Materials of Dialogue 2026
Anastasia Vladimirovna Kolmogorova
Doctor of Philological Sciences, Professor in the Department of Philology at the School of Humanities and Arts at the National Research University Higher School of Economics in St. Petersburg, Academic Director of the Master’s Program “Language Technologies in Business and Education,” and Head of the Research Laboratory of Language Convergence.
Visual language models in the task of describing paintings
Abstract
This report describes the results of a comparative analysis of two samples: descriptions of 900 paintings from a renowned museum’s collection, created by informants under specially defined experimental conditions, and descriptions obtained from three visual language models, for which similar experimental conditions were set in the prototype. By annotating the obtained data using other language models, we compared which objects, for example, are assigned color by models and people, and which of these two types of “observers” use a more perceptual vocabulary, while which use an interpretive vocabulary, suggesting social symbolism (e.g., uniform or awards) or, for example, abstract terminology. The obtained results allow us to draw some preliminary conclusions about the specifics of the conceptualization of “scenes” by natural and artificial “observers” in the context of the task of describing paintings.
Ruslan Ukrainsky
Head of Research and Innovation, Production Department
Garant From Model to Agent: How Generative AI is Changing Legal Information Management and the Development of Legal Reference Systems
Abstract
Generative AI is changing not only the way we work with legal information but also the approaches to creating legal services themselves. However, experience shows that the LLM model alone is insufficient to address professional challenges that require timely, reliable, and transparent sources.
This report examines the transition from a model as a standalone solution to an agent-based architecture, where LLM is just one component of the system, along with search, memory, personalization, and external knowledge sources. It will demonstrate why the development of LegalTech is not about choosing one “perfect” model, but rather about building flexible intelligent systems capable of utilizing various models and tools.
The second part of the report examines the impact of generative AI on software development using the Iskra AI service as an example. Particular attention is paid to the transition from the traditional approach of detailed technical specifications to the creation of working prototypes and the use of AI agents as a new design and development tool.
The report draws on practical experience in the creation and development of intelligent services for working with legal information.
Alexander Vladimirovich Podobryaev
PhD, Associate Professor at the School of Linguistics of Higher School of Economics
Semantic Competencies of Large Language Models
Abstract
A number of recent papers (e.g., Dentella et al. 2023, Leivada et al. 2023, Kallini et al. 2024, Mahowald et al. 2024, Hardt 2025, Lan et al. 2026) discuss the extent to which the syntactic competencies of modern large language models (LLMs) are close to those of humans. In this talk, we will look at the semantic competencies of LLMs, primarily at how well they cope with the tasks of establishing logical consequence between utterances in natural language. Based on a study of LLM responses to the prompts of a conventional formal semanticist, we will attempt to answer the question of whether LLMs can be considered a theory of natural language (Piantodosi 2023) from a semantics perspective.
