PaperlayerView exact revision on arXiv

Source-linked AI summary

Agri-SAGE: Simulation-Grounded Multi-Agent LLM for Context-Aware Agricultural Advisory Generation

Vedant Balasubramaniam, Geetha Charan, Manojkumar Patil, Rohit P Suresh, V Priyanka, Kodur Sai Vinay Sathvik, Y. Narahari

arXiv:2607.00454v1cs.AIcs.MA

TL;DR

Static agronomic guidance cannot adapt to in-season variability, and LLM advisory systems may generate plausible but incorrect recommendations. Agri-SAGE combines retrieval-grounded multi-agent reasoning with APSIM simulation; in a 10-year retrospective evaluation, its reasoning methods outperform static PoP, with Tree of Thoughts yielding the highest improvements and Reflexion using fewer simulator interactions.

  • Problem

    Static PoPs cannot adapt to seasonal and local conditions, while LLM-based advisory systems may generate plausible but incorrect recommendations.

    Source: §I · p. 1

  • Method

    Agri-SAGE combines retrieval-grounded multi-agent LLM reasoning with APSIM simulation to generate and validate context-aware advisories.

    Source: §I · p. 1, A. The Retrieval Agent · p. 2

  • Results

    Across the 10-year retrospective evaluation, all reasoning methods outperform static PoP, Tree of Thoughts achieves the highest yield improvements, and Reflexion achieves competitive yields with fewer simulator interactions.

    Source: §VI · pp. 5–6

  • Takeaways & Limitations

    The evaluation supports simulation-grounded adaptive advisory generation as an approach that improves on static agronomic schedules across the tested climatic conditions.

    Source: §VI · pp. 5–6

Abstract

from arXiv · show

Agricultural advisory systems face a fundamental tension: static agronomic guidelines offer consistent, evidence-based recommendations, yet remain blind to in-season variability and dynamic uncertainties. Recent advisory systems powered by LLMs are liable for a different risk of generating recommendations that are agronomically credible but physiologically unconvincing. Agri-SAGE is a closed-loop framework designed to resolve the above two limitations by integrating retrieval-grounded multi-agent LLM reasoning with APSIM-based biophysical simulation, to generate and validate agronomic advisories. To assess this framework, we evaluate three reasoning approaches, namely Plan-and-Solve, Tree of Thoughts, and Reflexion, over a 10-year retrospective analysis. All three significantly outperform static PoP (Package-of-Practice) baselines, with Tree of Thoughts achieving impressive peak yields. At the same time, Reflexion achieves comparable agronomic outcomes at substantially lower computational cost by leveraging cross-seasonal episodic memory.

I. INTRODUCTION

Static Packages of Practices cannot adapt to in-season and local variability, while LLM-based systems may produce plausible but incorrect advice. Agri-SAGE combines retrieval-grounded multi-agent reasoning with APSIM simulation, and its tested reasoning approaches outperform static PoP baselines.

I. INTRODUCTION

  • Static PoPs are issued before the season and updated infrequently, limiting their ability to adapt to weather shocks, pest outbreaks, and local soil heterogeneity.The resulting recommendations may not suit specific field conditions.Source: §I · p. 1
  • LLM advisory systems can improve access to agronomic knowledge but may generate plausible, incorrect recommendations.Source: §I · p. 1
  • Agri-SAGE reframes the Package of Practices as a dynamic, context-aware plan generated by specialized agents using retrieved agronomic knowledge and real-time environmental inputs.The framework is designed to produce season-long management strategies.Source: §I · p. 1
  • APSIM simulation validates generated plans against crop physiology, addressing the need for recommendations to remain biophysically feasible.Source: §I · p. 1
  • All three reasoning approaches significantly outperform static PoP baselines in the 10-year comparison, with Tree of Thoughts achieving the highest yield.The comparison evaluates Plan-and-Solve, Tree of Thoughts, and Reflexion.Source: §I · p. 1
  • In Mandya simulations, optimized fertilizer and irrigation splits and other operations produced better outcomes than static PoP practices.Source: §I · p. 1

II. RELATED WORK

Prior LLM advisory systems use retrieval and domain-specific methods to improve factual grounding, but their knowledge remains bounded by available corpora. Related work also identifies reliability challenges and explores specialized-agent approaches.

