A few days ago, the third edition of the International Artificial Intelligence Olympiad (IOAI 2026) came to a close in Astana, Kazakhstan (2–8 August 2026).
The IOAI is a major event both from an IT perspective and as a sporting competition, particularly in these times. The International Olympiad in Artificial Intelligence (IOAI 2026) in Astana was attended by 440 secondary school pupils from over 50 countries. Most of the major nations fielded a delegation of eight competitors (divided into two sub-teams during the team competition). Each sub-team tackled complex programming and artificial intelligence challenges, putting their skills to the test. Specifically, in the Individual Contest, participants tackled six practical tasks (tasks 1–3 on the first day, tasks 4–6 on the second). The challenges did not simply require theoretical concepts, but the actual writing and optimisation of machine learning models:
- Computer Vision (CV): tasks centred on visual recognition and image processing (e.g. segmentation, classification of atypical or noisy images, object tracking).
- Natural Language Processing (NLP): problems relating to text comprehension and generation using language models (LLMs, sentiment analysis, classification or specialised translation).
- Machine Learning & Data Science: predictive algorithms, hyperparameter optimisation and the management of complex or imbalanced datasets.
- Refinement and Prompt Engineering/Fine-Tuning: the creation of specific ‘anti-brute-force’ solutions capable of running and training models within tight timeframes and with limited computing resources.
For those who are interested, the rules for taking part can be found online here, whilst a collection of photographs can be found here.
The medal table
The winner of the competition was Russia, followed by Kazakhstan and China; specifically:
| Location | Name | Country | Score |
|---|---|---|---|
| 1 | Artem Gorokhov | Russia | 428.49 |
| 2 | Dauzhan Beketov | Kazakhstan | 405.58 |
| 3 | Shao Zix | China | 375.45 |
It is worth noting that Gorokhov was the only participant to achieve a perfect score of 100/100 on Task 6. Gorokhov is active in the open-source community on GitHub (username gotheartem), where he shares repositories relating to machine learning solutions, data science competitions and projects on Large Language Models (LLMs) and computer vision. He has taken part in several hackathons and competitive AI challenges (including those organised by Yandex and AIIJC on language and vision models).
Italy took part, achieving the following results:
| Location | Name | Country | Score |
|---|---|---|---|
| 102 | Andrea Rosso | Italy | 112.5340 |
| 120 | Samuele Benassai | Italy | 94.1288 |
| 121 | Fabrizio Giacomi | Italy | 93.4160 |
| 166 | Fabio Cigaina | Italy | 65.3077 |
The medal tally worked as follows:
- Gold medal for those ranked 1st to 37th
- Silver medal for those ranked 38th to 110th
- Bronze medal for those ranked 111th to 220th
Andrea Rosso, therefore, won a silver medal, whilst Benassai, Giacomi and Cigaina each won a bronze. There were also honourable mentions awarded to certain students who, although they did not win, distinguished themselves in specific tasks.
