Paper deep dive

AgriWorld:A World Tools Protocol Framework for Verifiable Agricultural Reasoning with Code-Executing LLM Agents

Zhixing Zhang, Jesen Zhang, Hao Liu, Qinhan Lv, Jing Yang, Kaitong Cai, Keze Wang

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 52

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 7/21/2026, 4:03:23 AM

Summary

The paper introduces AgriWorld, an executable agricultural environment and a World-Tools-Protocol framework designed to enable verifiable reasoning for large language models (LLMs) in agriculture. It addresses the gap between LLMs' text capabilities and the need for high-dimensional spatiotemporal data analysis by providing unified tools for geospatial queries, remote sensing analytics, and crop simulation. The framework employs an LLM agent, Agro-Reflective, which uses an execute-observe-refine loop to self-correct errors and produce auditable, code-grounded results. The authors also present AgroBench, a verifiable evaluation suite for agricultural QA tasks.

Entities (7)

AgriWorld · software-framework · 98%Agro-Reflective · ai-agent · 97%AgroBench · datasetbenchmark · 96%World-Tools-Protocol · methodology · 94%Execute-Observe-Refine · process · 92%Parcels · data-entity · 90%Remote-Sensing Time Series · data-entity · 90%

Relation Signals (6)

Agro-Reflective → runson → AgriWorld

confidence 96% · On top of this environment [AgriWorld], we design a multi-turn LLM agent, Agro-Reflective...

AgriWorld → implements → World-Tools-Protocol

confidence 95% · Our design follows a World–Tools–Protocol abstraction. World provides a Python execution environment, AgriWorld...

Agro-Reflective → usesmethod → Execute-Observe-Refine

confidence 94% · Agro-Reflective uses an execute-observe-refine loop to self-correct errors...

AgroBench → evaluates → Agro-Reflective

confidence 93% · We introduce AgroBench... Experiments outperform text-only and direct tool-use baselines, validating execution-driven reflection...

AgriWorld → supports → Parcels

confidence 90% · AgriWorld abstracts this complexity into a unified state space S composed of five typed entities: Parcels...

AgriWorld → supports → Remote-Sensing Time Series

confidence 90% · AgriWorld abstracts this complexity into a unified state space S composed of five typed entities: Raster Time Series...

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Foundation models for agriculture are increasingly trained on massive spatiotemporal data (e.g., multi-spectral remote sensing, soil grids, and field-level management logs) and achieve strong performance on forecasting and monitoring. However, these models lack language-based reasoning and interactive capabilities, limiting their usefulness in real-world agronomic workflows. Meanwhile, large language models (LLMs) excel at interpreting and generating text, but cannot directly reason over high-dimensional, heterogeneous agricultural datasets. We bridge this gap with an agentic framework for agricultural science. It provides a Python execution environment, AgriWorld, exposing unified tools for geospatial queries over field parcels, remote-sensing time-series analytics, crop growth simulation, and task-specific predictors (e.g., yield, stress, and disease risk). On top of this environment, we design a multi-turn LLM agent, Agro-Reflective, that iteratively writes code, observes execution results, and refines its analysis via an execute-observe-refine loop. We introduce AgroBench, with scalable data generation for diverse agricultural QA spanning lookups, forecasting, anomaly detection, and counterfactual "what-if" analysis. Experiments outperform text-only and direct tool-use baselines, validating execution-driven reflection for reliable agricultural reasoning.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

50,452 characters extracted from source content.

