At their current basic capabilities, neural networks can’t handle the tasks associated with science and business: LLMs can’t read the emotional tone of emails, adapt to new conditions, or cooperate with other neural networks to solve a task. These factors directly affect the quality of service they provide, especially in retail, where it’s important to understand consumer moods, quickly adapt to change, and coordinate the work of several services.
As user scenarios get more complicated, the demand grows for smart AI systems capable of solving complex tasks. These are the systems that will be developed at the new laboratory by ITMO and AIRI. Here, researchers will assemble system from multiple agents: virtual assistants that can adapt to changing circumstances (e.g., new user requests) while interacting with one another.
To meet this goal, the lab will work in several areas. The first is multimodal contextual search systems; instead of looking for key words, they search for meaning in text, images, and audio. For instance, they can help retailers select products to fulfill a client’s demand based on the weather, season, or previous purchases. The second area is agent cooperation and adaptation methods; this means that virtual assistants will learn to work as a team: negotiating, dividing tasks, and adapting to each other in real time.
The third area is evolutionary optimization. The researchers will use algorithms similar to biological principles to make agents automatically improve the accuracy of their replies. Scientific knowledge generation will be the fourth focus area. For this, the lab will develop a pipeline of several agents that will work as a research team. One of them will write a literature review, another – come up with experiments, the third one will prepare code for verification, the fourth will present the results in an article, and the fifth one will act as a reviewer.
“We will use LLMs, multimodal systems, and multiagent architectures to make agents communicate with each other, negotiate, and solve tasks together. At the moment, these are cutting-edge technologies – and smart AI systems will be in great demand, especially in business. ITMO has great expertise in developing intelligent systems and training young researchers, while AIRI flourishes in AI research and complex multimodal models. Thanks to this partnership, we get access to powerful computing clusters for LLM training, evolutionary prompt optimization, and resource-intensive experiments,” says Sergey Muravyov, the head of the new laboratory on ITMO’s side.
Particular focus at the lab will be paid to developing multiagent systems for automatic peer review. According to the researchers, existing agents can already handle writing code for experiments, meaning that they can write a program to test a hypothesis. They are also good at prototyping, creating quick drafts of experiments to see if an idea is feasible. However, the greatest challenge is posed by reviewing articles: this task is not just about fact-checking, but also evaluating the research logic, the approach, and conclusions. This means that the agent needs to understand the author’s idea, find subtle errors, and offer constructive criticism. For current LLMs, this task is still too hard. That’s why developing a full-scale reviewer agent will become a separate development area for the lab.
“For quite a while already, AIRI and ITMO have been productively cooperating on multiagent system research. The new laboratory will become a point of attraction for talented students and young scientists who’ll be able to join our work on frameworks for evolutionary multiagent systems and accelerate their practical applications in real-world tasks,” shares Ilya Makarov, a senior researcher and the head of the AI in Industry group at AIRI Institute.
The new laboratory has opened at ITMO’s Kronverksky campus in room 2232, with the team made up of ITMO and AIRI researchers. The lab is open to students of all levels with experience in programming, machine learning, and neural networks. Those looking to join need to send their CVs to Dr. Muravyov (smuravyov@itmo.ru) and pass an interview.
