VK presented three student projects from EMS — from a multimodal benchmark to text analysis

VK discussed three projects by students of the Engineering and Mathematics School (IMS), which the company is developing together with HSE University. All three cases sit at the intersection of science and applied development: from evaluating multimodal models in Russian to reducing the cost of language architectures and analyzing psychological states through text.

What Projects Were Showcased

The main emphasis in the selection is on developments that solve quite practical problems rather than look like educational demos. The first project is a Russian-language benchmark for evaluating visual-language models. Its author tackled a problem familiar to almost all teams testing multimodal models in a local context: English-language datasets poorly reflect Russian realities, cultural references, and historical background. As a result, VK now has a tool already being used to compare open source models and assess how well they understand Russian-language scenarios.

• Russian-language VLM benchmark for more honest evaluation of multimodal models
• Linear attention mechanism for processing sequences with lower computational costs
• Psycholinguistic analysis system that tracks emotions and states in text

The second project focuses on a fundamental problem of modern transformers: attention mechanisms remain too expensive computationally, and their complexity often grows quadratically. The IMS team is working on a linear bidirectional mechanism that should accelerate training and inference without heavyweight low-level implementation.

The third case is a psycholinguistic analysis tool. It should distinguish direct emotional expression from indirect signals, notice state changes within dialogue, and rely on data annotated by specialists rather than only on general text corpora.

Why the Workshop Is Useful

From the material, it's clear that IMS for VK is not just an educational brand, but a format for early hiring and developing specialists for real tasks. One student after the VLM project joined VK's team, another deepened his research into neural network architectures, a third brought his case to a research paper that took first place in HSE's NIRS competition in psychology. Such results appear not from formal internships, but from the combination of "applied task + mentor from industry + requirement to bring work to working state."

"Those who can improve models' understanding of psychology will gain a

significant expansion of tasks for LLM."

Based on students' descriptions, the workshop's value lies not only in the project topic but also in the process. Participants read primary sources, discuss solutions with VK curators, build their own hypotheses, and work not on toy prototypes but on systems with architecture, data annotation, quality criteria, and potential integration into products. This immediately provides several practical benefits: better understanding of ML architectures, full-stack development experience, the skill to formulate research questions, and material that truly strengthens a resume.

How Enrollment Works

The program is open to students of specialized IT fields at HSE University. Participants can work in workshops and take educational courses from VK, receiving 40,000 rubles monthly in return. Participation itself is free, but places are limited, so enrollment is competitive.

For candidates, it looks like a rare combination: on one hand, an academic environment and research tasks, on the other—a clear path into industry and a chance to work with teams developing search, recommendations, advertising, moderation, and large language models. It's also important that VK showcases such stories publicly not as HR marketing but as a demonstration of concrete results. The article lists other school projects—from adapting neural network algorithms to domestic processors to MLSecOps research.

This sends a signal to the market: companies increasingly view university workshops as a way to do R&D faster than through the classic cycle of "job opening—interview—onboarding." If the model works, the business gets not only new people but also ready solutions for products.

What This Means

For the Russian AI market, this is a good example of how a combination of a university and a large IT company can produce not abstract diplomas but useful tools, research results, and future employees. If there are more such programs, local teams will have more of their own benchmarks, architectural experiments, and applied solutions for Russian language and local scenarios.