AI Insights

Dziyana Valenta, a data scientist at Godel, has recently won a Gold Medal for placing 3rd in the 2026 NeuroGolf Championship on Kaggle.

Kaggle is a global data science and machine learning platform where companies and research organisations host real-world competitions, and participants worldwide compete to build the best solutions.

As members of “The Final Ensemble” team, Dziyana, Christof, Fritz Cremer, Matheus Pedroza, and Tony Li finished in 3rd place.

The 2026 NeuroGolf Championship, hosted by Google software engineer Michael Moffitt, challenged participants to develop highly efficient AI models capable of solving 400 ARC-AGI tasks and visual reasoning puzzles. Designed to assess an AI system’s ability to recognise patterns and tackle unfamiliar problems, the competition ran over three months and featured 400 independent abstract reasoning challenges. Success required not only discovering accurate solutions and generating valid ONNX models, but also optimising those models for minimal memory consumption and compact size, placing a strong emphasis on efficiency alongside performance.

The challenge required participants to create functionally correct ONNX models for 400 unique ARC-AGI tasks while minimising both parameter count and memory usage. To tackle this, the team adopted an agent-based workflow using ChatGPT and Codex, experimenting with different models and reasoning strategies to analyse, solve and optimise individual tasks. Dedicated sessions focused on specific objectives, including improving task accuracy, reducing runtime and resolving bugs. A key challenge was balancing model efficiency with execution speed, as all 400 tasks had to run within a strict 30-minute time limit. As a result, even highly accurate solutions were not viable if they consumed excessive processing time. Team members initially worked independently, developing complementary approaches to the problem before combining their solutions into a final ensemble, which delivered additional performance gains and improved overall scores.

Dziyana initially competed independently in the NeuroGolf Championship, consistently placing between 9th and 15th on the leaderboard among approximately 3,000 teams. Following the competition’s verification process, around 100 teams were disqualified, further highlighting the strength of her performance. Based on her results, she was invited to join a four-person team and combine her solution with theirs. The team included Kaggle competitors ranked #2, #15, #30 and #252 globally, from a pool of around 200,000 ranked participants. To receive an official Kaggle ranking, Dziyana was required to participate in a set number of competitions, but her success in the championship marked only the beginning of her journey. Her performance demonstrated the skill and expertise needed to compete alongside some of the world’s leading data scientists.

The competition was highly research-focused, exploring how abstract reasoning challenges can be represented and solved using exceptionally compact AI models. Beyond the competition itself, the techniques developed have broader implications for the field of artificial intelligence, contributing to research into more resource-efficient AI systems and new approaches to tackling unfamiliar reasoning problems. By demonstrating how complex tasks can be solved with minimal computational resources, the work highlights potential pathways towards more efficient and scalable AI solutions.

“This was an intense and highly competitive experience, and I’m very proud of this achievement and of the contribution I made to our final result. What made the experience even more meaningful was working with such smart and talented teammates. Even under significant time pressure and across different time zones, communication and collaboration within the team were always smooth, open, and enjoyable.”

Dziyana Valenta.

Posted 17 Aug 2026

share