Control Point Detection for Network Alignment Using Large Language Model
Workshops of 26th International Conference on Computational Science and its Applications, ICCSA 2026, Braga, Portekiz, 30 Haziran - 03 Temmuz 2026, cilt.16766 LNCS, ss.365-373, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 16766 LNCS
- Doi Numarası: 10.1007/978-3-032-30539-8_28
- Basıldığı Şehir: Braga
- Basıldığı Ülke: Portekiz
- Sayfa Sayıları: ss.365-373
- Anahtar Kelimeler: Alignment, Coordinate Transformation, GeoAI, Georeferencing, Large Language Model, Road Network
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Control point (CP) selection is a crucial preprocessing step in network alignment workflows in GIS and spatial data integration. CP-based registration commonly depends on selecting corresponding intersections between a source network and a target network, after which a geometric transformation such as similarity, affine, projective, or locally adaptive warping can be estimated. This paper evaluates GPT-5.2 “thinking” mode, a multimodal Large Language Model (LLM) for CP detection and correspondence establishment between two network inputs that represent the same geographic extent. We analyze twelve experimental runs organized by three factors: input modality as rasterized network imagery versus vector networks, availability of explicit intersection node markers and identifier labels versus line-only inputs, and distortion level of the target network including regular affine-like distortions, globally irregular distortions, and local distortions. The results indicate that regular affine-like deformations enable strong LLM-assisted CP detection, particularly when nodes are provided and the raster/vector representation is chosen appropriately. This advantage diminishes under irregular distortions (global or local), where achieving consistently good performance becomes more difficult in vector-based settings. Finally, when deformation exists by localized detail-level differences, neither raster nor vector yields reliable results, making local distortions a modality-agnostic limitation for the current LLM-assisted CP detection setup.