Research track
All 10 directions are included in EVA’s first-cohort pilot. Teams choose a focused question, test a hypothesis and develop a research contribution with a supervisor.
Choose a research project instead of an own-product or corporate case project. Teams have 3–5 members. Before research begins, we agree the exact question, supervisor, data access, resources and success criteria. Supervisors are not yet announced; laboratory topics depend on a suitable partner.
Outputs may include a literature review, a permitted dataset, method, prototype, experiment and manuscript. Null or negative findings are valuable when the study is sound.
/ 01All pilot topics
Each direction needs a specific testable hypothesis. Novelty must be assessed against existing literature and prior solutions.
R01
How can people retain control over multi-step agents?
Compare checkpoints, least privilege and error detection in an isolated environment.
What we measure: Task success, permission violations and human intervention time.
R02
How can we verify that a research claim is supported by its source?
Study claim-to-source verification, citation quality and abstention when evidence is insufficient.
What we measure: Source support, fabricated citations, completeness and verification cost.
R03
How can useful AI require fewer resources?
Compare small and large models, adaptive routing and context processing under comparable budgets.
What we measure: Quality, cost per correct answer, latency and energy where measurement hardware is available.
R04
How can AI assistance support lasting independent skills?
Compare guided hints with ready-made answers, followed by a delayed unassisted test.
What we measure: Skill transfer, unassisted performance and learning time. Requires an approved protocol, participant consent and an adequate sample.
R05
How can we respect data rights and prevent leakage?
Evaluate source and license traceability; an alternative project tests RAG access separation using synthetic data.
What we measure: Traceability, license conflicts, cross-user leakage and audit cost.
R06
How can content provenance survive transformations?
Study how image provenance verification survives compression, cropping and republication.
What we measure: Verification retention, robustness and false alarms. Provenance does not prove that content is true.
R07
How can applications avoid dependence on a single model provider?
Compare migration across cloud providers and a local model, including simulated outages in a test environment.
What we measure: Quality after switching, recovery time, manual adaptation and migration cost.
R08
How can people be supported as occupations change?
Compare retraining and support scenarios for one occupation using public data.
What we measure: Scenario sensitivity, income distribution and policy cost. Task exposure does not equal job loss.
R09
How can robots learn when real-world data is scarce?
Test how simulation diversity affects the robustness of robotic grasping.
What we measure: Task success, robustness and real-data requirements. Physical-world transfer requires a laboratory partner.
R10
How should AI startups be evaluated and funded as technology changes?
Study capital intensity, inference cost, model dependence, durability of advantages and follow-on funding needs.
What we measure: Cost scenarios, runway, dilution and sensitivity. Unavailable financial data must not be invented.
/ 02Process
Weeks 1–2
Agree a topic, supervisor and research question. Review prior work and clarify rights, data and budget.
Weeks 3–4
Prepare the protocol, evaluation criteria and data. Reproduce a baseline and begin prior-art search for a potential technical invention.
Weeks 5–8
Run experiments, repeated evaluations and error analysis. Record model versions, parameters and execution conditions.
Weeks 9–10
Analyse uncertainty and limitations. Obtain expert review and decide on IP protection before public disclosure.
Weeks 11–12
Prepare a report, reproducible package and manuscript where results support it. Use a closed defense if needed; journal review may continue after the cohort.
Plan for 8–12 hours per week. Experiment complexity, adequate sample sizes and publication timelines are agreed separately; research may continue after the cohort.
/ 03Research output
The aim is an original manuscript for an appropriate journal on the Russian VAK list or indexed by Web of Science / Scopus. These are distinct criteria; journal scope, specialty and active coverage must be checked before submission. Acceptance and indexing are not guaranteed.
Prior-work review, a falsifiable hypothesis, baselines, appropriate data and analysis of uncertainty and limitations. References and findings are checked, author contributions recorded and AI use disclosed under the journal’s rules. Submit a manuscript to one journal at a time.
VAK documentsWeb of Science Master Journal ListScopus Sources
/ 04Intellectual property
Results with technical novelty may be considered for patent protection in Russia. This requires a specific technical result, prior-art search and a specialist’s patentability assessment. A patent grant is not guaranteed. Economic or educational research can be valuable without a patent outcome.
Agree contributors’ rights and any university or employer rights before work begins. Assess protection before a preprint, code release or public Demo Day; where appropriate, file first and then agree disclosure. Software or database registration is a separate procedure from a patent. Compute, publication and IP expenses are agreed separately.
Choose a research project in the application