Jerzy Kamiński
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jerzy Kamiński has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Anna Kalyuzhnaya
ITMO University, St. Petersburg, Russian Federation
2 shared publications - Irina Deeva
ITMO University
2 shared publications - Alexander V. Boukhanovsky
ITMO University
1 shared publication - Andrey V. Savchenko
Sber AI Lab, Moscow, Russian Federation, HSE University, Moscow, Russian Federation, and ITMO University, Saint Petersburg, Russian Federation
1 shared publication - Artem Kuznetsov
Artificial Intelligence Technologies Faculty, ITMO University, Saint-Petersburg 197101, Russia
1 shared publication - Danil Fedorov
ITMO University, St. Petersburg, Russian Federation
1 shared publication
Latest Publications
- Open Access
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1344-1352
Evolving Multi-Agent Systems with Large Language Models: Surrogate Fitness Limitations and Empirical Insights
We investigate the use of large language models (LLMs) as evolutionary operators for optimizing the architecture of multi-agent systems (MAS) - a representative instance of LLM-driven optimization over attributed-graph artifacts that arise in generative design. Given a task description, an initial MAS configuration is generated and then refined on a per-task basis through two optimization strategies: (1) a population-based evolutionary approach where LLMs serve as mutation, crossover, and selection operators, with an LLM-as-a-Judge ensemble providing the surrogate fitness signal; and (2) a simpler iterative regeneration scheme that uses textual feedback from LLM judges to guide single-solution improvement without maintaining a population. We evaluate both approaches on the GAIA benchmark. Our experiments reveal a critical sensitivity to the judge model's quality: a capable judge (Claude 4 Sonnet) yields +14.3 and +6.67 percentage-point accuracy improvements for evolutionary and iterative approaches, respectively, whereas a weaker judge (Gemini-2.5-Flash) provides no benefit or even degrades performance below the unoptimized baseline. Both approaches incur substantial computational costs, raising questions about practical cost-effectiveness. These results highlight fundamental challenges of surrogate fitness functions in LLM-driven evolutionary optimization and offer practical insights for applying evolutionary methods to MAS architecture design.
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1344-1352
Open AccessEvolving Multi-Agent Systems with Large Language Models: Surrogate Fitness Limitations and Empirical Insights
We investigate the use of large language models (LLMs) as evolutionary operators for optimizing the architecture of multi-agent systems (MAS) - a representative instance of LLM-driven optimization over attributed-graph artifacts that arise in generative design. Given a task description, an initial MAS configuration is generated and then refined on a per-task basis through two optimization strategies: (1) a population-based evolutionary approach where LLMs serve as mutation, crossover, and selection operators, with an LLM-as-a-Judge ensemble providing the surrogate fitness signal; and (2) a simpler iterative regeneration scheme that uses textual feedback from LLM judges to guide single-solution improvement without maintaining a population. We evaluate both approaches on the GAIA benchmark. Our experiments reveal a critical sensitivity to the judge model's quality: a capable judge (Claude 4 Sonnet) yields +14.3 and +6.67 percentage-point accuracy improvements for evolutionary and iterative approaches, respectively, whereas a weaker judge (Gemini-2.5-Flash) provides no benefit or even degrades performance below the unoptimized baseline. Both approaches incur substantial computational costs, raising questions about practical cost-effectiveness. These results highlight fundamental challenges of surrogate fitness functions in LLM-driven evolutionary optimization and offer practical insights for applying evolutionary methods to MAS architecture design.
- Open Access
- 1 June 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (6) , 447
Too Many Tools, Too Much Confusion? Navigating Agentic Tool Selection at Scale
This paper addresses the critical scalability challenge that large language model agents face when operating over massive tool repositories. As tool catalogs expand to hundreds or thousands of functions, current architectures exhibit substantial performance degradation caused by semantic collisions between similar tools and ineffective handling of complex multi-tool scenarios. To address these bottlenecks, we propose a recall-first Retrieval–Plan–Select (RPS) framework that combines context-aware query decomposition with synthetic tool description augmentation. The proposed approach explicitly separates retrieval, planning, and final selection through step-local candidate generation, while augmented tool descriptions enriched with expanded summaries and synthetic user questions reduce representation collisions in dense embedding spaces. Evaluation across Ultratool, ToolLinkOS, and ToolRet demonstrates that contextual decomposition consistently improves end-to-end recall under large tool catalogs, increasing recall from 0.340 to 0.494 on Ultratool, from 0.208 to 0.323 on ToolLinkOS, and from 0.300 to 0.347 on ToolRet. Description augmentation further improves retrieval quality, increasing Recall@10 from 0.288 to 0.403 and reducing high-similarity semantic collisions by 41.9% at the 0.90 cosine-similarity threshold. The proposed framework highlights that scalable tool use should be approached primarily as a recall-oriented retrieval and planning problem rather than as a flat in-context selection task, providing practical guidance for building large-scale tool-augmented agents over modern API and MCP-based ecosystems.
