GeoX: Geometric & Geospatial Reasoning

• GeoX is a family of specialized models and benchmarks for geometric, geospatial, and cross-view reasoning in vision-language systems.
• The framework utilizes a three-stage pre-training pipeline, reinforcement learning, and geometry-preserving distillation to achieve state-of-the-art performance.
• GeoX benchmarks enable precise evaluations in geo-localization, pose estimation, and 3D reconstruction, driving research in autonomous driving and spatial logic.

GeoX refers to a family of models, frameworks, and benchmarks for geometric, geospatial, and cross-view reasoning in vision-language systems. The term encompasses specialized vision-LLMs for geometry problem-solving, self-play reinforcement learning approaches for image-grounded spatial logic, large-scale 3D representation learning for autonomous driving, and datasets targeted at cross-view geo-localization and pose estimation.

1. Geometric Problem Solving via Vision-Language Pre-Training

GeoX in the context of geometric problem solving designates a multi-modal large model specifically architected for automatic Geometry Problem Solving (GPS). The model ingests a plane-geometry diagram and a textual question, and outputs a formalized solution—an executable program—interpreted by an external symbolic solver. Addressing the domain gap between natural images and geometric diagrams, GeoX introduces a three-stage formalized vision–language pre-training pipeline: (1) unimodal pre-training of specialized diagram encoder (Geo-ViT) and language decoder (Geo-LLM), (2) geometry-language alignment using a Generator-and-Sampler Transformer (GS-Former), and (3) visual instruction tuning on diagram–question pairs for interpretable, verifiable output (Xia et al., 2024).

Key architectural elements include:

• Geo-ViT: a ViT-B encoder-decoder trained using masked autoencoding on 120k diagrams, focusing on robust representation of monochrome, symbolic geometric visual data.
• Geo-LLM: a decoder-only LLM (LLEMMA-7B backbone), fine-tuned over 100M geometry-focused tokens, to reliably generate formal solution programs.
• GS-Former: a module aligning diagram features with formal geometric captions, employing sparsification losses and dynamic query generation for selective attention to key diagrammatic features.

The end-to-end pipeline concatenates geometry-aware visual embeddings with question tokens, and auto-regressively generates a formal solution program, achieving state-of-the-art accuracy across multiple benchmarks (e.g., GeoQA, UniGeo, Geometry3K, PGPS9K) with performance gains over both generalist and specialist baselines.

2. Geospatial Reasoning through Self-Play and Verifiable Reward

A distinct instantiation of GeoX addresses geospatial reasoning: the framework models geospatial logic acquisition as programmatic self-play, using automatically generated, executable spatial reasoning problems over satellite or aerial images (Ahn et al., 19 May 2026). The core of this approach is a single multimodal policy $\pi_\theta$ that alternately proposes and solves spatial problems without reliance on annotated QA data:

• Proposer role: constructs executable spatial programs $(p, a)$ combining spatial function primitives (e.g., area, centroid, argmax) and segmentation tools, and computes the outcome $o = p(a;I)$ via a verifier.
• Solver role: given (possibly masked) elements of $(p, a, o)$, solves abduction (infer input), deduction (predict output), or induction (synthesize a consistent program), with correctness automatically evaluated by the verifier.

Rewards are precisely quantified through execution: exact-match for abduction/induction, type-aware scores for deduction (e.g., normalized error for numeric outputs). Proposer rewards are shaped for problem learnability, peaking for intermediate difficulty. Reinforcement learning jointly optimizes proposer and solver parameters, yielding consistent accuracy improvements (up to +5.5 pp) on held-out geospatial VQA benchmarks—without any curated data.

The GeoX self-play process constructs extensive, auto-verified geospatial reasoning benchmarks covering nine reasoning dimensions and seventeen downstream tasks, expanding evaluation options beyond manually written questions.

3. Geometry-Centric Scene Understanding for Autonomous Driving

LiAuto-GeoX extends the GeoX paradigm to ego-centric 3D scene understanding for autonomous driving (Lian et al., 4 Jun 2026). It introduces a real-time, efficient framework for dense, grounded 3D reconstruction from surround-view imagery, leveraging sparse LiDAR depth as geometric priors during training:

• Teacher Model ($F^T$): a 24-layer Vision-Geometry-Grounded Transformer trained on large-scale surround-view data plus LiDAR depth supervision to produce per-view depth maps, fused point clouds, and geometry tokens.
• Student Model ($F^S$): a compact 155M-parameter transformer (DINOv2-Small encoder plus 12-layer decoder), distilled via two novel regularizers:
  • Mask-Guided Depth-Aware Distillation: injects depth-aware, geometry-sensitive biases using teacher activation masks, focusing capacity on object boundaries and ambiguous regions.
  • Relative-Pose Relational Distillation (RPR): enforces cross-view geometric consistency by matching pairwise camera-conditioned token relations among views.

