Geo-Expert: Domain-Aligned Geoscience Models

• Geo-Expert is a class of models fine-tuned on authoritative geological content to perform rigorous reasoning in stratigraphy, tectonics, and basin analysis.
• It integrates geographic priors, iterative self-refinement, and tool-grounding to mitigate systematic hallucinations in complex geoscience queries.
• Empirical benchmarks show that parameter-efficient Geo-Expert models outperform larger general-purpose models by recalibrating semantic and deductive geological logic.

Geo-Expert denotes a class of domain-aligned geoscience expert systems centered on rigorous, task-specific reasoning rather than generic language fluency. In the narrow sense, it is a family of parameter-efficient geological LLMs fine-tuned for subsurface and deep-time reasoning, especially in stratigraphy, structural geology, tectonics, and basin analysis (Guo et al., 24 May 2026). In a broader systems sense, related work uses the term as a conceptual backbone for geospatial expert agents that inject geographic priors, executable tools, multimodal evidence, and verifiable evaluation into prediction, localization, disaster reasoning, and image-grounded spatial analysis (Tang et al., 6 Aug 2025, Hasan et al., 15 Jun 2026).

1. Scope, genealogy, and problem setting

The defining motivation of Geo-Expert is that general-purpose models often fail on geoscience tasks that require chaining physical principles across spatial and temporal scales. In the geological setting, the central failure mode is systematic hallucination under subsurface and deep-time queries. The reported causes include multi-step deductive demands such as cross-cutting relationships, stratigraphic cyclicity, stress/strain rheology, and plate-scale tectonic context, as well as polysemous terminology for which generalists default to non-geological meanings (Guo et al., 24 May 2026).

Related work broadens the same idea from solid-Earth reasoning to geospatial intelligence. In that literature, a Geo-Expert is not merely a model with geographic vocabulary; it is a system that reasons with locality, topology, covariates, tools, and explicit validation. GeoSR frames this through iterative self-refinement with Tobler’s First Law as a prior (Tang et al., 6 Aug 2025). GeoDisaster operationalizes it through tool-grounded, contract-verified disaster reasoning (Hasan et al., 15 Jun 2026). SpotAgent and GeoAgent apply closely related principles to verifiable geo-localization through tool use or geographically faithful rewards (Jia et al., 10 Feb 2026, Jin et al., 13 Feb 2026). GeoX extends the idea to executable spatial programs with self-play and verifiable rewards over overhead imagery (Ahn et al., 19 May 2026).

System	Domain	Core mechanism
Geo-Expert	Geological reasoning	LoRA-tuned textbook-aligned LLMs
GeoSR	Geospatial prediction	Agentic spatial self-refinement
GeoDisaster	Disaster geo-intelligence	Orchestrated agents with execution contracts
SpotAgent / GeoAgent	Image geo-localization	Tool grounding or geo-similarity RL
GeoX	Overhead-image reasoning	Executable programs with self-play

This lineage suggests that Geo-Expert is best understood as a design objective: expert-level behavior emerges when domain structure, data curation, and evaluation are matched to the target geoscience task rather than delegated to generic pretrained priors.

2. Geological Geo-Expert models and instruction tuning

The core Geo-Expert family is built by parameter-efficient fine-tuning of decoder-only Transformers: Qwen3-8B, Qwen3-32B, and Gemma-3-27B (Guo et al., 24 May 2026). “Domain-aligned” in this setting means instruction-tuned on authoritative geological content with Chain-of-Thought supervision and domain-structured tags. The stated aim is to recalibrate semantic priors toward solid-Earth concepts and the deductive logic of geology rather than mere factual retrieval.

Its data pipeline converts five canonical textbooks into 11,518 CoT-enhanced instruction pairs. The reported workflow is MinerU PDF→Markdown extraction with structure preservation, Python cleaning, hash-based global deduplication, chapter-aware recursive chunking, hierarchical domain-tree construction, dynamic question generation for coverage, and CoT answer construction aligned to source material (Guo et al., 24 May 2026). The covered domains include subsurface geological reasoning, stratigraphy, tectonics, structural kinematics, petrology, paleontology, basin analysis, and deep-time evolution.