- 1 June 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (6) , 447
Open AccessToo Many Tools, Too Much Confusion? Navigating Agentic Tool Selection at Scale
This paper addresses the critical scalability challenge that large language model agents face when operating over massive tool repositories. As tool catalogs expand to hundreds or thousands of functions, current architectures exhibit substantial performance degradation caused by semantic collisions between similar tools and ineffective handling of complex multi-tool scenarios. To address these bottlenecks, we propose a recall-first Retrieval–Plan–Select (RPS) framework that combines context-aware query decomposition with synthetic tool description augmentation. The proposed approach explicitly separates retrieval, planning, and final selection through step-local candidate generation, while augmented tool descriptions enriched with expanded summaries and synthetic user questions reduce representation collisions in dense embedding spaces. Evaluation across Ultratool, ToolLinkOS, and ToolRet demonstrates that contextual decomposition consistently improves end-to-end recall under large tool catalogs, increasing recall from 0.340 to 0.494 on Ultratool, from 0.208 to 0.323 on ToolLinkOS, and from 0.300 to 0.347 on ToolRet. Description augmentation further improves retrieval quality, increasing Recall@10 from 0.288 to 0.403 and reducing high-similarity semantic collisions by 41.9% at the 0.90 cosine-similarity threshold. The proposed framework highlights that scalable tool use should be approached primarily as a recall-oriented retrieval and planning problem rather than as a flat in-context selection task, providing practical guidance for building large-scale tool-augmented agents over modern API and MCP-based ecosystems.
- Open Access
- 1 August 2024
- Published by ITMO University in Scientific and Technical Journal of Information Technologies, Mechanics and Optics
- Vol. 24 (4) , 608
Flexible and tractable modeling of multivariate data using composite Bayesian networks
The article presents a new approach to modeling nonlinear dependencies called composite Bayesian networks. The main emphasis is on integrating machine learning models into Bayesian networks while maintaining their fundamental principles. The novelty of the approach is that it allows us to solve the problem of data inconsistency with traditional assumptions about dependencies. The presented method consists in selecting a variety of machine learning models at the stage of training composite Bayesian networks. This allows you to flexibly customize the nature of the dependencies in accordance with the requirements and dictated characteristics of the modeled object. The software implementation is made in the form of a specialized framework that describes all the necessary functionality. The results of experiments to evaluate the effectiveness of modeling dependencies between features are presented. Data for the experiments was taken from the bnlearn repository for benchmarks and from the UCI repository for real data. The performance of composite Bayesian networks was validated by comparing the likelihood and F1 score with classical Bayesian networks trained with the Hill-Climbing algorithm, demonstrating high accuracy in representing multivariate distributions. The improvement in benchmarks is insignificant since they contain linear dependencies that are well modeled by the classical algorithm. An average 30 % improvement in likelihood was obtained on real UCI datasets. The obtained data can be applied in areas that require modeling complex dependencies between features, for example, in machine learning, statistics, data analysis, as well as in specific subject areas.
- 1 August 2024
- Published by ITMO University in Scientific and Technical Journal of Information Technologies, Mechanics and Optics
- Vol. 24 (4) , 608
Open AccessFlexible and tractable modeling of multivariate data using composite Bayesian networks
The article presents a new approach to modeling nonlinear dependencies called composite Bayesian networks. The main emphasis is on integrating machine learning models into Bayesian networks while maintaining their fundamental principles. The novelty of the approach is that it allows us to solve the problem of data inconsistency with traditional assumptions about dependencies. The presented method consists in selecting a variety of machine learning models at the stage of training composite Bayesian networks. This allows you to flexibly customize the nature of the dependencies in accordance with the requirements and dictated characteristics of the modeled object. The software implementation is made in the form of a specialized framework that describes all the necessary functionality. The results of experiments to evaluate the effectiveness of modeling dependencies between features are presented. Data for the experiments was taken from the bnlearn repository for benchmarks and from the UCI repository for real data. The performance of composite Bayesian networks was validated by comparing the likelihood and F1 score with classical Bayesian networks trained with the Hill-Climbing algorithm, demonstrating high accuracy in representing multivariate distributions. The improvement in benchmarks is insignificant since they contain linear dependencies that are well modeled by the classical algorithm. An average 30 % improvement in likelihood was obtained on real UCI datasets. The obtained data can be applied in areas that require modeling complex dependencies between features, for example, in machine learning, statistics, data analysis, as well as in specific subject areas.
- Open Access
- 15 July 2023
- Published by Association for Computing Machinery (ACM)
- p. 2366-2369
LSevoBN: a structure learning algorithm for large Bayesian networks
The issue of structure learning in large Bayesian networks has gained significant attention from the academic community in recent years due to its wide-ranging applications in various fields. Efficient algorithms for structure learning are crucial for enabling the practical use of Bayesian networks in real-world scenarios where the number of variables can be substantial. In previous approaches, addressing this problem has resulted in either prolonged convergence times or significant degradation of the problem. Although evolutionary algorithms have been effective in training Bayesian Networks (BNs), their application to large BNs has been impractical due to lengthy convergence times. In this paper, we introduce a modified evolutionary algorithm for training large BNs by dividing the network into smaller ones, coding and connecting them through evolutionary optimization. Our proposed approach significantly accelerates the training process, achieving a time acceleration of more than twenty times compared to existing solutions, while maintaining comparable or slightly degraded quality.
- 15 July 2023
- Published by Association for Computing Machinery (ACM)
- p. 2366-2369
Open AccessLSevoBN: a structure learning algorithm for large Bayesian networks
The issue of structure learning in large Bayesian networks has gained significant attention from the academic community in recent years due to its wide-ranging applications in various fields. Efficient algorithms for structure learning are crucial for enabling the practical use of Bayesian networks in real-world scenarios where the number of variables can be substantial. In previous approaches, addressing this problem has resulted in either prolonged convergence times or significant degradation of the problem. Although evolutionary algorithms have been effective in training Bayesian Networks (BNs), their application to large BNs has been impractical due to lengthy convergence times. In this paper, we introduce a modified evolutionary algorithm for training large BNs by dividing the network into smaller ones, coding and connecting them through evolutionary optimization. Our proposed approach significantly accelerates the training process, achieving a time acceleration of more than twenty times compared to existing solutions, while maintaining comparable or slightly degraded quality.