On the occasion of the 200th anniversary of its founding, the Hungarian Academy of Sciences is organizing a number of high-quality events. Each month, a different scientific department will present the history, achievements, and future challenges of the given field. In January 2026, the Department of Mathematical Sciences of the Hungarian Academy of Sciences awaits interested parties with a rich program, lectures, film screenings, and an exhibition that can be visited throughout the month. On the opening day of the program series at the Academy headquarters, on January 5, the topic will be: Artificial Intelligence through a Mathematical Eye. The first lecture will be given by Anna Christ Miranda (University of California – UC Berkeley), a member of the research team at the Rényi Institute, entitled The Structure of Relation Decoding Linear Operators in Large Language Models.
The young researcher graduated from Fazekas High School in Budapest in 2024, and the night before his oral exam, he was still working on the scientific article that he, along with Adrián Csiszárik, Gergely Becsó, and Dániel Varga, presented at the world's most prestigious artificial intelligence conference, the NeurIPS 2025 Spotlight program, in early December. Their study could bring significant progress in understanding the functioning of large language models. During their research, they realized that language models do not store isolated facts, but capture the relationships of the world along interconnected conceptual patterns. In other words, the knowledge of the models is not a collection of independent facts, but a system of overlapping relationships organized into common structures.
This study forms the backbone of Anna Christ Miranda's current lecture, in which she will talk about the foundations for understanding large language models and how relational knowledge can be surprisingly easily modeled with so-called linear relation decoders. (One relation type, for example, is the capital of which country, and several relations belong to this, including that Paris is the capital of France.) Her lecture will reveal how these decoders were collectively modeled – and this is particularly important, since they were previously only examined individually – and how they were interpreted in order to see how knowledge of multiple relation types and multiple different topics is organized in large language models. This is important because it is a key question for the interpretability and reliability of artificial intelligences, in terms of the internal structure of their knowledge.
The young researcher said that the NeurIPS conference was particularly important for him because he could experience the serious interest in his work. He was able to talk to many foreign researchers about how the methods of the four-member research group of the Rényi Institute could be linked to their topics. So, in addition to building relationships, these meetings can also be sources of new ideas and potential collaborations.
The detailed program of the program series of the Department of Mathematical Sciences of the Hungarian Academy of Sciences can be viewed HERE