The distillation framework is two-stage: first, training the teacher with full supervision (including LiDAR); second, distilling knowledge into the student with geometry-preserving objectives. At inference, only the student model operates, achieving 223.8 FPS on KITTI with sub-meter 3D reconstruction accuracy (DDAD Acc=1.012m, Comp=1.174m). The reconstructed geometry is natively transferable to downstream tasks such as trajectory prediction (PDMS=90.6), occupancy prediction (mIoU=24.63), and future-frame prediction (IoU=47.67), making it a scalable backbone for modern autonomy stacks.

4. Cross-View Geo-Localization and Pose Estimation Benchmark

GeoX-Bench constitutes a large-scale, multimodal benchmark targeting the cross-view geo-localization and pose estimation capabilities of large multimodal models (LMMs) (Zheng et al., 17 Nov 2025). The dataset comprises 10,859 aligned ground-satellite image pairs sampled from 128 cities globally, and 755,976 QA pairs systematically covering:

• Geo-localization (ground-to-satellite and intra-map retrieval): top-K accuracy of identifying the correct satellite tile from a gallery.
• Pose estimation: discrete heading classification and continuous orientation error, including geodesic distance, with tasks varying in the spatial configuration of the ground view relative to the satellite center.

Benchmark results show that while recent LMMs (e.g., GPT-4o, Gemini-2.5-Pro, o3) achieve 80–90% localization accuracy with or without priors, pose estimation remains a fundamental weakness (≤30% for open-source, ≤50% for closed-source, fixed-center). Instruction-tuning with GeoX-Bench QA pairs significantly boosts performance (+20–30% in pose estimation).

Limitations include heading quantization (four bins only), possible temporal/visual misalignment, and underrepresentation of certain terrains. Future extensions call for integration of geometric priors, continuous orientation regression, expanded environmental diversity, and advanced multimodal fusion architectures.

5. Experimental Insights, Ablations, and Evaluation

GeoX models and datasets consistently demonstrate that domain-specific representation, geometry-focused pre-training, and structured programmatic outputs are critical for geometric and geospatial machine reasoning:

• Domain-adapted encoders and instruction-tuned decoders outpace generalist MLLMs (e.g., GeoX GeoQA Top-1: 54.9% vs GPT-4V 43.4%).
• Geometry-language alignment and sparsification regularizers (in GS-Former) enable selective focus and verifiable, interpretable solutions.
• Self-play geospatial models autonomously generate and verify a combinatorial space of reasoning problems, while maintaining or exceeding accuracy of data-hungry baselines.
• In autonomous driving, geometry-preserving distillation delivers compact, real-time models with high-fidelity 3D reconstruction and effective transfer to downstream planning tasks.
• GeoX-Bench reveals that while large-scale training confers robust spatial localization, orientation invariance and precise heading reasoning are not yet solved.

Ablation studies affirm that the specialized architectural components (e.g., GS-Former in geometry, mask-guided distillation in 3D learning) produce non-trivial gains (e.g., Geometry3K Completion: full alignment/GS-Former 58.6% vs no alignment/GS-Former 33.1%) (Xia et al., 2024, Lian et al., 4 Jun 2026).

6. Limitations and Prospects for Future Research

Current GeoX approaches, despite substantial advances, retain several open issues:

• Dependence on external symbolic solvers in geometry; integration into end-to-end differentiable pipelines remains aspirational.
• Manual annotation bottlenecks for formal relationship descriptions in geometric diagrams.
• 3D and transformation geometry extension, automatic mining of formal captions, and learning from partially labeled or automatically paired data are active research directions.
• In geospatial reasoning, further work is needed in continuous pose regression, robustness to environmental and temporal shifts, and richer multimodal logic.
• Scaling instruction-tuning and integrating explicit geometric priors are recommended pathways according to benchmark findings (Zheng et al., 17 Nov 2025).

GeoX, in its various specialized forms, establishes a suite of architectures, learning paradigms, and resources addressing fundamental algorithmic and representational bottlenecks in machine geometric, geospatial, and cross-view reasoning.