Large Language Models (LLMs) for Insurance Intelligence: Automating Multimodal Damage Evaluation and Claim Decisions
10th International Conference on Computer Science and Engineering, UBMK 2025, İstanbul, Türkiye, 17 - 21 Eylül 2025, ss.288-292, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/ubmk67458.2025.11206911
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.288-292
- Anahtar Kelimeler: Gemma LLM, Insurance Claim Automation, Large Language Models (LLMs), Vehicle Damage Assessment, Zero-shot Learning
- Boğaziçi Üniversitesi Adresli: Evet
Özet
Policy creation and claim operations are insurance-related procedures that have direct interaction with customers. It is a quick and uncomplicated approach to fulfill the customer's policy requirements. However, claim processes are extensive and time-consuming. When the insured car is damaged, the customer is asked to provide documentation and photographs. Experts study these photographs and documentation to determine who is responsible for the accident. Additionally, the cost of the car accident is calculated. All of these are lengthy and complicated processes. This paper presents a vision-language large language model (LLM)-based framework for automated assessment of automobile damage and insurance claim decisions. Three vision-language LLMs - Gemma, Llama, and Mistral - were evaluated in this study. Gemma 3-27B parameter model gives the best result among these models. In this way, a faster and more accurate process is managed. Processes below a certain accident amount result in automatic decisions. This amount is valid for minor damages that are low in cost. In this way, the operational costs of insurance companies are reduced. The Gemma 3-27B model outperformed the others in producing contextually accurate and detailed insurance claim reports. However, the more lightweight Gemma 3-12B was used for deployment via local inference for efficiency considerations.