Expand or collapse full text

AgriWorld: A World–Tools–Protocol Framework for Verifiable Agricultural Reasoning with Code-Executing LLM Agents Zhixing Zhang Equal contribution. jesen Zhang11footnotemark: 1 Hao Liu Qinhan Lv Jing Yang Kaitong Cai Keze Wang Sun Yat-sen University Corresponding author: kezewang@gmail.com Abstract Foundation models for agriculture are increasingly trained on massive spatiotemporal data (e.g., multi-spectral remote sensing, soil grids, and field-level management logs) and achieve strong performance on forecasting and monitoring. However, these models lack language-based reasoning and interactive capabilities, limiting their usefulness in real-world agronomic workflows. Meanwhile, large language models (LLMs) excel at interpreting and generating text, but cannot directly reason over high-dimensional, heterogeneous agricultural datasets. We bridge this gap with an agentic framework for agricultural science. It provides a Python execution environment, AgriWorld, exposing unified tools for geospatial queries over field parcels, remote-sensing time-series analytics, crop growth simulation, and task-specific predictors (e.g., yield, stress, and disease risk). On top of this environment, we design a multi-turn LLM agent, Agro-Reflective, that iteratively writes code, observes execution results, and refines its analysis via an execute–observe–refine loop. We introduce AgroBench, with scalable data generation for diverse agricultural QA spanning lookups, forecasting, anomaly detection, and counterfactual “what-if” analysis. Experiments outperform text-only and direct tool-use baselines, validating execution-driven reflection for reliable agricultural reasoning. 1 Introduction Large language models (LLMs) are increasingly used as general-purpose assistants for scientific work: they can follow complex instructions, explain concepts, summarize documents, and generate code that automates parts of an analysis Kaddour et al. (2023). Yet, in many scientific domains, answering even a seemingly simple question requires operating over high-dimensional, heterogeneous, and spatiotemporal numerical data—not purely textual contextHu et al. (2025); Zhang et al. (2026b). This mismatch has motivated tool-augmented and code-executing agents that connect LLMs to external programs and datasets, enabling conclusions to be grounded in computation and verified through intermediate results rather than surface-form text Schick et al. (2023a); Wang et al. (2023); Zhang et al. (2025b, a, i, f). Agriculture is a particularly compelling and challenging setting for executable scientific assistants Liakos et al. (2018). Modern agronomic decision making relies on diverse data products, including multi-spectral remote sensing time series, soil and terrain grids, field-parcel boundaries, field-level management logs, and local weather streamsZhang et al. (2002, 2025c, 2025j). Practical questions are inherently spatiotemporal and unit-sensitive: they require parcel–pixel–grid alignment, stage-specific temporal windows, and careful handling of measurement scalesCai et al. (2023); Khanna and Kar (2019); Zhang et al. (2026c, d). Meanwhile, foundation models trained on large agronomic corpora can be effective for narrow predictive objectives (e.g., forecasting and monitoring)Sawyer et al. (2024); Zhang et al. (2025g), but they often lack language-facing interfaces for interactive analysis, “what-if” exploration, and transparent, inspectable reasoning stepsKamilaris and Prenafeta-Boldú (2018). Conversely, LLMs excel at natural-language interaction but cannot natively manipulate structured agronomic datasets, and thus may produce answers that are plausible yet difficult to verify Thoppilan et al. (2022); Zhang et al. (2026a); Yao et al. (2024) . The gap becomes evident in everyday workflows. Consider a query such as: “Which parcels are likely to experience water stress next week, and what irrigation adjustment would mitigate the risk?” A reliable solution requires more than recalling facts. It must (i) perform geospatial filtering and joins across entities (fields, administrative regions, sensor footprints)Mishra et al. (2020); Zhang et al. (2025d); Li et al. (2021), (i) compute and visualize remote-sensing indicators over time with correct aggregation windowsGhorbanian et al. (2022), (i) incorporate soil properties and exogenous drivers such as weatherCampbell and Norman (1998); Wang et al. (2020), and (iv) run a simulator or predictor to evaluate counterfactual interventionsJin et al. (2018b, a); Zhang et al. (2025e). In practice, agronomists orchestrate an ecosystem of tools—GIS operations, time-series analytics, crop growth simulators, and task-specific modelsKooistra et al. (2024) —and translate computational artifacts into actionable recommendationsKamilaris and Prenafeta-Boldú (2018); Sorensen et al. (2010). Without a programmatic interface and an execution loop, an LLM has no principled way to check coordinate systems, validate intermediate computations, or correct subtle errors in spatial joins, temporal windows, and unit conversions, making purely text-based responses brittle and hard to trust for scientific useQiao et al. (2023a); Mialon et al. (2023). We propose an agentic framework for agricultural science that makes LLM-based analysis executable, auditable, and reproducible. Our design follows a World–Tools–Protocol abstraction. World provides a Python execution environment, AgriWorld, which exposes unified APIs for core agricultural operations: (1) geospatial querying over field parcels and administrative regions, (2) remote-sensing time-series analytics and anomaly statistics, (3) crop growth simulation with support for counterfactual interventions, and (4) task-specific predictors such as yield, stress, and disease risk. Tools in AgriWorld return inspectable intermediate artifacts (e.g., tables, plots, and derived masks), allowing analyses to be grounded in concrete computations and audited step-by-step. Protocol standardizes task specification and evaluation through executable references and checkers whenever the underlying question admits programmatic validation, enabling scalable and reproducible comparisons. On top of AgriWorld, we introduce Agro-Reflective, a multi-turn LLM agent that alternates between writing code, executing it, and refining its analysis based on execution feedbackQiao et al. (2023b). This execution-driven reflection is crucial in agriculture, where minor mistakes in coordinate systems, aggregation windows, or units can invalidate the final conclusionShinn et al. (2023); Wang et al. (2023). Rather than treating tool calls as one-shot add-ons to generation, Agro-Reflective uses intermediate artifacts as first-class signals to diagnose errors, revise hypotheses, and converge to a checkable solution. To evaluate such agents systematically, we