A. LLM-Based Agricultural Advisory Systems

  • Farmer.Chat [2], ShizishanGPT, AgriGPT [4], and AgriRegion [5] use curated knowledge, weather APIs, knowledge graphs, region-aware retrieval, or domain-specific fine-tuning for agricultural advice.Source: A. LLM-Based Agricultural Advisory… · p. 2
  • Structured retrieval pipelines improve factual grounding and reduce hallucinations [6], but corpus limits prevent reasoning about agronomic scenarios absent from retrieved documents.Source: A. LLM-Based Agricultural Advisory… · p. 2
  • Prior evaluations report that GPT pest-management performance improves with explicit domain context [7], while LLMs can give confident but incorrect advice outside their training distribution; AgroAskAI uses specialized agents and reviewers [9].Source: A. LLM-Based Agricultural Advisory… · p. 2

B. Simulation Models and Multi-Agent Systems

Prior work combines language models and simulation in several settings, while Agri-SAGE targets the joint gap of physiological validation and complete season-long plans. Its methodology organizes the framework around retrieval, generation, and verification/feedback agents.

B. Simulation Models and Multi-Agent Systems

III. METHODOLOGY

  • The framework has Retrieval, Generation, and Verification and Feedback modules.Source: §III · p. 2

A. The Retrieval Agent

The Retrieval Agent grounds advice in regional agronomic materials and environmental context. The Generation Agent uses APSIM feedback with three reasoning approaches to iteratively formulate or refine agricultural actions.

A. The Retrieval Agent

  • Regional university PoPs ground recommendations in locally accessible and affordable crop inputs, fertilizer types, and operational methods.Source: A. The Retrieval Agent · p. 2
  • The Retrieval Agent combines hierarchical chunking and hybrid dense–sparse retrieval, using weather, soil, phenology, and solar-radiation inputs to find relevant agronomic literature.Source: A. The Retrieval Agent · p. 2

B. The Generation Agent

  • The Generation Agent combines retrieved local context with real-time environmental state to produce detailed, actionable advisories.Source: B. The Generation Agent · p. 2
  • The task is framed as iterative optimization: APSIM supplies deterministic observations, and the LLM generates actions from those observations using Plan-and-Solve, Tree of Thoughts, or Reflexion.Source: B. The Generation Agent · p. 2
  • Plan-and-Solve diagnoses biophysical failures in the previous action, revises its plan, and executes a new action, using APSIM feedback and baseline yields for self-correction.Source: B. The Generation Agent · p. 2
  • Tree of Thoughts generates three mutually exclusive intervention paths, critiques their biophysical pros and cons against the weather forecast, and sends its highest-scoring path to APSIM.Source: B. The Generation Agent · p. 3
  • Reflexion stores distilled heuristics from prior simulated seasons in a persistent textual memory and conditions later-year policies on that history.After each season, the agent reflects on macro-level success or failure and adds the reflection to memory.Source: B. The Generation Agent · p. 3

C. The Verification and Feedback Agent

The Verification and Feedback Agent translates generated advice into executable APSIM inputs, then uses crop-simulation metrics to support iterative critique and refinement.

  • The translation layer converts generated recommendations into executable APSIM calls for sowing, fertilizer, irrigation, mulch, and tillage operations.Source: C. The Verification and Feedback Agent · p. 3
  • APSIM simulates the full crop life cycle, and the Feedback Agent extracts grain yield, water stress, and nitrogen stress as feedback metrics.Grain yield is the optimization objective; water and nitrogen stress correspond to irrigation and fertilizer decisions.Source: C. The Verification and Feedback Agent · p. 3
  • By exposing physical drivers of crop outcomes, such as nitrogen leaching after premature application before heavy rainfall, the LLM can critique and refine prior advisories.Source: C. The Verification and Feedback Agent · p. 3

IV. DATASETS AND EXPERIMENTAL SETUP

The evaluation uses maize simulations for Mandya, Karnataka, across a decade of variable weather, with a single-crop, locally parameterized soil setup.

A. Weather, Region, and Crop Selection

B. Agronomic Knowledge Base

The setup combines a maize knowledge corpus with a fixed regional PoP benchmark and repeated simulations across feedback depths.

B. Agronomic Knowledge Base

  • The maize knowledge base combines about 1,000 open-access research papers with official Karnataka extension materials on sowing windows and cultivar selection.Source: B. Agronomic Knowledge Base · p. 4

C. Baseline Formulation

  • The regional Gold Standard PoP is encoded as fixed APSIM sowing dates and fertilizer splits, simulated across 10 years to provide annual yield thresholds.Source: C. Baseline Formulation · p. 4

D. Experimental Setup

  • Each configuration was run 10 times, and feedback depths N ∈ {0, 2, 3, 4} were tested, with four loops as the cap because agents reportedly plateau within three to four.The 0-loop setting is a zero-shot baseline using retrieval context alone; metrics are reported as run means.Source: D. Experimental Setup · p. 4

V. EXPERIMENTAL RESULTS

Across the 10-year comparison, all three reasoning approaches beat the static baseline, with Tree of Thoughts achieving the highest average yield.

