. Chen, K. Ji, L. Zhou, W. Chen, B. Wang, RAG-Instruct: Boosting LLMs with Diverse Retrieval-Augmented Instructions, arXiv preprint arXiv:2501.00353, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    C. Mai, Y. Wang, Z. Gong, H. Wang, Y. Huang, KnowRA: Knowledge Retrieval Augmented Method for Document-level Relation Extraction with Comprehensive Reasoning Abilities, Proceedings of the International Joint Conference on Artificial Intelligence, 2025

    2025

  13. [13]

    J. Xu, H. Fei, H. Zhou, X. Quan, Q. Huang, S. Wu, W. Y. Wang, M.-L. Lee, W. Hsu, LogicReward: Incentivising LLM Reasoning via Step-Wise Logical Supervision, International Conference on Learning Representations, 2026

    2026

  14. [14]

    J. C. Aker, Dial “A” for agriculture: a review of information and communication technologies for agricultural extension in developing countries, Agricultural Economics, 42(6), 631-647, 2011

    2011

  15. [15]

    Fabregas, M

    R. Fabregas, M. Kremer, J. Zahl, Realising the potential of digital agricultural extension, Science, 366(6471), eaay3058, 2019

    2019

  16. [16]

    S. A. Cole, A. N. Fernando, Mobile agricultural extension and farmer behaviour at scale, American Economic Journal: Applied Economics, 12(4), 1-29, 2020

    2020

  17. [17]

    Kamilaris, F

    A. Kamilaris, F. X. Prenafeta-Boldú, Deep learning in agriculture: A survey, Computers and Electronics in Agriculture, 147, 70-90, 2018

    2018

  18. [18]

    Abade, P

    A. Abade, P. A. Ferreira, F. de Barros Vidal, Plant diseases recognition on images using convolutional neural networks: A systematic review, Computers and Electronics in Agriculture 185 (2021) 106125. 23

    2021

  19. [19]

    E. T. Lwoga, C. Stilwell, P. Ngulube, Access and use of agricultural information and knowledge in Tanzania, Library Hi Tech, 29(2), 272-295, 2011

    2011

  20. [20]

    D. J. S. Ravindran, I. Skarga-Bandurova, S. V, M. Awais, M. S, AgroLLM: Connecting Farmers and Agricultural Practices through Large Language Models for Enhanced Knowledge Transfer and Practical Application, AgriEngineering 8 (2026) 130-145

    2026

  21. [21]

    L. Liu, S. Li, J. Qi, Z. Yuan, P. Yang, TARAG: Time-aware retrieval-augmented generation for precision crop pest and disease management, Computers and Electronics in Agriculture, 2026

    2026

  22. [22]

    L. C. Marques Junior, J. R. R. Manesco, J. P. Papa, AgriRAG: Training-Free Retrieval-Augmented Generation for Agricultural Disease Diagnosis with Vision-Language Models, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026

    2026

  23. [23]

    Singh, J

    N. Singh, J. Wang’ombe, N. Okanga, T. Zelenska, J. Repishti, G. K. Jayasankar, S. Mishra, R. Manokaran, V. Singh, M. I. Rafiq, R. Gandhi, A. Nambi, Farmer.Chat: Scaling AI-Powered Agricultural Services for Smallholder Farmers, arXiv preprint arXiv:2409.08916, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    A. S. Patne, S. S. Ranjane, O. Walke, R. Owhal, AgriSaathi: A Comprehensive Survey on Large Language Models and Multimodal AI for Agricultural Advisory Systems, International Journal for Multidisciplinary Research (IJFMR) 7(6) (2025)

    2025

  25. [25]

    S. Yang, Z. Liu, W. Mayer, N. Ding, Y. Wang, Y. Huang, P. Wu, W. Li, L. Li, H.-Y. Zhang, Z. Feng, ShizishanGPT: An Agricultural Large Language Model Integrating Tools and Resources, arXiv preprint arXiv:2409.13537, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Davis, F

    K. Davis, F. Landini, J. van Niekerk, K. Green, Extension and advisory services during and after COVID-19: digital solutions and challenges, Journal of Agricultural Education and Extension, 26(3), 223-228, 2020

    2020

  27. [27]

    Baumüller, The little heads-up: Mobile phone-enabled agricultural advice services in Africa, Development Policy Review, 36(S1), O225-O245, 2018

    H. Baumüller, The little heads-up: Mobile phone-enabled agricultural advice services in Africa, Development Policy Review, 36(S1), O225-O245, 2018

    2018

  28. [28]

    Munthali, S

    N. Munthali, S. P. Katengeza, J. H. Mangisoni, Factors influencing the adoption of digital extension services among smallholder farmers in Malawi, Computers and Electronics in Agriculture, 188, 106312, 2021

    2021

  29. [29]

    J. R. Landis, G. G. Koch, The measurement of observer agreement for categorical data, Biometrics, 33(1), 159–174, 1977. 24

    1977

  30. [30]

    Zhang, J

    H. Zhang, J. Chen, L. Huang, X. Shen, D. Khashabi, Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements, International Conference on Learning Representations, 2025

    2025

  31. [31]

    J. Ji, T. Zhang, L. Zhou, S. Huang, S. Wu, InferenceGuard: Protecting LLMs from Safety Failures at Test-Time, International Conference on Learning Representations, 2025

    2025

  32. [32]

    Y. Jin, L. Gao, J. Han, X. Wan, Case-Augmented Deliberative Alignment: Dynamic Inference-Time Policy Grounding for Large Language Models, Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2026

    2026

  33. [33]

    Agribusiness

    Jackrong, Qwen3.5-9B-DeepSeek-V4-Flash, HuggingFace Model Hub, https://huggingface.co/ Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash, 2026. 25 Appendix A. Full Text of the ESO Questionnaire The complete set of questions and statements from the digital survey template is detailed below. Appendix A.1. Section 1: Informed Consent • Do you agree to participate?Yes, I...

    2026

  34. [34]

    Pest detection output: {pest_species} (density: {pest_density}) 29

  35. [35]

    Crop and growth stage: {crop_type} (stage: {growth_stage})

  36. [36]

    Local weather and region: {weather_conditions}, {region_district}

  37. [37]

    You MUST adhere strictly to the active ingredients, dosages, and safety precautions specified in the retrieved policy documents

    Retrieved policy documents: {policy_context} Write a clear, structured management advice draft. You MUST adhere strictly to the active ingredients, dosages, and safety precautions specified in the retrieved policy documents. Do not recommend chemicals not explicitly listed. [User Input] Draft suggestions for: - Pest: Fall Armyworm (3 larvae per plant) - C...

  38. [38]

    All chemical pesticides recommended in the draft are registered in the {ghana_epa_registry}

  39. [39]

    The dosage recommended is within the safe thresholds for the specific crop growth stage

  40. [40]

    status":

    Any regional chemical bans (e.g., restricted active ingredients in water-catchment districts) are respected. If any safety rules are violated, output a structured violation report in JSON format with the violation code (e.g., ERR_PST_UNAPPROVED_CROP) and rewrite instructions. If the draft is safe, output {"status": "APPROVED"}. [User Input] Audit the foll...