build a verifiable evaluation suite on top of an existing agricultural benchmark by converting questions into tool-grounded instances with deterministic reference code and executable checker functions. This suite covers a broad spectrum of capabilities: basic lookups (indices, phenology, and resource use), forecasting and monitoring, anomaly detection, and counterfactual intervention analysis. Importantly, the executable design enables fine-grained diagnosis of failure modes (e.g., spatial misalignment, incorrect temporal windows, unit errors, and ungrounded claims), beyond coarse end-answer accuracy Chen et al. (2021); Liu et al. (2023). The main contributions are three-fold: • AgriWorld: an executable agricultural environment. We introduce a Python environment that unifies geospatial querying, remote-sensing time-series analytics, crop simulation, and task-specific predictors through consistent APIs, enabling grounded and auditable agronomic analyses. • Agro-Reflective: an execution-driven reflective agent. We propose a multi-turn agent that iterates execute–observe–refine, using intermediate artifacts to self-correct and produce tool-grounded analysis traces rather than one-shot text. • A verifiable evaluation suite with executable checkers. Built on top of an existing benchmark, we provide a scalable task generation and scoring pipeline with deterministic reference programs and executable checkers spanning lookups, forecasting, anomaly detection, and counterfactual interventions, enabling reproducible comparisons against text-only and direct tool-use baselines. 2 Related Work Tool-augmented LLM agents and executable reasoning. A growing line of work augments LLMs with external tools (APIs, code interpreters, and symbolic modules) Karpas et al. (2022); Zhang et al. (2025k, h)to ground answers in computation rather than purely textual generationSchick et al. (2023b); Song et al. (2023); Ma et al. (2024, 2025e, 2025d, 2025a, 2025b). Such systems typically couple a planner with tool invocation, enabling multi-step problem solving over structured resources and producing intermediate results that can be inspected. Our work aligns with this direction but targets a domain where correctness hinges on spatiotemporal alignment, unit consistency, and reproducible execution traces. We thus emphasize (i) a domain-grounded environment that exposes typed operations and auditable artifacts, and (i) a protocol that makes evaluation executable whenever possible, beyond text-only answer matching. Scientific and data-centric agents for domain workflows. Recent “scientific assistants” and “data analysis agents” aim to automate parts of research workflows by writing code, operating over datasets, and iteratively refining hypotheses using execution feedbackBran et al. (2023); Ghafarollahi and Buehler (2024); Lála et al. (2023); Ma et al. (2025c); Long et al. (2025). Parallel efforts in geospatial analytics, remote sensing pipelines, and decision-support systems have developed mature toolchains for spatial joins, raster processing, and time-series modeling, but these pipelines are typically engineered for experts and are not language-interactive18; 52; 48; 10. We bridge these threads by providing an agricultural world model (AgriWorld) that unifies geospatial, remote-sensing, weather, and management data with simulation/prediction backends, and by designing an execution-driven agent (Agro-Reflective) that treats debugging signals (e.g., CRS mismatches, window drift) as first-class feedback for self-correction. Benchmarks and verifiable evaluation for agentic systems. Evaluating agents remains challenging because free-form responses can be plausible yet incorrectDurante et al. (2024), and failures often stem from latent mistakes in intermediate steps. Existing benchmarks commonly score final textual answers, sometimes with reference solutions, but often without executable checkers that can validate constraints, tolerances, and counterfactual claimsAustin et al. (2021). Complementary work has proposed programmatic evaluation, unit tests, or environment-based scoring to enable reproducible assessment and failure diagnosisDrummond (2009). Our protocol follows this verifiable-evaluation philosophy by compiling questions into tool-grounded instances with deterministic reference programs and executable checker functions, enabling scalable comparison across agent variants and fine-grained attribution of errors (spatial misalignment, temporal window errors, unit mismatches, and ungrounded claims). 3 Method 3.1 The World–Tools–Protocol Framework Current LLMs lack the intrinsic capability to process high-dimensional spatiotemporal dataCui et al. (2023). To bridge this gap, we formalize an agricultural assistant not merely as a chatbot, but as an agent A operating within a rigorous World–Tools–Protocol framework (see Figure 1). This framework consists of an executable environment ℰE (AgriWorld), a set of grounded functional tools ℱF, and a verifiable specification protocol V. Figure 1: Overview of the proposed Framework. The pipeline begins with a heterogeneous State Space S (e.g., geographical locations, wind fields). Data flows through functional modules such as Spatial Query, Spatio-temporal Analysis, and Crop Simulation. A critical Coordinate Alignment step transforms the raw Pre-training Sequence into a unified Aligned Sequence (g→kg→ kg). Finally, the Agent validates the timeline for Temporal Errors and enforces constraints like Pattern Validity and Numerical Conservation before producing the Final Output. As illustrated in Figure 1, the workflow initiates from the State Space S, which encapsulates diverse agricultural data types including Geographical Location and Wind Field Representation. These inputs are processed through specialized tools for Spatial Query, Spatio-temporal Analysis, Crop Simulation, and Task Prediction. To ensure consistency across heterogeneous sources, the system employs a Coordinate Alignment mechanism that converts the initial Pre-training Sequence into a standardized Aligned Sequence (denoted as g→kg→ kg). Crucially, the reasoning process is not open-ended; an Agent actively monitors the execution timeline to detect and Fix potential Temporal Errors. The agent enforces strict verification protocols, checking for Pattern Validity, Numerical Conservation, and Semantic Consistency to guarantee that the Final Output is physically and logically sound. Task instance definition. A task instance is formally defined as a tuple ℐ=⟨q,ℬ,out,,B⟩I= q,\ B,\ S_out,\ V,\ B . Here, q represents the natural-language query. ℬB denotes the bindings to concrete data handles (e.g., specific parcel IDs idP_id, time intervals [tstart,tend][t_start,t_end], and sensor products). outS_out specifies the rigid output schema required for downstream applications. V is an