A. Quantitative Analysis

  • Plan-and-Solve averaged 9,045 kg/ha (+935 kg/ha) and Reflexion 9,002 kg/ha (+892 kg/ha), reaching similar yield ceilings.Plan-and-Solve uses critique and revision; Reflexion accumulates heuristic rules across simulated growing seasons.Source: A. Quantitative Analysis · p. 4
  • 9,262 kg/ha (+1,152 kg/ha over the baseline) was Tree of Thoughts’ 10-year average maximum yield, highest among the methods.Figure 3 presents yield performance for Plan-and-Solve, Reflexion, and Tree of Thoughts versus baseline over 10 years, averaged over 10 runs.Source: A. Quantitative Analysis · p. 4, Figure 3 · p. 4
  • During the 2019 drought and 2024 floods, Tree of Thoughts outperformed the baseline by 1,384 kg/ha and 1,375 kg/ha, respectively.Its three-pathway lookahead critiques alternative interventions against the current weather forecast before simulation.Source: A. Quantitative Analysis · p. 5

B. Convergence Dynamics

Feedback iterations improve yields most early on, but gains flatten after three to four loops. ToT and Reflexion often converge faster than Plan-and-Solve, while ToT uses more tokens per decision stage.

B. Convergence Dynamics

  • ToT often reaches optimal yield in iteration 1, while Reflexion’s episodic memory enables faster convergence and Plan-and-Solve gains flatten after 3 to 4 feedback loops.Forcing these agents beyond 3 iterations yields diminishing returns.Source: B. Convergence Dynamics · p. 5
  • ToT consumes k times more tokens per decision stage because it explores k parallel paths, whereas Reflexion and Plan-and-Solve refine a single path.Source: B. Convergence Dynamics · p. 5

C. Qualitative Analysis

  • The 2019 simulation’s severe late-season drought provides the case study for comparing how the architectures handle extreme weather.Source: C. Qualitative Analysis · p. 5

1) Baseline:

In the 2019 drought, the fixed mid-June PoP sowing date exposed maize to severe reproductive-stage water stress. Plan-and-Solve recovered through simulator feedback, while Tree of Thoughts selected earlier sowing and achieved the strongest reported yield.

1) Baseline:

  • 7,022 kg/ha followed the PoP’s hard-coded mid-June sowing date, which placed maize’s sensitive reproductive stage at the peak of the late-season drought.Source: 1) Baseline · p. 5

2) Plan-and-Solve (PS):

  • 8,168 kg/ha was reached by Plan-and-Solve after APSIM feedback prompted surface mulch and a switch to easily dissolving nitrate fertilizer.Its first attempt followed standard timing and suffered the baseline’s drought penalties.Source: 2) Plan-and-Solve (PS) · p. 5

3) Tree of Thoughts (ToT):

  • 8,825 kg/ha was the best 2019 result, achieved by ToT after comparing sowing strategies and shifting planting to May 15 to finish before extreme heat.ToT found the solution in one attempt; the PoP yield was 7,022 kg/ha.Source: 3) Tree of Thoughts (ToT) · p. 5, 1) Baseline · p. 5
  • Reflexion’s iterations fell to one in 2017–2018 after memory-building in 2015–2016, then rose to four during the 2019–2020 drought anomaly.Source: 3) Tree of Thoughts (ToT) · p. 5

4) Reflexion:

Reflexion uses summaries of past seasons to anticipate drought and can achieve competitive yields with fewer simulator interactions. The paper’s broader evaluation supports adaptive, simulation-grounded advisories, while field validation and objectives beyond yield remain open.

4) Reflexion:

  • Reflexion achieved 7,988 kg/ha on its first 2019 attempt by using season memory to apply organic manure and schedule deep tillage ahead of drought.The authors describe heavy manure mixed deep into soil as increasing water-holding capacity; later-season first-iteration yields increased, with minimal subsequent-loop gains.Source: 4) Reflexion · p. 5

VI. CONCLUSION AND FUTURE WORK

  • Across the 10-year evaluation, all reasoning-based methods improved yield over static schedules; ToT led yields, while Reflexion delivered competitive yields with fewer simulator interactions.Source: §VI · pp. 5–6
  • Field validation remains necessary because APSIM predictions are a proxy for agronomic outcomes amid real-world soil and microclimate variability.Future work also includes costs, profit, sustainability metrics, real-time forecasts, and fine-tuned models for novel climate scenarios.Source: §VI · p. 6

Related papers

Loading 2607.00454v1…