Event
At the IOCMA 2026 we presented our latest research on Graph Neural Networks: we convert graph-based datasets into more expressive hypergraph structures, thereby addressing the so-called "over-squashing" problem, a central hurdle when reliable learning is required from highly interconnected data such as supply chains or social networks. In initial benchmarks, our approach improves the predictive accuracy of common models in up to 80% of test cases.
Here the link to the poster
Event
Liliya Imasheva presented a validation pipeline for evaluating AI summaries at Conf42 Large Language Models 2026.
Event
Perelyn presented a method at Helmholtz AI Conference for researching cultural values in Large Language Models, which not only reads the models' responses but directly examines their internal representations.
News
The contributions from the "Knowledge Management" workshop at KI2025 have been published as a collected volume. Among them is a paper by our team on a novel knowledge graph-based RAG method, developed to provide accurate, verifiable answers with fewer errors and lower costs.
