GeoFM: Geospatial & Geometric Reasoning

• GeoFM is a class of foundation models that integrate formal language synthesis for creating high-fidelity synthetic geometric and geospatial data.
• It employs transformer architectures pre-trained on heterogeneous spatial datasets, ensuring robust transferability for tasks like segmentation and urban analytics.
• GeoFM’s pipeline uses conditional declaration language and symbolic proof generation to optimize diagram accuracy and minimize label noise.

GeoFM refers to a class of Geospatial Foundation Models—large, transformer-based neural architectures pre-trained on massive, heterogeneous spatial data to produce general-purpose, transferable representations for a variety of geographic or geometric reasoning tasks. The term also designates a specific model—GeoFM for geometric reasoning—proposed as a synthetic data-generation pipeline for enhancing the geometric problem-solving capabilities of multi-modal LLMs (MLLMs), as well as a broader family of methods and models for geospatial data fusion, representation learning, and downstream task generalization. This entry focuses on the technical foundations, data-centric innovations, representative architectures, benchmarking approaches, and open research challenges associated with GeoFM in both the geometric and geospatial sense, referencing the canonical GeoFM pipeline as described in "GeoFM: Enhancing Geometric Reasoning of MLLMs via Synthetic Data Generation through Formal Language" (Zhang et al., 31 Oct 2025), alongside linked work in Earth observation GeoFMs.

1. Definition and Scope of GeoFM

GeoFM, in the geometric context, is an algorithmic framework for generating synthetic, formally-specified, high-fidelity geometric data to improve the mathematical geometric reasoning abilities of MLLMs. It synthesizes diverse geometry QA pairs via formal language manipulations and ensures correctness with a symbolic engine. More broadly, GeoFM in the geospatial context refers to foundation models pre-trained on spatial data—satellite imagery, remote sensing (RS), multimodal and temporal environmental data—designed for strong transferability and sample efficiency across varied downstream geospatial tasks, from semantic segmentation to resource mapping to urban analytics (Simumba et al., 19 Nov 2025, Mahlasi et al., 11 Feb 2026, Du et al., 1 Dec 2025, Ghamisi et al., 30 May 2025).

2. Formal Language Foundations and Synthetic Data Generation

GeoFM's technical core is a formal language, the Conditional Declaration Language (CDL), which encodes all geometric entities, relations, and problem statements in four segments:

• Construction CDL: declarative definitions of primitives (e.g., Point(A), Line(A,B), Circle(O,r), On(C,AB)).
• Text CDL: explicit metric constraints stated in the problem (e.g., Dist(A,B)=5, Angle(A,B,C)=90°).
• Image CDL: metric information visible only in the geometric diagram, used to ground image understanding.
• Goal CDL: the quantity or property to solve for (e.g., Dist(A,B), Angle(A,B,C)).

Formally, geometric constraints are specified as CDL expressions, with explicit equality and relation syntax. For example,

• Dist(A,B)=5 defines $d(A,B)=5$,
• Parallel(AB,CD) expresses $AB \parallel CD$,
• Perp(AB,BC) denotes $AB \perp BC$.

GeoFM parses seed geometry problems and diagrams via a combination of LLM-based text parsing and OCR-assisted diagram parsers, extracting full metric information through the FormalGeo symbolic solver (breadth-first search over geometric theorem libraries). It then systematically generates new problems by swapping conditions in the metric space, guaranteeing solvable, non-redundant instances.

GeoFM’s synthetic dataset pipeline involves the following:

1. Formalization of seed data to CDL from raw text and diagrams.
2. Metric enumeration and manipulation: Enumerate all metric statements $M_{\text{all}}$ consistent with the geometry, then randomly add/delete subsets to produce new problem statements.
3. Symbolic proof generation with FormalGeo, enforcing only problems with provable and correct solutions are retained.
4. Natural-language conversion: Formal problems and proofs are translated to linguistically rich Q&A statements via template-based conversion and LLM rewriting (e.g., Qwen2.5-72B).
5. Diagram synthesis: Generate high-fidelity images from CDL by translating into GMBL (geometry markup), optimizing for constraint satisfaction; invalid diagrams are filtered by constraint-violation losses.
6. Noise and diversity enhancement: Conditions are randomly assigned to text or image CDL, varying resolution, and point ordering.

This approach contrasts with prior methods that rely on problem rephrasing or fixed template expansion, yielding limited diversity and substantial label noise.

3. Model Architectures and Training Protocols

GeoFM-generated data is used to train MLLMs such as LLaVA-NeXT-8B and InternVL2-8B-MPO in a full-parameter fine-tuning regime.

• GeoFM80K: The synthetic dataset comprises 80,000 formally verified geometry QA pairs.
• Architecture specifics:
  • LLaVA-NeXT-8B: tuned for 2 epochs, batch size 64, learning rates for LLM, adapter, and vision modules set at $3 \times 10^{-5}$ and $2 \times 10^{-6}$ respectively, with a cosine LR schedule and 3% warmup.
  • InternVL2-8B-MPO: similar protocol, but with fixed vision and adapter modules.