The adaptation mechanism is LoRA over “all linear” modules across attention and MLP blocks. The paper gives the standard update

$\Delta W = BA,\qquad W = W_0 + \Delta W,$

with $A \in \mathbb{R}^{r \times d_{in}}$, $B \in \mathbb{R}^{d_{out} \times r}$, and trainable parameter count

$P = r(d_{in}+d_{out}).$

This is paired with teacher-forced cross-entropy

$L = -\sum_{t=1}^{T}\log p(y_t \mid y_{<t}, x).$

No RLHF or preference optimization is reported (Guo et al., 24 May 2026).

The hyperparameter regime is scale-dependent. Qwen3-8B-geo uses $r=32$, $\alpha=32$, LoRA dropout $0.05$, learning rate $2e{-5}$, FP16 precision, AdamW, and a single NVIDIA RTX 5090. Gemma-3-27B-geo and Qwen3-32B-geo use $r=64$, $A \in \mathbb{R}^{r \times d_{in}}$0, dropout $A \in \mathbb{R}^{r \times d_{in}}$1, learning rate $A \in \mathbb{R}^{r \times d_{in}}$2, BF16, AdamW, gradient checkpointing, gradient accumulation $A \in \mathbb{R}^{r \times d_{in}}$3, batch size $A \in \mathbb{R}^{r \times d_{in}}$4 per device, and $A \in \mathbb{R}^{r \times d_{in}}$5 RTX 5090 (Guo et al., 24 May 2026). The training recipe is explicitly positioned as a reproducible, consumer-grade route to expert-level geological specialization.

3. Geo-Eval and empirical performance

Geo-Expert is evaluated on Geo-Eval, a benchmark built to probe “hard boundary” geological reasoning rather than broad factual recall. The construction pipeline begins with 2,591 complex open-ended questions mined from the same authoritative sources, then applies comparative inference with Qwen3-8B-geo and DeepSeek-R1, difficulty-aware LLM-as-a-Judge mining using GLM-4.5, and expert vetting by geology professors to finalize 387 questions. The taxonomy is Concept, Process, and Engineering (Guo et al., 24 May 2026).

The scoring protocol uses GPT-4o as a reference-guided judge with strict prompts that enforce semantic correctness and penalize hallucinations. Mean category scores are reported, and the gains over base models are backed by a paired $A \in \mathbb{R}^{r \times d_{in}}$6-test with $A \in \mathbb{R}^{r \times d_{in}}$7 (Guo et al., 24 May 2026).

The main quantitative result is that the domain-aligned 8B model outperforms much larger generalists on specialized geology. Qwen3-8B-geo reaches an average score of 6.27, versus 4.12 for Llama-3.1-70B-Instruct and 5.93 for GPT-4o. Qwen3-32B-geo reaches 6.82, approaching GPT-5.4 at 7.15. Gemma-3-27B-geo reaches 6.59. The absolute gains over the untuned bases are +1.64 for Qwen3-8B, +1.82 for Qwen3-32B, and +1.43 for Gemma-3-27B-IT (Guo et al., 24 May 2026).

Model	Average score	Note
Qwen3-8B-geo	6.27	Beats GPT-4o and Llama-3.1-70B-Instruct
Gemma-3-27B-geo	6.59	Strong domain-aligned mid-scale model
Qwen3-32B-geo	6.82	Approaches GPT-5.4
GPT-4o	5.93	Strong generalist baseline
GPT-5.4	7.15	Highest reported reference

A representative failure analysis concerns the question “What phenomena does the local thickening of a wedge cause, and how is this deformation accommodated?” GPT-4o is reported to drift into civil engineering reinforcement, scoring 0/10, whereas Qwen3-8B-geo anchors the answer in structural geology, including stress concentration, thrust reactivation, and accommodation via thrust sliding, extensional collapse, and lateral flow, scoring 9/10 (Guo et al., 24 May 2026). The broader implication is that scaling alone is not decisive in vertical domains; domain alignment changes the meaning of the vocabulary on which reasoning depends.