Click to expand the table
| Location | Name | Country |
|---|---|---|
| 221 | Vadim Vozmitel | Belarus |
| 222 | Ayan Zeeshan | Pakistan |
| 223 | Varun Sharma | India |
| 224 | Rastsislau Svechnikau | Belarus |
| 225 | Artyom Hovhannisyan | Armenia |
| 226 | Aivar Tynchtykbekov | Kyrgyzstan |
| 227 | Li Chunting | Australia |
| 228 | Gurgen Bayburdyan | Armenia |
| 229 | Troi Qendro | Albania |
| 230 | Soteris Savva | Cyprus |
| 231 | Raphael Nguebou Newa-Kamga | Benin |
| 232 | Wang Suilun | Macau |
| 233 | Kouonang Jules Simon | Cameroon |
| 234 | Kadirbergenov Atabek Tangirbergenovich | Uzbekistan |
| 235 | Felipe Rivera Durango | Ecuador |
| 236 | Prasiddha Mainali | Nepal |
| 237 | Mugisha Pacifique | Rwanda |
| 238 | Nazar Korshomnyi | Ukraine |
| 239 | Ilknur Yaren Karakoc | Germany |
| 240 | Abdurakhimov Asliddin Adkhamjon Ugli | Uzbekistan |
| 241 | Vincenzo Ribeiro da Cunha Caserta | Portugal |
| 242 | Cesar Murat Cepeda Beltran | Mexico |
| 243 | Luka Jovanovic | Serbia |
| 244 | Varun Aditya Agarwal | Spain |
| 245 | Kaloyan Kalinov Ivanov | Bulgaria |
| 246 | Abedalaziz (Mohammad Saeed) M. Alzghoul | Jordan |
| 247 | Mohamed Nimaga | Mali |
| 248 | Andrii Bilan | Ukraine |
| 249 | Sebastian Hugo Ochoa Machaca | Peru |
| 250 | Emma Sotomayor Cepeda | Mexico |
| 251 | Kareem Imad Omar Kayid | Palestine |
| 252 | Artyom Konukhov | Cyprus |
| 254 | Itay Karny | Israel |
| 255 | Hector Villanueva Davila | Spain |
| 256 | Abdallah Mohammed Mahmoud Sherbini | Palestine |
| 257 | Rukundo Promises | Rwanda |
| 258 | Davit Ghazaryan | Armenia |
| 259 | Albert Rodrigo Alvarez Jove | Peru |
| 260 | Giorgi Tskitishvili | Georgia |
| 261 | Pitambar Lamichhane | Nepal |
| 262 | Julian Berg-Larsen | Norway |
| 263 | Nyan Lin Pyae | Myanmar |
| 264 | Alexandar Lyubomirov Slavov | Bulgaria |
| 265 | Kodiraliev Abrorjon Ikboljon-Ugli | Uzbekistan |
| 268 | Sebastian Enrique Somoza Portillo | El Salvador |
| 269 | Edith Gateretse | Rwanda |
| 271 | Roberto Lloréns Dueñas | Spain |
| 272 | Chen Vincent Yingxi | Australia |
| 275 | Fanny Chryssie Irishura | Burundi |
| 291 | Sun Qiyuan | South Africa |
| 301 | Denagama Vidanelage Chanitha Sendinu Denagama | Sri Lanka |
| 307 | Aayush Yadav | Nepal |
| 309 | Danil Vorobyev | Israel |
| 310 | Tan Zen Ee | Malaysia |
| 313 | Maamoune Ben Ameur | Tunisia |
| 314 | Rafael da Silva Pulcinelli | Brazil |
| 326 | Kunakorn Chaiyara | Thailand |
| 327 | Jesus Alejandro Aguirre Vasquez | Venezuela |
| 329 | Murodjon Khairakov | Tajikistan |
| 338 | Arseni Ivanou | Latvia |
| 343 | Khalil Théo | Lebanon |
| 350 | Hsu Wutt Yee Lin | Myanmar |
The challenges
As already mentioned, the challenges have been spread over several days; here is some useful information.
Day 1
- Find the Order. The aim was to reconstruct the correct chronological order of an audio dialogue in English between two speakers. Each conversation was divided into several audio clips (.wav) that had been shuffled at random; only the first two clips were provided to indicate the start of the track. Contestants had to combine speech recognition and audio representation models (Whisper, wav2vec 2.0) with language models (LLMs such as Qwen) to analyse semantic coherence and the flow of speech. It was one of the most challenging problems on the first day.
- Robot Chasing. The challenge involved predicting trajectories and controlling the pursuit of a target by an autonomous robotic agent. Using image sequences and motion sensor data, the algorithm had to estimate the future position of an object or robot in continuous motion. It had to manage the accumulation of error in temporal predictions and balance high-frequency vision models with real-time state-tracking filters.
- Potato (Potato Dilemma). This was a complex modelling and classification problem involving structured/tabular data that was either imbalanced or affected by noise. It was the task that recorded the lowest average score on the first day and required advanced techniques of feature engineering, data pre-processing and the calibration of machine learning models under severe memory constraints (GPU limited to ~16 GB VRAM).