Training batches are augmented with diagrams at varying resolutions (short edge 112, 224, 336 px, aspect ratio 4:3).

• Answer verification: LLM-predicted answers are cross-checked against FormalGeo’s symbolic solution to eliminate hallucinated or incorrect data.

4. Experimental Results and Quantitative Benchmarks

GeoFM-trained MLLMs deliver substantial improvements on standard geometry reasoning tasks:

• MathVista-GPS: GeoFM-8B achieves 79.3% accuracy; GPT-4o baseline at 60.6% (Δ=+18.7%).
• GeoQA: GeoFM-8B achieves 77.9%; GPT-4o at 61.4% (Δ=+16.5%).
• Open-source baseline comparison: InternVL2-8B-MPO baseline at 73.6% (MathVista) and 75.2% (GeoQA); GeoFM delivers +5.7 pp and +2.7 pp improvements, respectively.
• Dataset ablations: Adding GeoFM data to open-source sets (Geo170K-QA / MathV360K-GPS) yields further performance gains (e.g., MathVista +1.9% for LLaVA-NeXT-8B).

Ablation studies show that models trained solely on seed data exhibit a notable drop (up to 17.3%) in generalization when evaluated on novel condition combinations—underscoring the importance of synthetic diversity.

Qualitative comparisons indicate that GeoFM-augmented models resolve geometric queries with precise, interpretable solutions, whereas strong proprietary baselines (e.g., GPT-4o) may hallucinate geometric properties (e.g., false collinearity or incorrect angles).

5. Significance of Formal-Language Synthesis and Symbolic Guarantees

Formal-language-driven synthesis in GeoFM offers multiple advantages:

• Systematic metric space exploration: Enumerates latent geometric constraints beyond template-based synthetic data pipelines, substantially increasing diversity.
• Symbolic correctness: The formal–symbolic–proof pipeline eliminates labeling errors and LLM hallucinations, as only verifiably provable problems are admitted.
• Diagram fidelity: GMBL-based diagram optimization ensures diagrams are both numerically consistent with formal constraints and visually diverse.

The model thereby produces data that is both more varied and less noisy than prior template or rule-based synthetic pipelines. This contributes directly to superior generalization and robustness on previously unseen tasks or problem structures (Zhang et al., 31 Oct 2025).

6. Limitations and Prospective Extensions

Despite strong empirical gains, current GeoFM pipelines have important limitations:

• Manual seed curation: The pipeline depends on manually collected seed problems, reducing scalability.
• Formalization gaps: Problems lacking explicit point identifiers or with highly nonstandard diagrams elude formalization and symbolic parsing.
• Word-problem coverage: Word-heavy geometric problems without formal element mapping remain insufficiently addressed.
• Data domain constraints: Extensions to 3D geometry, circle/conic theorems, and formalization of non-planar or projective settings are currently in development.
• Future work: Target directions include fully automated formal problem generation, integration of learned parsers to reduce manual intervention, and coverage expansion to additional geometric subfields and spatial reasoning settings.

7. Relationship to Broader Geospatial Foundation Models

The term GeoFM is frequently used across geospatial AI and Earth Observation to denote large, self-supervised, transformer-based models pre-trained on satellite and multimodal spatial datasets (Mahlasi et al., 11 Feb 2026, Simumba et al., 19 Nov 2025, Du et al., 1 Dec 2025, Ghamisi et al., 30 May 2025, Romero et al., 11 Jun 2026, Gilch et al., 11 Mar 2026). These models enable generalization, data efficiency, and robust adaptation across a wide spectrum of tasks, including:

• Semantic segmentation, pixelwise regression, and change detection in Earth Observation imagery (Simumba et al., 19 Nov 2025, Ghamisi et al., 30 May 2025).
• Ecological trait mapping and cross-modal geospatial reasoning (Mahlasi et al., 11 Feb 2026).
• Resource-constrained, on-orbit model deployment (Du et al., 1 Dec 2025).
• Multimodal urban analytics (fusing imagery, POI, maps) (Mühlematter et al., 15 Oct 2025).
• Embedding-centric, decoupled workflows exploiting GeoFM representations for efficient inference and transfer (Gilch et al., 11 Mar 2026, Romero et al., 11 Jun 2026).

In all these domains, principles of broad pretraining, architecture scaling, multimodal fusion, and rigorous benchmarking (e.g., GEO-Bench-2, SustainFM, NeuCo-Bench) underlie the ongoing development and assessment of GeoFM variants.

---

In summary, GeoFM, as both a formal-synthesis pipeline and a family of geospatial foundation models, occupies a pivotal technical position in contemporary geometric and spatial reasoning with large neural architectures. Its formal-language-driven data generation, symbolic validation, and demonstration of downstream gains set a reference standard for designing and evaluating robust, generalist models for mathematical and spatial AI (Zhang et al., 31 Oct 2025).