4. Geo-Expert as an agentic geospatial architecture

Outside geology proper, the Geo-Expert idea is instantiated as agentic reasoning over geographic structure. GeoSR is the clearest formalization of this pattern. It embeds Tobler’s First Law into an iterative self-refinement loop with a Variable-Selection Agent, a Point-Selection Agent, and a Refine Agent, while a Predict Agent supplies initial outputs. The update uses selected covariates at the same location, prior predictions from nearby locations, and optional farther references for global context. The paper reports that performance often peaks at 2–3 rounds, while bias tends to decrease monotonically with more rounds, producing a controllable accuracy–fairness trade-off (Tang et al., 6 Aug 2025). Across tasks and models, GeoSR improves both Spearman’s rank correlation and the composite bias score; for example, GPT-3.5-Turbo on Infant Mortality improves from 0.445 to 0.747 in Spearman and from −0.188 to −0.006 in Bias (Tang et al., 6 Aug 2025).

GeoDisaster pushes the same principle into operational disaster geo-intelligence. It defines a benchmark with 2,921 verified instances, 43 question types, and five task families, then couples it to an orchestrated multi-agent system with 18 disaster-oriented tools and typed execution contracts of the form

$A \in \mathbb{R}^{r \times d_{in}}$8

Alignment is achieved through Role-Contract Expectation Alignment, combining failure-aware supervised fine-tuning with contract-grounded reinforcement learning over dense step-level signals. The aligned system reaches Ans 90.11, TSR 94.24, ToolAnyOr 98.56, ToolUni 98.72, and perfect step-wise Inst./Tool/ArgN/ArgV scores of 100.00 on GeoDisaster (Hasan et al., 15 Jun 2026). Here, Geo-Expert behavior is defined not by free-form explanation but by valid tool calls, correct evidence handles, CRS consistency, and decision generation under deterministic checks.

Geo-localization work provides another operational reading of the concept. SpotAgent models geo-localization as a POMDP with ReAct-style tool use, using WebSearch, GeoCoding, and ImageTool within a six-call budget. On Im2GPS3k, the full toolset reaches 14.12 / 40.36 / 57.80 / 73.43 / 85.75 across Accuracy@$A \in \mathbb{R}^{r \times d_{in}}$9 km, outperforming ablations with any single tool (Jia et al., 10 Feb 2026). GeoAgent instead emphasizes human-authored CoT, geo-similarity rewards, and a consistency agent; after SFT and GRPO, it reaches 40.75 / 58.57 / 76.21 / 89.90 on IM2GPS3K across City/Region/Country/Continent thresholds, exceeding GeoCLIP, PIGEON, and GLOBE (Jin et al., 13 Feb 2026). GeoX generalizes the same theme to overhead imagery, training a single multimodal policy by self-play over executable spatial programs with verifiable rewards; its base VLMs improve by up to 5.5 points on average and blackQ reaches 51.2 on EarthVQA and 43.3 on GEOBench-VLM (Ahn et al., 19 May 2026).

Taken together, these systems show a recurring Geo-Expert architecture: domain-aligned representations, an explicit spatial interface, iterative or multi-agent planning, and rewards or evaluations that are verifiable in geographic space rather than purely linguistic.

5. Benchmarks, interpretability, and geospatial data interfaces

A mature Geo-Expert stack depends not only on model training but also on domain-specific evaluation and interpretable interfaces. GeoPro-Net addresses interpretability in spatiotemporal forecasting through statistically guided Geo-concepts, interpretable channel fusion, geographic-based pooling, and prototype learning. The model uses local and global significance tests to create concept channels, then projects learned prototypes to real-world cases. On Chicago crime it reports approximately CrsEnt $B \in \mathbb{R}^{d_{out} \times r}$0, ACC $B \in \mathbb{R}^{d_{out} \times r}$1, and F1 $B \in \mathbb{R}^{d_{out} \times r}$2; spatial concept pooling outperforms both no pooling and max pooling while preserving interpretability (An et al., 2024).

GeoGrid-Bench evaluates whether foundation models can reason over multimodal gridded geospatial data. It contains approximately 3,200 question-answer pairs generated from 8 expert-curated templates over 16 climate variables and 150 locations. The reported conclusion is that vision-LLMs perform best overall, annotated heatmaps outperform raw tabular inputs, and code-based setups underperform because generated scripts are often incomplete or non-executable without agentic orchestration (Jiang et al., 15 May 2025). This benchmark is diagnostic rather than operational: it isolates trend reasoning, spatial references, coordinate references, and label references.

