Skoltech hosted an international foresight session on artificial intelligence that brought together researchers, engineers, and industry partners and set an agenda spanning fundamental research through pilot implementations with measurable KPIs. The session was coordinated by the Strategic Agency for the Support and Formation of AI Developments (SAPFIR) as part of the International AI Foresight initiative conducted in 2025.
The foresight exercise included a plenary session and three thematic working groups: architectures, algorithms, and mathematics; foundational and generative models; governance, decision-making, and multi-agent systems. Each group involved international experts, with a total of nine participants from abroad, affiliated with institutions in countries such as China, the United Kingdom, and India.
Opening the session, Grigory Kabatyansky emphasized foresight as a tool to align the fundamental research agenda with industry demand: results should convert into pilots with clear quality metrics and infrastructure requirements. Alexander Panchenko outlined an approach to trustworthy LLMs: multilingual resources, rigorous evaluation protocols, and control of toxicity and factual errors down to the token level — so that models not only “write nicely” but reliably solve applied tasks, from adaptive knowledge retrieval (RAG) to expert support in medicine and engineering. Mirko Farina highlighted reliability practices in responsible applications: verification and validation, documentation of limitations, mandatory human-in-the-loop control at critical steps, and regular stress-testing by an independent “red team.”
In the architecture and optimization group, participants prioritized methods that improve reproducibility and stability under real-world loads and accelerate training and inference for large models. The “Foundation and Generative Models” track focused on multimodal LLMs, standardization of benchmarks and trust metrics, reduction of hallucinations and toxicity, and adaptive knowledge-retrieval pipelines tailored to domain tasks. The control and multi-agent systems group formalized requirements for production-grade AI architectures: quality monitoring and decision traceability, mandatory human oversight at risk-bearing steps, regular stress tests, and operations planning.
The foresight session concluded by consolidating one-year “roadmaps,” which included: a package of benchmarks and trust metrics; a list of pilot projects for industry, energy, transport, and healthcare with defined KPIs for accuracy, robustness, and total cost of ownership; and a description of the requirements for training and inference infrastructure. “Foresight is an excellent opportunity to align scientific priorities with industry needs,” emphasized Skoltech Vice President Grigory Kabatyansky. “We conducted the session following the Ministry of Economic Development’s methodology: in person, across three working groups, and with colleagues holding foreign affiliations — to independently validate approaches and ensure metric comparability with global practice. These results will form the basis for practical deployments and joint analytical materials and will be updated annually together with our partners.”
Skoltech is Russia’s leading center for AI development, uniting fundamental and applied agendas. More than 300 researchers, engineers, and graduate students develop platform solutions for industry, energy, transport, and sustainable development, building on generative models, 3D computer vision, predictive analytics, and signal processing. The Skoltech AI Center leads Russian universities in the number of publications at A* conferences; its algorithms are used by leading companies. Five of the Center’s professors are among the global top 2% most-cited scientists in their fields. Talent development is also central to the Center’s mission: master’s and PhD programs, continuing education courses in AI and robotics, and the international SMILES Summer School.