executable checker function, and B is the computational budget (max steps). Validity maximization. The agent interacts with the environment to generate a final answer y y and a set of auditable artifacts O. The protocol evaluates the validity of the solution: (z,d)=(y^,∣ℐ),z∈0,1,(z,d)=V( y,O ), z∈\0,1\, (1) where z is a binary correctness indicator and d is a structured diagnostic report detailing specific constraint violations (if any). The agent’s objective is to find a policy π that maximizes the expected validity [z]E[z] subject to the budget constraint B. 3.2 AgriWorld: The Executable Environment 3.2.1 Data Model and State Space Agricultural data is inherently heterogeneous. AgriWorld abstracts this complexity into a unified state space S composed of five typed entities: • Parcels (P): Vector-based entities where each p∈p encapsulates a polygon geometry gpg_p, a coordinate reference system (CRS), and attribute metadata (e.g., crop type). • Raster Time Series (ℛR): 4D tensors representing remote sensing data. A raster r∈ℛr is defined by a grid mapping function M:ℝ2×T→ℝkM:R^2× T ^k, carrying spectral bands and a validity mask (e.g., cloud cover). • Grid Fields (G): Static spatial layers (e.g., soil texture, DEM) aligned to a canonical spatial grid. • Weather Streams (W): Time-series of exogenous drivers, formalized as w=(t,t)t=1Tw=\(t,v_t)\_t=1^T, where t∈ℝdv_t ^d includes variables like precipitation and radiation. • Management Logs (ℳM): Discrete event sequences m=(ti,ai,qi)m=\(t_i,a_i,q_i)\, recording actions aia_i (e.g., irrigation) with quantities qiq_i. 3.2.2 Unified Tool Interface and Artifacts To ground reasoning in computation, AgriWorld exposes a set of functional tools ℱF. We model a tool execution not as a side-effect-free query, but as a state transition that produces an artifact: (v^,α)=f(sin;θ),f∈ℱ,( v,α)=f(s_in;θ), f , (2) where sin⊆s_in is the input data subset, v v is the return value (e.g., a time-series DataFrame), and α is the Inspectable Artifact. We rigorously define an artifact as α=⟨type,payload,meta,prov⟩α= type, payload, meta, prov . Crucially, meta contains physical constraints (units, CRS, resolution), and prov=H(f,sin,θ) prov=H(f,s_in,θ) is a cryptographic hash of the inputs and tool version, ensuring strict reproducibility of scientific claimsAltintas et al. (2011). Tool Family Illustrative Signatures Geospatial Querying geo.filter_parcels(region, crop, area_range) geo.sjoin(parcels, sensor_footprints) Remote-Sensing Analytics rs.load(index, time_range, region) rs.zonal_stats(raster_ts, parcels, stat=mean,...) rs.anomaly(ts, baseline_window, method) Soil/Terrain Access grid.sample(gridfield, parcels) grid.aggregate(gridfield, region) Weather Variables wx.get(station/region, time_range) wx.rolling_sum/degree_days(...) Simulation & Predictors sim.run(parcels, wx, soil, mgmt, intervention) pred.stress/yield/disease(...) Table 1: The AgriWorld Toolset. Grouping of core APIs by functional domain. The system exposes unified interfaces for heterogeneous spatiotemporal data. 3.2.3 Spatiotemporal Alignment and Unit Safety A core challenge in agricultural AI is the misalignment of heterogeneous data sources. AgriWorld explicitly handles this via a canonical alignment operator. Canonical alignment operator. We define a composite alignment operator Π that maps arbitrary inputs to a unified working space: x~=Π(x)=(resample∘reproject)(x). x= (x)=(T_resample _reproject)(x). (3) Here, reprojectS_reproject transforms vector/raster geometries to a task-specific CRS (e.g., EPSG:3857 for metric calculation), and resampleT_resample aligns temporal data to a canonical frequency (e.g., daily) using strictly defined interpolation policies (nearest/linear)Longley et al. (2015). Robust Zonal Statistics. To handle noise in satellite data (e.g., clouds), we formalize the zonal statistics operation. For a parcel p and a raster r with pixel values rir_i and validity mask mi∈0,1m_i∈\0,1\, the aggregated signal μ and the validity ratio ν are computed as: μ(p,r)=∑i∈Z(p)mi⋅ri∑i∈Z(p)mi,ν(p,r)=∑i∈Z(p)mi|Z(p)|,μ(p,r)= _i∈ Z(p)m_i· r_i _i∈ Z(p)m_i, ν(p,r)= _i∈ Z(p)m_i|Z(p)|, (4) where Z(p)Z(p) is the set of pixels spatially intersecting p. The system raises a LowCoverageError if ν(p,r)<τminν(p,r)< _min, preventing hallucinations based on corrupted data. 3.3 Agro-Reflective: Execution-Driven Agent Most generic agents fail in scientific domains due to "silent errors" Qiao et al. (2023b)(e.g., wrong units). We introduce Agro-Reflective, which utilizes an iterative Execute–Observe–Refine loop driven by structured feedback. Policy and Trace. At step t, the agent acts according to a policy πθ(ct∣mt−1) _θ(c_t m_t-1) conditioned on the memory mt−1m_t-1. The execution yields an observation tuple (ot,t)←ℰ(ct)(o_t,O_t) (c_t). The interaction trace is formalized as a sequence τt=⟨(c1,o1,1),…,(ct,ot,t)⟩ _t= (c_1,o_1,O_1),…,(c_t,o_t,O_t) . Reflection mechanism. We model reflection as a diagnostic function ρ. When a failure (runtime error or protocol violation) occurs, the agent inspects the artifact metadata ℳ(α)M(α) to synthesize a corrective patch: ct+1=ρ(ct,dt,ℳ(αt)).c_t+1=ρ(c_t,d_t,M( _t)). (5) Specifically, ρ categorizes errors into types (e.g., SpatialMisalignment, UnitError) and proposes targeted fixes. The critical importance of this reflection loop, alongside the canonical alignment mechanism (Π ), is empirically validated in our design space exploration (see Table 2), where removing either component significantly degrades performance. Table 2: Ablation and design space exploration of Agro-Reflective. We compare against text-only and direct-execution baselines. ‘Alignment’ denotes the spatiotemporal canonical alignment mechanism. Bold indicates best performance. Method Interaction Tool Scope Alignment Avg. Turns Acc (%) Baselines (1) Text-Only (Kimi/GPT) Chat None – 1.0 21.9 (2) AgriWorld-Direct One-shot All ✓ 1.0 48.2 Agro-Reflective Variants (Ours) (3) w/o Remote Sensing Reflective w/o RS ✓ 2.4 41.5 (4) w/o Alignment Reflective All × 3.1 51.3 (5) w/o Reflection One-shot All ✓ 1.0 50.6 (6) Agro-Reflective (Full) Reflective All ✓ 3.5 57.6 Algorithm 1 Agro-Reflective Execution Loop 0: Task ℐI, Env ℰE, Budget TmaxT_ 1: m0←Encode(ℐ)m_0← Encode(I) 2: for t=1t=1 to TmaxT_ do 3: ct∼πθ(⋅∣mt−1)c_t _θ(· m_t-1) Propose Code 4: (ot,t)←ℰ.exec(ct)(o_t,O_t) .exec(c_t) Execute 5: mt←Update(mt−1,ct,ot)m_t← Update(m_t-1,c_t,o_t) 6: (z,d)←(y^t,1:t∣ℐ)(z,d) ( y_t,O_1:t ) 7: if z=1z=1 then 8: return Finalize(mt)(m_t) 9: end if 10: cpatch←ρ(mt,d)c_patch←ρ(m_t,d) Self-Correction 11: mt←Append(mt,cpatch)m_t← Append(m_t,c_patch) 12: end for 13: return BestEffort(mTmax)(m_T_ ) 3.4 Protocol: Verifiable Specification The protocol V ensures that the agent’s reasoning is grounded. Unlike open-ended chat evaluation, we employ executable checkers with hierarchical constraints: 1. Schema Validity: y y must satisfy the strict JSON schema outS_out (e.g., keys for parcel_id, value, unit). 