🧠 MARSAIL Research Highlights (2023–2026)
MARSAIL drives the next generation of automotive AI through foundational transformer research led by Dr. Kao (Teerapong Panboonyuen). Our work unifies damage-centric computer vision, knowledge distillation at scale, and document intelligence into deployable systems.
🌌 MARS — Mask Attention Refinement with Sequential Quadtree Nodes
Introduced the Sequential Quadtree Attention mechanism for precise instance-level damage mask refinement. Demonstrated clear gains over Mask R-CNN, PointRend, and Mask Transfiner — MARSAIL's first official research publication.
🚗 ALBERT — Transformers for Automotive Damage Evaluation (Teacher Model)
Large-scale transformer backbone delivering rich contextual embeddings, multi-level localization, and robust attention signals — the foundational teacher model for downstream automotive AI systems.
🧩 SLICK — Knowledge-Enhanced Vehicle Damage Segmentation (Student Model)
Distills ALBERT's knowledge into a lightweight, production-ready model delivering real-time, fine-grained damage segmentation optimized for insurance workflows and automated repair pipelines.
🔍 DOTA — Enhanced OCR for Automotive Services
Integrates deformable attention with retrieval-augmented generation to achieve human-level robustness in real-world automotive documents — claims, invoices, and inspection reports.
🎓 Final Research Series (2026) — The concluding academic contributions of MARSAIL Lab
🌏 Deconstructing GPT Geopolitical Bias with Fairness in the 2025 Thailand–Cambodia Border Conflict
✅ Accepted — ICCAI 2026 · Okinawa, Japan
Investigates stance distortion in LLMs during geopolitically sensitive events. Introduces GeoFACT, a counterfactually calibrated bias mitigation framework, plus 10K-THAC, a 10,000-statement multilingual dataset.
Research Lead: Dr. Teerapong Panboonyuen
🧠 DABertAttn — Diversity-Aware BertAttention for Reducing the Systematicity Gap in VQA
✅ Accepted — ICCAI 2026 · Okinawa, Japan
Reduces the systematicity gap by 44% on Minimal splits and 60% on Complex splits, demonstrating that architectural sensitivity to attribute diversity enhances compositional generalization beyond data scaling alone.
Research Lead: Dr. Teerapong Panboonyuen
🧩 SPLINT — SParse Learning for INterpretable Tuning
📝 Under Review — ACL Rolling Review
Integrates entropy-guided learning with sparsity-aware regularization to jointly optimize prediction accuracy and interpretability. Outperforms LoRA and AdaLoRA in low-resource and fully supervised settings.
Research Lead: Dr. Teerapong Panboonyuen
🏠 HOMEY — Heuristic Object Masking with Enhanced YOLO for Property Insurance Risk Detection
✅ Accepted — arXiv 2026 · March 19, 2026
Detects 17 property risk categories including structural damages and maintenance neglect. Leverages risk-aware loss calibration to balance class skew and severity weighting for scalable AI-driven property insurance workflows.
🚙 AutoAI-Motor — Foundations and Architectures of AI for Motor Insurance
✅ Accepted — arXiv 2026 · March 19, 2026
Formalizes a vertically integrated AI paradigm for end-to-end vehicle risk assessment and claims processing, introducing domain-adapted transformer architectures for structured visual understanding and multimodal document intelligence.
✍️ Citation
@inproceedings{panboonyuen2023mars,
title = {MARS: Mask Attention Refinement with Sequential Quadtree Nodes},
author = {Panboonyuen, Teerapong},
booktitle = {ICIAP},
address = {Università degli Studi di Udine, Italy},
pages = {28--38},
year = {2023},
publisher = {Springer Nature Switzerland}
}
@article{panboonyuen2025albert,
title = {ALBERT: Advanced Localization and Bidirectional Encoder Representations from Transformers for Automotive Damage Evaluation},
author = {Panboonyuen, Teerapong},
journal = {arXiv preprint arXiv:2506.10524},
year = {2025}
}
@article{panboonyuen2025slick,
title = {SLICK: Selective Localization and Instance Calibration for Knowledge-Enhanced Car Damage Segmentation},
author = {Panboonyuen, Teerapong},
journal = {arXiv preprint arXiv:2506.10528},
year = {2025}
}
@inproceedings{marsail2025dota,
title = {DOTA: Deformable Optimized Transformer Architecture for End-to-End Text Recognition with RAG},
author = {N., N. and Panboonyuen, Teerapong},
booktitle = {KST},
pages = {301--306},
year = {2025},
publisher = {IEEE}
}
@inproceedings{marsail2026geofact,
title = {Deconstructing GPT Geopolitical Bias with Fairness in the 2025 Thailand--Cambodia Border Conflict},
author = {N., N. and Panboonyuen, Teerapong},
booktitle = {ICCAI},
address = {Okinawa, Japan},
year = {2026}
}
@inproceedings{marsail2026dabertattn,
title = {DABertAttn: A Diversity-Aware BertAttention for Reducing the Systematicity Gap in VQA},
author = {N., N. and Panboonyuen, Teerapong},
booktitle = {ICCAI},
address = {Okinawa, Japan},
year = {2026}
}
@article{marsail2026splint,
title = {SPLINT: SParse Learning for INterpretable Tuning},
author = {N., N. and Panboonyuen, Teerapong},
journal = {ACL Rolling Review},
year = {2026},
note = {Under Review}
}
@article{panboonyuen2026homey,
title = {HOMEY: Heuristic Object Masking with Enhanced YOLO for Property Insurance Risk Detection},
author = {Panboonyuen, Teerapong},
journal = {arXiv preprint arXiv:2603.18502},
year = {2026}
}
@article{panboonyuen2026autoai,
title = {AutoAI-Motor: Foundations and Architectures of AI for Motor Insurance},
author = {Panboonyuen, Teerapong},
journal = {arXiv preprint arXiv:2603.18508},
year = {2026}
}
🌟 Fun Facts About MARSAIL
🇺🇸 Fun Facts — English Version
MARSAIL is the dedicated Artificial Intelligence Laboratory under
MARS (Motor AI Recognition Solution).