GeoR-Bench focuses on geoscience visual reasoning through reasoning-informed visual editing. It contains 440 curated samples across 6 geoscience categories and 24 task types, scored on Reasoning, Consistency, and Quality, with strict accuracy requiring all three to pass simultaneously. The best-performing model reaches 42.7% overall strict accuracy, while the best open-source model reaches 10.3%; the paper’s central finding is that visual consistency and image quality frequently surpass scientific accuracy (Zheng et al., 12 May 2026). This sharply distinguishes visually plausible outputs from geoscientifically valid ones.

At the data-infrastructure level, Open Data Cube work shows how a regional Geo-Expert can be operationalized over imagery pipelines rather than model prompts alone. The reported system is a Dockerized, ODC-based service with PostgreSQL/PostGIS that automates ingestion, polygon-centric loading, per-pixel extraction, and export for approximately 20,000 AOIs in the Basque Country using Sentinel-2. It emphasizes ROI loading restricted to pixels strictly inside polygons, automatic metadata inference, GeoPandas outputs, and routine generation of NDVI and EVI products (Otamendi et al., 2022). This kind of infrastructure is not itself a reasoning model, but it defines the evidentiary substrate on which Geo-Expert systems depend.

Benchmark or system	Focus	Key signal
GeoPro-Net	Interpretable spatiotemporal prediction	Geo-concepts and prototypes
GeoGrid-Bench	Gridded multimodal reasoning	Heatmaps outperform tables
GeoR-Bench	Geoscience visual reasoning	Strict accuracy bottleneck
ODC regional pipeline	Geo-imagery operations	Polygon-centric ingestion and export

6. Limitations, misconceptions, and future directions

Several limitations recur across the Geo-Expert literature. In the geological LLM setting, coverage is uneven: structural geology, stratigraphy, and tectonics are over-represented, while mineralogy, geochemistry, and geophysics remain under-represented. Geo-Eval contains 387 expert-vetted items, which suffices for boundary stress testing but remains smaller than broad general benchmarks. The current models are text-only, while geological practice is inherently multimodal, involving cross-sections, logs, and imagery (Guo et al., 24 May 2026).

In geospatial prediction and agentic systems, the main assumptions are structural. GeoSR assumes spatial autocorrelation; where Tobler’s Law is weak, nearest-neighbor references can mislead. More refinement rounds reduce Bias but may plateau or slightly reduce Spearman, and distant reference points can increase Bias unless tuned carefully (Tang et al., 6 Aug 2025). GeoDisaster identifies non-uniform regional coverage, ambiguity in visual evidence, and the operational risk of overconfident routing or exposure assessment (Hasan et al., 15 Jun 2026). SpotAgent notes sparse indoor scenes, semantic traps, and external API dependence, while GeoAgent identifies low-cue scenes and residual geographic bias despite stratified sampling (Jia et al., 10 Feb 2026, Jin et al., 13 Feb 2026).

Two common misconceptions are directly contradicted by the reported evidence. The first is that larger generic models necessarily outperform smaller domain models. Geo-Expert shows that a domain-aligned 8B model can outperform both GPT-4o and an open-weight 70B generalist on specialized geological reasoning (Guo et al., 24 May 2026). The second is that visually polished outputs imply scientific competence. GeoR-Bench shows the opposite: models often achieve strong Consistency and Quality while failing Reasoning, which indicates that superficially plausible geoscience artifacts can still be scientifically wrong (Zheng et al., 12 May 2026).

The future directions are correspondingly concrete. Geo-Expert proposes integrating vision-LLMs for stratigraphic profiles and remote sensing, as well as Retrieval-Augmented Generation for dynamic literature grounding (Guo et al., 24 May 2026). GeoSR suggests kriging-inspired refinements, spatiotemporal extensions, graph-based priors, uncertainty calibration, and human-in-the-loop overrides (Tang et al., 6 Aug 2025). GeoDisaster calls for uncertainty quantification in exposure and routing, richer multi-temporal evidence, more sensor modalities, probabilistic consistency checks, and expanded task families such as landslide and cyclone wind fields (Hasan et al., 15 Jun 2026). GeoR-Bench implies that future progress will require models that internalize Earth-system process constraints rather than merely improving image-editing fidelity (Zheng et al., 12 May 2026).

A plausible implication is that the term “Geo-Expert” will continue to denote not one monolithic model class but a convergence of four design commitments: authoritative domain alignment, explicit spatial structure, verifiable reasoning, and benchmark regimes that separate scientific correctness from surface plausibility.