Iterative Agentic Loop in GeoSR

• Iterative Agentic Loop is a modular framework that coordinates specialized agents to iteratively refine predictions using contextual feedback and geostatistical priors.
• GeoSR implements distinct agents for prediction, variable selection, and spatial anchoring to enhance accuracy and fairness with significant empirical gains.
• The iterative process enforces spatial consistency and systematically reduces bias, as evidenced by improved Spearman correlations and reduced prediction errors.

An Iterative Agentic Loop is a modular, multi-step reasoning and self-correction framework wherein specialized agents coordinate to iteratively generate, assess, and refine outputs by explicitly leveraging contextual feedback in a closed-loop configuration. In GeoSR, this paradigm is instantiated for geospatial prediction by orchestrating distinct variable selection, local spatial anchoring, and iterative refinement agents, embedding geostatistical priors and domain heuristics (e.g., Tobler’s First Law). This agentic sequence progressively updates zero-shot LLM predictions, enforces spatial consistency, and systematically reduces bias, with well-characterized convergence and empirically validated gains over standard LLM prompting (Tang et al., 6 Aug 2025).

1. Structural Decomposition and Agent Roles

The GeoSR realization of the Iterative Agentic Loop comprises three lightweight, interleaved LLM-driven agent roles:

• Predict Agent ($\pi$): Initializes with base LLM prompts per location, generating initial zero-shot predictions that may include (optionally) map-based or external covariate context.
• Variable-Selection Agent ($\varphi$): Selects a compact, task-dependent subset of auxiliary variables at each location, conditioned on local context and predicted correlations, to enrich the refinement prompt.
• Point-Selection Agent ($\psi$): Identifies spatially proximate reference locations serving as empirical anchors via discrete neighbor selection (instantiating Tobler's Law).
• Refine Agent ($\phi$): Synthesizes local evidence—including prior prediction, selected covariates, and neighboring reference predictions—then evaluates the need for update and issues refined values if appropriate.

This modular decomposition isolates each aspect of geospatial reasoning and correction, with each agent implemented as a prompt-programmed LLM query. Agents are strictly separated: the prediction agent is only responsible for initial value generation, while refinement is exclusively mediated by evidence aggregation and explicit decision rules within the Refine Agent.

2. Iterative Loop Workflow and Algorithmic Formalization

The agentic loop operates through well-defined stages at each iteration $k$ (pseudocode as in the original source):

1. Initialization: For each spatial locus $\ell_t$, produce starting prediction $$\hat{y}_t^{(0)} = \pi(\mathrm{prompt}_\mathrm{base}(\ell_t, \tau))$$.
2. For $k = 1 \ldots K$ (max rounds):
   • For each $\ell_t$:
     • $S_t \leftarrow \varphi(\ell_t, \tau)$ — select covariate subset.
     • $\varphi$0 values of $\varphi$1 at $\varphi$2.
     • $\varphi$3 — select neighborhood anchors.
     • refs $\varphi$4
     • covs $\varphi$5
     • decision $\varphi$6
     • If update needed: $\varphi$7 and set update flag.
   • If no updates across all $\varphi$8: terminate early.
3. Output: Converged solution $\varphi$9.

This process can be formalized as the recursive update: $\psi$0

Convergence is determined by the early-exit condition when no further updates are flagged during iteration, or by reaching a fixed iteration upper bound ($\psi$1).

3. Mathematical Principles and Geostatistical Priors

A central organizing principle is Tobler’s First Law of Geography: “everything is related to everything else, but near things are more related than distant things.” GeoSR operationalizes this law discretely, always including a core of $\psi$2 nearest neighbors in $\psi$3 for each target location.

Variable selection is formalized by mapping $\psi$4 (with $\psi$5 the set of all covariates), typically employing prompt-based LLM reasoning to select task-relevant covariates. Unlike classic geostatistical kernels, GeoSR relies on discrete anchor selection rather than explicit spatial weighting functions.

For performance evaluation, the framework employs:

• Spearman Rank Correlation ($\psi$6):

$\psi$7

capturing monotonic association of prediction and truth.

• Bias Metric ($\psi$8):

$\psi$9

with $\phi$0 an anchor distribution vector (e.g., population density), $\phi$1 median absolute deviation, and $\phi$2 answer rate, quantifying unwanted correlation between predictions and demographic anchors.

4. Empirical Dynamics, Convergence, and Quantitative Impact

Empirical results indicate that most models’ Spearman correlation rises sharply within the first two or three refinement rounds, plateauing or slightly declining thereafter. Bias (as defined above) steadily decreases with successive rounds, even when accuracy gains stall—demonstrating an explicit fairness-accuracy tradeoff in later iterations. The optimal number of loop iterations is therefore architecture- and domain-dependent, but $\phi$3–3 is optimal for both generalist and domain-specialized models.

Quantitatively, the GeoSR iterative agentic loop produces substantial improvements:

• GPT-3.5-Turbo’s Spearman correlation on infant mortality improves from 0.445 (base) to 0.747 after iteration (+67.9%).
• Bias reductions reach up to 96.8% for some model-task pairs.
• Removing the point-selection anchor (nearest-10) or the variable selection agent reduces both accuracy and fairness, confirming the necessity of iterative, context-anchored refinement and covariate-dependent evidence construction.

5. Architectural Modularity and Generalization Properties

GeoSR’s loop is strictly modular, making each agent independently replaceable or upgradable:

• The point-selection and variable-selection agents can be swapped to incorporate different heuristics (e.g., population density vs. topographic distance).
• The system operates at prompt time, with no gradient-based fine-tuning or model modification required, supporting full compatibility with off-the-shelf LLMs.
• The protocol admits domain-specialized enrichment (e.g., variable selection for bioclimatic tasks vs. socioeconomic ones) by modifying only the variable-selection prompt logic.

Empirical results, including spatial rank-maps pre- and post-GeoSR, demonstrate not only uplift in accuracy but also correction of severe conceptual errors such as “inverted” clusters (e.g., infant mortality over Sub-Saharan Africa). Table 4 and Table 5 in the source demonstrate, via ablation, that each module is essential to the robustness and equity of the final solution (Tang et al., 6 Aug 2025).

6. Implications for LLM-Based Geospatial Reasoning and Broader Agentic Systems

The iterative agentic loop, as instantiated in GeoSR, evidences that recurrently contextualizing LLM predictions with local evidence and multi-agent collaboration can transcend the static limitations of prompt engineering—including systematic mitigation of spatial bias and amplification of multi-hop spatial reasoning potential.

Three general properties emerge:

• The modular, LLM-agent loop paradigm is directly extensible to any setting where predictions can be anchored to inter-related, context-sensitive evidence.
• By decoupling refinement triggers and update logic into separate agents, the loop can integrate domain-specific stopping criteria or convergence safeguards.
• The explicit, interpretable structure of evidence and reference selection provides avenues for theoretical analysis, auditability, and fairness control that are not possible in opaque, one-shot LLM prediction regimes.

In summary, the Iterative Agentic Loop in GeoSR represents a principled, empirically validated method for decomposing, contextualizing, and self-correcting LLM-driven predictions, providing a reproducible, extensible foundation for agentic systems in geospatial, scientific, and other structured domains (Tang et al., 6 Aug 2025).