While MARS was founded earlier as a tech company, the AI-focused MARSAIL was later established and shaped by
Dr. Teerapong Panboonyuen (Dr. Kao),
who leads the AI Research & Development division.
After completing his
Ph.D. at Chulalongkorn University,
Dr. Kao joined MARS in January 2022 to spearhead the next era of AI innovation.
Upon reviewing the existing "first-generation AI system" from 2021, he made a decisive move —
he removed the entire legacy system and rebuilt everything from the ground up.
This total reboot resulted in the new
MARS model
(Mask Attention Refinement with Sequential Quadtree Nodes), inspired partly by the company name.
It became MARSAIL's first official research publication and was presented on the international stage at
ICIAP 2023 in Udine, Italy.
From 2022–2026, under Dr. Kao's leadership, MARSAIL evolved into a next-generation AI research lab,
producing advanced models including the flagship transformer
ALBERT.
Today, MARSAIL powers innovation across
car insurance, damage analytics, intelligent document understanding,
and AI assistance for vehicle damage assessment.
🇹🇭 เกร็ดน่ารู้ — เวอร์ชันภาษาไทย
MARSAIL คือห้องปฏิบัติการปัญญาประดิษฐ์ของบริษัท
MARS – Motor AI Recognition Solution (บริษัท มอเตอร์ เอไอ เรคอกนิชั่น โซลูชั่น จำกัด)
โดยบริษัท MARS มีการก่อตั้งมาก่อน แต่ห้องแลป MARSAIL ถูกสร้างและพัฒนาโดย
ดร.ธีรพงศ์ ปานบุญยืน (ดร.เก้า)
ในฐานะ Head of AI Research เพื่อผลักดันการวิจัย AI ให้เติบโตระดับองค์กร
หลังจบ
ปริญญาเอกจากจุฬาลงกรณ์มหาวิทยาลัย
ดร.เก้าเริ่มเข้ามาทำงานที่ MARS ในช่วง มกราคม 2022 เพื่อสานต่องานพัฒนา AI จากทีมรุ่นแรกในปี 2021
แต่แทนที่จะปรับแต่งแบบเดิม เขากลับตัดสินใจครั้งใหญ่ —
ลบระบบ AI เดิมทั้งหมด และสร้างระบบใหม่ตั้งแต่ศูนย์
เพื่อให้ทันสมัย แข็งแรง และรองรับงานระดับอุตสาหกรรม
โมเดลใหม่ที่เกิดจากการรีบูตนี้คือ
MARS
(Mask Attention Refinement with Sequential Quadtree Nodes) ซึ่งสอดคล้องกับชื่อบริษัท
และได้เป็นงานวิจัยลำดับแรกของห้องแลป MARSAIL ที่ถูกนำเสนอในเวทีนานาชาติ
ICIAP 2023 ณ ประเทศอิตาลี
ตั้งแต่ปี 2022–2026 ดร.เก้า ได้พัฒนา MARSAIL ให้เติบโตเป็นห้องวิจัย AI ระดับแนวหน้า สร้างโมเดลเด่น เช่น
ALBERT
ที่เป็นหัวใจของงานด้านประกันภัยรถยนต์ การประเมินรอยความเสียหาย
การทำความเข้าใจเอกสารอัจฉริยะ และ
AI assistance สำหรับงานประเมินราคาความเสียหายของรถยนต์
ผลักดันอุตสาหกรรมยานยนต์ไทยให้ล้ำหน้าด้วยพลังของ AI