2. Numeric Tolerance: For quantitative queries, the result must lie within a dynamic tolerance δ of the reference simulator output y∗y^*: |y^−y∗|<δ| y-y^*|<δ. 3. Counterfactual Consistency: For "what-if" analysis (e.g., "Does irrigation I reduce stress?"), we verify the causal direction. Let M(s,a)M(s,a) be the simulator function. The checker asserts: M(s,aintervene)−M(s,abaseline)≤−Δ,M(s,a_intervene)-M(s,a_baseline)≤- , (6) ensuring the agent correctly identified a strategy with significant impact Δ . 4. Physical Sanity: All intermediate artifacts are checked for unit consistency (dimensional homogeneity) and sufficient valid-pixel coverage (ν>τminν> _min). This rigorous protocol transforms vague agricultural advice into verifiable, reproducible computational science. 4 Experiments In this section, we conduct a comprehensive evaluation to verify whether executable, reflective agents can fundamentally improve agricultural reasoning over heterogeneous spatiotemporal data. Specifically, we aim to answer three research questions: • RQ1 (Efficacy): Does tool-grounding and reflective execution outperform state-of-the-art text-only LLMs in domain-specific tasks? • RQ2 (Ablation): What is the individual contribution of each component (e.g., remote sensing tools, alignment protocol, reflection loop) to the overall system performance? • RQ3 (Robustness): Can the proposed agent generalize to unseen geographic regions and temporal windows with extreme weather conditions? 4.1 Experimental Setup 4.1.1 Benchmark and Metrics We utilize AgroBench, a verifiable evaluation suite constructed from real-world agronomic workflows. Unlike standard QA datasets, AgroBench requires agents to interact with the AgriWorld environment to derive answers. We employ a rigorous set of metrics for different task types: 1. Quantitative Accuracy: For lookup tasks, we measure exact matches on retrieved values. 2. Forecasting Error: For time-series prediction (e.g., yield, moisture), we use the Normalized Root Mean Square Error (NRMSE), defined as NRMSE=1N∑(y^i−yi)2y¯NRMSE= 1NΣ( y_i-y_i)^2 y. 3. Spatial IoU: For anomaly detection (e.g., identifying water stress zones), we calculate the Intersection over Union (IoU) between the predicted polygon mask and the ground truth derived from Sentinel-2 satellite imagery. 4. Causal Consistency: For counterfactual analysis, we verify the causal direction using the simulator function M(s,a)M(s,a). A success is recorded only if the agent’s proposed intervention aintervenea_intervene satisfies the physical constraint M(s,aintervene)−M(s,abaseline)≤−ΔM(s,a_intervene)-M(s,a_baseline)≤- , ensuring the decision is physically grounded. 4.1.2 Baselines and Implementation To isolate the impact of our framework, we control for the underlying Large Language Model (LLM). We finetune the Qwen3 series (8B and 32B parameters) using Low-Rank Adaptation (LoRA) Hu et al. (2021); Cai et al. (2025) on the training split. We compare three distinct interaction paradigms: • Text-only Baseline: The model answers queries using internal parametric knowledge only. This serves as a baseline to measure the hallucination rate of pure LLMs. • AgriWorld-Direct (One-shot): A standard “Code Interpreter” approach where the agent generates and executes Python code in a single pass without error handling. • Agro-Reflective (Ours): The proposed agent equipped with the Execute–Observe–Refine loop. We set the maximum interaction budget T=20T=20 steps. The agent receives feedback from standard error streams and our custom validator V. Additionally, we benchmark against strong open-source and proprietary baselines, including Qwen3 Team (2025),DeepSeek-V3 DeepSeek-AI (2024), Yi-1.5-34B 01.AI (2024), LLaMA 3 at Meta (2024),GPT-4o OpenAI (2024), and Gemini-2.0 DeepMind (2024) to assess whether model scale can substitute for tool grounding. 4.2 Main Results and Analysis 4.2.1 Overall Performance Comparison Table 3 presents the main evaluation results, including domain-specific QA performance on Animal, Plant, Aquatic, and Herb. The proposed Qwen3-32B-LoRA-Reflective achieves an aggregate QA score of 7.72/541 and an aggregate choice accuracy of 73.84%, establishing a new state-of-the-art on this benchmark. Beating the Giants with Grounding. A critical finding is that our specialized, grounded agent significantly outperforms general-purpose giant models. For instance, GPT-4o, despite its massive parameter count and RLHF training, only achieves 36.75% accuracy. This is primarily because general-purpose models lack the specific API knowledge of AgriWorld and often hallucinate plausible but incorrect parameter names (e.g., inventing a nonexistent get_soil_moisture() function instead of using the correct grid.sample() API). In contrast, our agent, through tool grounding, strictly adheres to the environment’s schema. The Failure of Text-Only Models. The text-only baselines (Qwen3-32B, Llama3-8B) struggle significantly, with accuracies hovering around 30-40%. This confirms that agricultural reasoning—which involves retrieving dynamic soil data and computing time-series aggregations—is fundamentally beyond the scope of parametric memory. LLMs cannot “memorize” the daily weather of every field parcel in the world; they must retrieve it. Table 3: Main Results on Agricultural Reasoning Benchmarks. We report the aggregated score (Score/Total num Tasks) for domain-specific QA tasks and the percentage accuracy (%) for multiple-choice tasks. Bold denotes the best performance. Model Animal Plant Aquatic Herb QA Total Choice (%) Ours (Agro-Reflective) Qwen3-32B-LoRA-Reflective 7.45/118 7.76/325 7.68/74 8.66/24 7.72/541 73.84 Qwen3-8B-LoRA-Reflective 6.08/118 6.41/325 5.92/74 5.98/24 6.25/541 41.34 Base Models (LoRA Tuned) Qwen3-32B-LoRA 7.21/118 7.45/325 7.28/74 8.46/24 7.42/541 67.25 Qwen3-8B-LoRA 5.97/118 5.71/325 5.82/74 7.04/24 5.84/541 57.55 Open-Source Baselines Qwen3-32B 5.87/118 5.55/325 5.58/74 6.58/24 5.67/541 39.67 DeepSeek-V3 3.36/33 4.31/90 4.08/20 5.21/7 4.17/150 57.43 Yi-1.5-34B-Chat 5.79/33 5.32/90 6.60/20 6.86/7 5.67/150 35.38 Proprietary Models GPT-4o 7.67/33 6.76/90 7.05/20 8.43/7 7.07/150 36.75 Gemini-2.0-Flash 2.88/33 4.02/90 4.30/20 6.07/7 3.90/150 55.33 4.2.2 Fine-grained Task Analysis To better understand where the improvements come from, we analyze the performance breakdown by task type (Table 4). 1. Lookups vs. Reasoning. The gap between baselines and our method is smallest in the Lookup tasks (82.5% vs 86.7%). This is expected, as lookups (e.g., "What is the crop type of parcel X?") require only a single API call. Most “Code Interpreter” models can handle this without complex reasoning. 