Day 2
- Double Agent Dilemma. The task involved identifying anomalous behaviour or ‘double-dealing’ within the dynamics of interaction between agents; uncovering hidden patterns and inconsistencies in decision-making within simulated cooperative/competitive environments. It was the task with the highest average score in the entire tournament, with several perfect scores (100/100) also recorded.
- Ghost of the Machine. The aim was to isolate or reconstruct signals and structures from heavily degraded and distorted ‘ghost’ images or data, and thereby develop models capable of removing complex visual artefacts or synthetic noise in order to recover the original information or object free from alterations.
- IOAI Field. The aim was to optimise the behaviour of agents within a virtual environment (a ‘field’ with obstacles and dynamic rules) and thus develop efficient reinforcement learning policies or heuristic search algorithms to maximise the score in the fewest possible steps. It was the task with the fewest ‘zero scores’, enabling many students to accumulate valuable points towards the overall ranking.
Echoes of war
Any readers who have interpreted these challenges in a military context are not mistaken; indeed, it could be stated without fear of contradiction that many of the techniques addressed in these challenges directly reflect the priorities of contemporary technological and military development. Although academic competitions are designed to drive theoretical research and software engineering, the mathematical and machine learning principles applied are entirely ‘dual-use’ (suitable for both civilian and military purposes). We could therefore reconsider the challenges listed above within a military context and attempt to assess their impact on the theatre of operations.
- Robot Chasing (tracking and pursuit): this is at the heart of tracking systems for autonomous drones and loitering munitions (suicide drones). It enables a platform to track a moving target (such as a vehicle) even in the event of visual interference or a temporary loss of the GPS signal. There has been much discussion of this in recent years in the context of the Russia–Ukraine conflict.
- The Double Agent Dilemma (game theory and countermeasures): it is applied in information warfare, cybersecurity and counter-intelligence. It helps to identify bots, agents infiltrating communication networks, or anomalous behaviour within command and control systems.
- Ghost of the Machine (signal and image reconstruction): it is used directly in reconnaissance and surveillance (ISR – Intelligence, Surveillance and Reconnaissance). It enables the clearing of satellite images or drone footage obscured by smoke, fog, camouflage or optical/electromagnetic interference.
- IOAI Field (navigation and autonomy in complex environments): this is essential for the autonomous operation of swarms of drones or unmanned ground vehicles (UGVs) that need to navigate hostile terrain, avoid obstacles and reorganise themselves in the field without direct human intervention.
- Find the Order (audio processing and wiretaps): this is used by intelligence agencies for SIGINT (Signals Intelligence) to analyse, reorder, transcribe and summarise large volumes of environmental or radio intercepts that have been interrupted or jammed.
- Potato Dilemma (optimisation of complex and noisy data): it is used in military logistics, in the predictive analysis of supply flows and in real-time modelling of operational risks.
Conclusions
The Artificial Intelligence Olympics in Astana were not only a showcase of extraordinary scientific and engineering excellence but also a clear snapshot of the rapid pace at which the field is advancing globally. The competitors tackled practical and extremely complex problems: from real-time optical tracking to the processing of degraded signals, right through to the management of complex cooperative and competitive environments. These are challenges that require a level of insight and technical mastery which, until just a few years ago, was the exclusive preserve of leading industrial and university research laboratories.
The picture that emerges is one of profound ambivalence: almost all the machine learning algorithms and architectures developed during these trials fall squarely within the category of dual-use technology. The very same solutions capable of optimising the navigation of drones for emergency deliveries, reconstructing damaged medical images or analysing conversations to protect a country’s cyber security can also be used for autonomous weapon systems, intelligence interception and electronic warfare tactics on an unprecedented scale. This inherently ambivalent nature of AI makes the boundary between civilian progress and military application extremely blurred and indistinct.
It is worth giving these implications careful thought, because the consequences in the future could be far from negligible.
