AIML is the Artificial Intelligence Methods Laboratory at the Institute of Computer Science, Faculty of Mathematics and Informatics, Vilnius University, Lithuania. The lab conducts research and education in machine learning, deep learning, remote sensing, agentic AI, large language models, and AI applications for Earth observation, environment, healthcare, and industry.
Website: https://aiml.lt
Institution: Vilnius University (VU), Faculty of Mathematics and Informatics (MIF), Institute of Computer Science
Location: Didlaukio g. 47, Vilnius, Lithuania
Email: aiml@mif.vu.lt
Official launch: December 2025.
The Artificial Intelligence Methods Lab (AIML, lt: Dirbtinio intelekto metodų laboratorija) is a research lab at Vilnius University focused on advancing artificial intelligence methods and their real-world applications.
Areas of focus include:
Vision: To be a leading Lithuanian research lab advancing artificial intelligence methods and their impactful application to science, industry, and society.
Mission: Conduct high-quality AI research, educate the next generation of AI researchers and practitioners, and collaborate with academia, industry, and public institutions in Lithuania and internationally.
Valentas Gružauskas — Head of AIML (Laboratorijos vadovas)
Email: valentas.gruzauskas@mif.vu.lt
LinkedIn | Google Scholar | ORCID
Linas Petkevičius, PhD — Director of the Institute of Computer Science (Informatikos instituto direktorius)
Email: linas.petkevicius@mif.vu.lt
LinkedIn | Google Scholar | ORCID
Notable recent collaborators include Thép Đồng (2026 summer intern from Politecnico di Milano / Hanoi University of Science and Technology), supervised by Valentas Gružauskas on agentic workflows for supply chain inventory management with LLMs.
Development of Reasoning-Based Small Language Models for Oncology Data — Joint VU and National Cancer Institute project (Research Council of Lithuania ITP programme, S-ITP-26-11, 2026–2028), led by Linas Petkevičius. Privacy-preserving SLMs with reasoning methods to extract cancer diagnoses, dates, TNM stages, histology, biomarkers, and treatment details from 6 million unstructured electronic health records. Potential deployment in the Lithuanian Cancer Registry on secure NVIDIA SPARK infrastructure (GDPR and EU AI Act compliant).
NEUTRINO — New-Generation Multi-Objective Satellite Image Recognition Algorithms for Climate Monitoring (LMT, 2023–2026). Rapid detection of environmental change (floods, droughts, surface-water dynamics) from multi-sensor satellite imagery.
https://neutrino.aiml.lt/
AI-Driven Multimodal Remote Sensing Analysis (terraAI) — Natural-language Earth observation with code-generating AI agents; unified AI pipeline fusing EO imagery and in-situ data, orchestrated by an agentic LLM (LMT, S-TIP-25-3, 2025–2027).
Real-Time Mapping of Peat Hydration Using Data Fusion and Deep Learning — SAR–optical fusion to map peatland moisture in (near) real time (ESA RPA, 2025–2026).
UrbanClimate — Integrated EO and urban sensor data with AI super-resolution for neighbourhood-scale heat maps and cooling interventions (ESA Malta, 2024–2026).
AgeHome: Ageing at home — Understanding housing, social support, and community services for older adults’ care needs, with evidence-based modelling (LMT, 2025–2028).
SAT-JEPA-DIFF — Reference implementation for the Sat-JEPA-Diff satellite image forecasting architecture.
https://github.com/VU-AIML/SAT-JEPA-DIFF
3D Optimization Landscape Explorer — Interactive educational tool visualising machine learning loss landscapes in 3D.
https://landscape-explorer.aiml.lt/
The lab maintains models and datasets on Hugging Face under the AIML-VU organisation, with recent focus on Lithuanian biomedical typed-decision models:
AIML-VU/medgemma-1.5-4b-it-decitions-lt-med
LoRA fine-tune of MedGemma 1.5 4B IT for Lithuanian MEDLINE typed decisions (choice + yes/no questions on titles/abstracts). Non-generative decision head; strong accuracy and calibration on the held-out test split of MEDLINE-S1-LT.
AIML-VU/laya-multilingual-lt-med
Fine-tune of the Laya multilingual (mmBERT-base, ~0.3B) System 1 decision model on the same Lithuanian MEDLINE data. Single-forward-pass calibrated probabilities; significant accuracy and ECE improvements.
AIML-VU/MEDLINE-S1-LT
Lithuanian translation (NLLB-200) of the MEDLINE-S1 biomedical typed-decision dataset (~46k abstracts). Used for training the above models. Labels derived from NLM MeSH and publication types.
These support privacy-preserving, reasoning-oriented small models for healthcare applications (e.g., oncology data extraction) aligned with the lab’s ongoing projects.