2. The Challenge of Forecasting. In Forecasting tasks, however, the gap widens significantly (NRMSE 0.34 vs 0.18). Forecasting requires multi-step data processing: fetching weather data, resampling it to a daily frequency, aligning it with crop growth stages, and running a predictive model. The One-shot baseline often fails at the alignment step—for example, mixing up monthly aggregated precipitation with daily temperature—leading to poor predictions. Agro-Reflective detects these dimensional mismatches via the execution feedback loop and corrects the resampling code, resulting in a 47% reduction in error. 3. Causal Reasoning. The most dramatic improvement is seen in Counterfactual Analysis (Success Rate 43.8% vs 71.4%). These questions ask "What if we reduce irrigation by 10%?". To answer this, the agent must run a simulation, observe the outcome, and verify if the change is significant. The Text-only models fail completely here (<15%), as they cannot simulate physical processes. Our reflective agent excels because it treats the simulation as an experimental sandbox, iteratively adjusting inputs until a valid causal inference is drawn. Table 4: Task-Specific Performance Breakdown. Agro-Reflective shows the largest gains in complex multi-step reasoning tasks (Forecasting, Counterfactual) compared to the One-shot baseline. Model Lookup Forecasting Anomaly Counterfactual Acc. NRMSE (↓ ) IoU (↑ ) Success (↑ ) Text-Only 41.2% 0.89 0.12 14.5% One-shot 82.5% 0.34 0.51 43.8% Reflective 86.7% 0.18 0.68 71.4% 4.3 Ablation Study: Dissecting the Agent To verify the necessity of each system component, we conducted an ablation study (referenced in Table 2 of the methodology). Impact of Remote Sensing Tools. When we removed the remote sensing module (w/o RS), the agent lost the ability to "see" the fields. Performance on crop monitoring tasks dropped by over 40%. Text-only descriptions of fields (e.g., "a green field") are insufficient for calculating precise indices like NDVI, proving that pixel-level data access is non-negotiable for agricultural AI. Impact of Canonical Alignment. Removing the canonical alignment operator (w/o Alignment) caused a surge in runtime errors. Without automatic Coordinate Reference System (CRS) transformation, the agent attempted to join GPS coordinates (WGS84) with projected satellite rasters (UTM), resulting in empty intersection sets. This confirms that the spatiotemporal alignment protocol is the "glue" that makes heterogeneous data usable. Impact of Reflection Loop. The comparison between AgriWorld-Direct and Agro-Reflective quantifies the value of self-correction. We observed that 35% of initial code generations contained bugs (e.g., syntax errors, wrong column names). The reflection loop allowed the agent to recover from these errors in 92% of cases, converting a potential failure into a successful retrieval. 4.4 Efficiency Analysis We further investigate the cost-performance trade-off. Figure 2 illustrates the performance scaling with respect to the inference budget T. The curve follows a logarithmic saturation pattern. The steepest performance gain occurs in the first few turns (T=1→4T=1→ 4), where the accuracy jumps from 48.2% to 68.5%. This initial boost corresponds to the agent fixing "low-hanging fruits"—simple syntax errors and API parameter mismatches. As T increases from 4 to 20, the gains become marginal (+5.3%), representing the correction of deeper semantic errors. Importantly, although we allocate a budget of T=20T=20, the average turns required for convergence is only 3.5. This indicates that our verifiable protocol V is highly effective at early stopping: once the agent produces an output that satisfies the schema and physical constraints, the loop terminates, avoiding wasteful computation. 1122448820204040505060607070808048.248.261.361.368.568.572.172.173.873.8Inference Budget T (Max Steps)Aggregate Accuracy (%)Agro-ReflectiveOne-shot Baseline Figure 2: Performance Scaling. Accuracy saturates logarithmically, showing efficient self-correction. 4.5 Robustness to Out-of-Distribution Scenarios Table 5: OOD Robustness Analysis. Comparison of accuracy degradation (Δ ) on unseen regions (Spatial) and years (Temporal). Text models fail to generalize, while tool-grounded agents remain robust. Method IID Spatial OOD Temporal OOD Acc. Acc. Δ Acc. Δ Text-Only 41.2 18.5 -55% 22.1 -46% One-shot 48.2 26.4 -45% 35.8 -26% Agro-Reflective 73.8 64.5 -12% 69.2 -6% Finally, we challenge the agents with Out-of-Distribution (OOD) scenarios to test their generalization capabilities (Table 5). Spatial OOD (Brazil vs. California). We test on fields in Brazil, which were excluded from the LoRA training set. Text-only models see a catastrophic performance drop (Δ=−55% =-55\%), likely because they overfit to the specific crop calendars of the US (e.g., assuming harvest in October, which is incorrect for the Southern Hemisphere). Agro-Reflective, however, maintains robustness (Δ=−12% =-12\%) because it does not guess; it queries the local weather stations and satellite data dynamically. Temporal OOD (The El Niño Year). Testing on data from 2023, a year marked by extreme weather anomalies, reveals that text models relying on historical averages fail to predict the unusual yield losses. Our agent, by accessing real-time sensor streams, captures the signal of the anomaly and adjusts its predictions accordingly (Δ=−6% =-6\%). 5 Conclusion Our study shows that simply scaling large language models is insufficient for agricultural science, where correctness depends on precise spatiotemporal alignment and executable validation. We introduce AgriWorld, a World–Tools–Protocol framework that grounds LLM reasoning in verifiable computation, and Agro-Reflective, an execution-driven agent that iteratively refines its analysis through feedback from intermediate results. Extensive experiments on AgroBench demonstrate that enforcing executable reasoning substantially improves accuracy and robustness over text-only and one-shot tool-use baselines, especially on complex multi-step tasks. This work points toward a practical pathway for building trustworthy, reproducible LLM agents for agriculture and other spatiotemporal scientific domains. References [1] 01.AI (2024) Yi-1.5: open foundation models by 01.ai. arXiv preprint arXiv:2403.04652. Cited by: §4.1.2. [2] I. Altintas, S. P. Callahan, E. Deelman, D. Gannon, J. Myers, and M. A. Parsons (2011) Provenance for scientific workflows: a survey. Conceiving the future internet-From vision to reality, p. 475–487. Cited by: §3.2.2. [3] A. at Meta (2024) The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Cited by: §4.1.2. [4] J. Austin, A. Odena, M. Nye, M. Bosma, H. Michalewski, D. Dohan, E. Jiang, C. Cai, J. Shin, C. Lindsey, et al. (2021) Program synthesis with large language models. arXiv preprint arXiv:2108.07732. Cited by: §2. [5] A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and P. Schwaller (2023) ChemCrow: augmenting large-language models with chemistry tools. External Links: 2304.05376, Link Cited by: §2. [6] K. Cai, J. Zhang, J. Yang, Y. Fan, P. Xie, J. Wang, and K. Wang (2025) FlashVLM: text-guided visual token selection for large multimodal models. External Links: 2512.20561, Link Cited by: §4.1.2. [7] Z. Cai, Q. Hu, X. Zhang, J. Yang, H. Wei, J. Wang, Y. Zeng, G. Yin, W. Li, L. You, B. Xu, and Z. Shi (2023) Improving agricultural field parcel delineation with a dual branch spatiotemporal fusion network by integrating multimodal satellite data. ISPRS Journal of Photogrammetry and Remote Sensing. External Links: Document, Link Cited by: §1. [8] G. S. Campbell and J. M. Norman (1998) Environmental biophysics. Springer Science & Business Media. Cited by: §1. [9] M. Chen, J. Tworek, H. Jun, et al. (2021) Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374. Cited by: §1. [10] Y. Chen, X. He, X. Ma, and Y. Ma (2025) ContextFlow: training-free video object editing via adaptive context enrichment. arXiv preprint arXiv:2509.17818. Cited by: §2. [11] S. Cui, Q. Han, X. Jiang, Q. Zhou, Y. Wei, Z. Yuan, C. Xie, X. Zhang, H. Zhu, and J. Chen (2023) Can chatgpt replace human annotators for data labeling? an experimental study. arXiv preprint arXiv:2303.11162. Cited by: §3.1. [12] G. DeepMind (2024) Gemini 2.0: flash and beyond. Technical Report. Cited by: §4.1.2. [13] DeepSeek-AI (2024) DeepSeek-v3 technical report. arXiv preprint arXiv:2412.19437. Cited by: §4.1.2. [14] C. Drummond (2009) Reproducible research in computational science. Computing in Science & Engineering 11 (5), p. 66–71. Cited by: §2. [15] Z. Durante, Q. Huang, N. Wake, R. Gong, J. S. Park, B. Sarkar, R. Taori, Y. Noda, D. Terzopoulos, Y. Choi, K. Ikeuchi, H. Vo, L. Fei-Fei, and J. Gao (2024) Agent ai: surveying the horizons of multimodal interaction. External Links: 2401.03568, Link Cited by: §2. [16] A. Ghafarollahi and M. J. Buehler (2024) SciAgents: automating scientific discovery through multi-agent intelligent graph reasoning. External Links: 2409.05556, Link Cited by: §2. [17] A. Ghorbanian, M. Rezaei, S. Alipour, S. Talebi, and M. Amani (2022) Remote sensing for water stress detection in agriculture: a review. Remote Sensing Applications: Society and Environment 26, p. 100720. Cited by: §1. [18] N. Gorelick, M. Hancher, M. Dixon, S. Ilyushchenko, D. Thau, and R. Moore (2017) Google earth engine: planetary-scale geospatial analysis for everyone. Remote Sensing of Environment. Cited by: §2. [19] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen (2021) LoRA: low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685. Cited by: §4.1.2. [20] M. Hu, C. Ma, W. Li, W. Xu, J. Wu, J. Hu, T. Li, G. Zhuang, et al. (2025) A survey of scientific large language models: from data foundations to agent frontiers. arXiv preprint arXiv:2508.21148. External Links: Link Cited by: §1. [21] X. Jin, L. Kumar, Z. Li, H. Feng, X. Xu, G. Yang, and J. Wang (2018) A review of data assimilation of remote sensing and crop models. Agricultural and Forest Meteorology. External Links: Document, Link Cited by: §1. [22] Z. Jin, G. Azzari, and D. B. Lobell (2018) A review of data assimilation in crop models. European Journal of Agronomy 92, p. 141–152. External Links: Document Cited by: §1. [23] J. Kaddour, J. Harris, B. Ma, H. Mazumder, X. Guo, Z. Yang, Y. Lin, M. Zhang, L. Liu, X. Sun, et al. (2023) LLMs for science: a survey. arXiv preprint arXiv:2308.11584. Cited by: §1. [24] A. Kamilaris and F. Prenafeta-Boldú (2018) Deep learning in agriculture: a survey. Computers and Electronics in Agriculture 147, p. 70–90. External Links: Document Cited by: §1. [25] E. Karpas, O. Abend, Y. Belinkov, B. Lenz, O. Lieber, N. Ratner, Y. Shoham, H. Bata, Y. Levine, K. Leyton-Brown, D. Muhlgay, N. Rozen, E. Schwartz, G. Shachaf, S. Shalev-Shwartz, A. Shashua, and M. Tenenholtz (2022) MRKL systems: a modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning. External Links: 2205.00445, Link Cited by: §2. [26] S. Khanna and G. Kar (2019) Data fusion in precision agriculture. Journal of the Indian Society of Remote Sensing 47 (7), p. 1127–1143. Cited by: §1. [27] L. Kooistra, K. Berger, B. Brede, L. V. Graf, H. Aasen, J. Roujean, M. Machwitz, M. Schlerf, C. Atzberger, E. Prikaziuk, D. Ganeva, E. Tomelleri, H. Croft, P. Reyes Muñoz, V. Garcia Millan, R. Darvishzadeh, G. Koren, I. Herrmann, O. Rozenstein, S. Belda, M. Rautiainen, S. R. Karlsen, C. F. Silva, S. Cerasoli, J. Pierre, E. Tanır Kayıkçı, A. Halabuk, E. Tunc Gormus, F. Fluit, Z. Cai, M. Kycko, T. Udelhoven, and J. Verrelst (2024) Reviews and syntheses: remotely sensed optical time series for monitoring vegetation productivity. Biogeosciences 21, p. 473–511. External Links: Document, Link Cited by: §1. [28] J. Lála, O. O’Donoghue, A. Shtedritski, S. Cox, S. G. Rodriques, and A. D. White (2023) PaperQA: retrieval-augmented generative agent for scientific research. External Links: 2312.07559, Link Cited by: §2. [29] X. Li, J. Zhang, and F. Safara (2021-03) Improving the accuracy of diabetes diagnosis applications through a hybrid feature selection algorithm. Neural Process. Lett. 55 (1), p. 153–169. External Links: ISSN 1370-4621, Link, Document Cited by: §1. [30] K. G. Liakos, P. Busato, D. Moshou, S. Pearson, and D. Bochtis (2018) Artificial intelligence in agriculture: a review. Sensors 18 (8), p. 2674. External Links: Document, Link Cited by: §1. [31] X. Liu, Z. Zheng, Z. Du, et al. (2023) AgentBench: evaluating llms as agents. arXiv preprint arXiv:2308.03688. Cited by: §1. [32] Z. Long, M. Zheng, K. Feng, X. Zhang, H. Liu, H. Yang, L. Zhang, Q. Chen, and Y. Ma (2025) Follow-your-shape: shape-aware image editing via trajectory-guided region control. arXiv preprint arXiv:2508.08134. Cited by: §2. [33] P. A. Longley, M. F. Goodchild, D. J. Maguire, and D. W. Rhind (2015) Geographical information science. John Wiley & Sons. Cited by: §3.2.3. [34] Y. Ma, K. Feng, Z. Hu, X. Wang, Y. Wang, M. Zheng, X. He, C. Zhu, H. Liu, Y. He, et al. (2025) Controllable video generation: a survey. arXiv preprint arXiv:2507.16869. Cited by: §2. [35] Y. Ma, K. Feng, X. Zhang, H. Liu, D. J. Zhang, J. Xing, Y. Zhang, A. Yang, Z. Wang, and Q. Chen (2025) Follow-your-creation: empowering 4d creation through video inpainting. arXiv preprint arXiv:2506.04590. Cited by: §2. [36] Y. Ma, Y. He, H. Wang, A. Wang, L. Shen, C. Qi, J. Ying, C. Cai, Z. Li, H. Shum, et al. (2025) Follow-your-click: open-domain regional image animation via motion prompts. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, p. 6018–6026. Cited by: §2. [37] Y. Ma, H. Liu, H. Wang, H. Pan, Y. He, J. Yuan, A. Zeng, C. Cai, H. Shum, W. Liu, et al. (2024) Follow-your-emoji: fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, p. 1–12. Cited by: §2. [38] Y. Ma, Y. Liu, Q. Zhu, A. Yang, K. Feng, X. Zhang, Z. Li, S. Han, C. Qi, and Q. Chen (2025) Follow-your-motion: video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207. Cited by: §2. [39] Y. Ma, Z. Yan, H. Liu, H. Wang, H. Pan, Y. He, J. Yuan, A. Zeng, C. Cai, H. Shum, et al. (2025) Follow-your-emoji-faster: towards efficient, fine-controllable, and expressive freestyle portrait animation. arXiv preprint arXiv:2509.16630. Cited by: §2. [40] G. Mialon, E. Dancette, and J. Dessalles (2023) Demystifying chatgpt: a survey of its capabilities, limitations, and ethical considerations. arXiv preprint arXiv:2303.08774. Cited by: §1. [41] D. Mishra, R. Singh, and R. Yadav (2020) Geographic information system (gis) applications in precision agriculture: a review. Current Science 118 (9), p. 1427–1436. Cited by: §1. [42] OpenAI (2024) GPT-4o system card. arXiv preprint arXiv:2405.18455. Cited by: §4.1.2. [43] S. Qiao, H. Gui, C. Lv, Q. Jia, H. Chen, and N. Zhang (2023) Making language models better tool learners with execution feedback. arXiv preprint arXiv:2305.13068. External Links: Link, Document Cited by: §1. [44] S. Qiao, H. Gui, C. Lv, Q. Jia, H. Chen, and N. Zhang (2023) Making language models better tool learners with execution feedback. arXiv preprint arXiv:2305.13068. Cited by: §1, §3.3. [45] D. P. Sawyer, J. R. Hudson, A. Lerchner, D. J. Rezende, T. P. Lillicrap, M. Mozer, and J. X. Wang (2024) Can foundation models actively gather information in interactive environments to test hypotheses?. arXiv preprint arXiv:2412.06438. External Links: Link Cited by: §1. [46] T. Schick, J. Dwivedi-Yu, R. Dessì, et al. (2023) Toolformer: language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761. Cited by: §1. [47] T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom (2023) Toolformer: language models can teach themselves to use tools. External Links: 2302.04761, Link Cited by: §2. [48] Y. Shen, J. Yuan, T. Aonishi, H. Nakayama, and Y. Ma (2025) Follow-your-preference: towards preference-aligned image inpainting. arXiv preprint arXiv:2509.23082. Cited by: §2. [49] N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao (2023) Reflexion: language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366. Cited by: §1. [50] Y. Song, W. Xiong, D. Zhu, W. Wu, H. Qian, M. Song, H. Huang, C. Li, K. Wang, R. Yao, Y. Tian, and S. Li (2023) RestGPT: connecting large language models with real-world restful apis. External Links: 2306.06624, Link Cited by: §2. [51] C. G. Sorensen, S. Fountas, D. Nash, S. M. Pedersen, B. Basso, and S. Blackmore (2010) Agricultural information systems for precision agriculture: a review. European Journal of Agronomy 33 (2), p. 103–113. Cited by: §1. [52] Spatial joins — introduction to postgis. Note: PostGIS WorkshopOnline documentation External Links: Link Cited by: §2. [53] Q. Team (2025) Qwen3 technical report. arXiv preprint arXiv:2501.12948. Cited by: §4.1.2. [54] R. Thoppilan, D. De Freitas, J. Hall, et al. (2022) LaMDA: language models for dialog applications. arXiv preprint arXiv:2201.08239. Cited by: §1. [55] L. Wang, W. Huang, X. Zhao, G. Wei, D. Su, Y. Zhang, M. Wen, Z. Xu, P. Zhao, and S. Zhang (2023) A survey on large language model based agents. arXiv preprint arXiv:2308.11432. Cited by: §1. [56] L. Wang, J. Dong, J. Shang, J. Liu, and Z. Qian (2020) Review of soil moisture retrieval from remote sensing: a focus on agricultural applications. Agriculture 10 (12), p. 611. Cited by: §1. [57] J. Yao, J. Zhang, X. Pan, T. Wu, and C. Xiao (2024) DepthSSC: monocular 3d semantic scene completion via depth-spatial alignment and voxel adaptation. External Links: 2311.17084, Link Cited by: §1. [58] J. Zhang, K. Cai, Y. Fan, N. Liu, and K. Wang (2025) MAT-agent: adaptive multi-agent training optimization. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, External Links: Link Cited by: §1. [59] J. Zhang, K. Cai, Y. Fan, J. Wang, and K. Wang (2025) CF-VLM: counterfactual vision-language fine-tuning. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, External Links: Link Cited by: §1. [60] J. Zhang, K. Cai, X. Guo, S. Liu, Q. Lv, R. Chen, J. Yang, Y. Fan, X. Sun, J. Wang, Z. Chen, L. Lin, and K. Wang (2025) M-cot:a benchmark for probing visual chain-of-thought reasoning in multimodal models. External Links: 2512.08228, Link Cited by: §1. [61] J. Zhang, K. Cai, J. Wang, Y. Zheng, K. Lam, and K. Wang (2026) Process-of-thought reasoning for videos. External Links: 2602.07689, Link Cited by: §1. [62] J. Zhang, K. Cai, J. Yang, and K. Wang (2025) Learning dynamics of vlm finetuning. External Links: 2510.11978, Link Cited by: §1. [63] J. Zhang, K. Cai, Q. Zeng, N. Liu, S. Fan, Z. Chen, and K. Wang (2025) Failure-driven workflow refinement. External Links: 2510.10035, Link Cited by: §1. [64] J. Zhang, Y. Fan, K. Cai, Z. Huang, X. Sun, J. Wang, C. Tang, and K. Wang (2025) DrDiff: dynamic routing diffusion with hierarchical attention for breaking the efficiency-quality trade-off. External Links: 2509.02785, Link Cited by: §1. [65] J. Zhang, Y. Fan, K. Cai, X. Sun, and K. Wang (2025) OSC: cognitive orchestration through dynamic knowledge alignment in multi-agent llm collaboration. External Links: 2509.04876, Link Cited by: §1. [66] J. Zhang, Y. Fan, K. Cai, J. Yang, J. Yao, J. Wang, G. Qu, Z. Chen, and K. Wang (2026) Why keep your doubts to yourself? trading visual uncertainties in multi-agent bandit systems. External Links: 2601.18735, Link Cited by: §1. [67] J. Zhang, Y. Fan, K. Cai, J. Yang, Y. Zheng, K. Lam, L. Lin, and K. Wang (2026) Spectral gating networks. External Links: 2602.07679, Link Cited by: §1. [68] J. Zhang, Y. Fan, W. Lin, R. Chen, H. Jiang, W. Chai, J. Wang, and K. Wang (2025) GAM-agent: game-theoretic and uncertainty-aware collaboration for complex visual reasoning. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, External Links: Link Cited by: §2. [69] J. Zhang, Y. Fan, Z. Wen, J. Wang, and K. Wang (2025) Tri-MARF: a tri-modal multi-agent responsive framework for comprehensive 3d object annotation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, External Links: Link Cited by: §1. [70] J. Zhang, X. Guo, K. Cai, Q. Lv, Y. Fan, W. Chai, J. Wang, and K. Wang (2025) HybridToken-vlm: hybrid token compression for vision-language models. External Links: 2512.08240, Link Cited by: §1. [71] J. Zhang, Z. Huang, Y. Fan, N. Liu, M. Li, Z. Yang, J. Yao, J. Wang, and K. Wang (2025) KABB: knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems. In Forty-second International Conference on Machine Learning, External Links: Link Cited by: §2. [72] J. Zhang, N. Liu, Q. Lyu, J. Yang, and K. Wang (2026) Rational anova networks. External Links: 2602.04006, Link Cited by: §1. [73] N. Zhang, M. Wang, and N. Wang (2002) Precision agriculture—a worldwide overview. Computers and Electronics in Agriculture 36 (2–3), p. 113–132. External Links: Document Cited by: §1.

Instant research discovery