Hello CodeForces!
My name is Max (msporyshev), I’m head of motion planning team at Yandex Autonomous Vehicles. We create neural networks that drive an autonomous car.
It happened that this year I lead Machine Learning track of Yandex Cup. And this year we have tried to significantly enhance this competition! The qualification started on October 15 and will last until November 5.
Problems
For the qualification round, we have prepared three very challenging and interesting problems. They are all connected to our bleeding-edge AI technologies. To get ahead of other participants, it might be useful to read up-to-date research papers on the relevant topics, and maybe you'll even end up writing your own!
- Generating realistic images for self-driving car cameras based on a vector description of the scene. This is a small step towards world modelling, which is one of the most relevant research topics in robotics right now.
Hallucinations in LLM models. You are required to train a model to respond well to fact-based queries, ensuring that it remains robust to paraphrased questions and is able to answer «I don’t know» to provocative questions that have been carefully selected to deceive even the most advanced models. All data is completely hidden from the participants, so you will have to work on the dataset on your own.
A VLM problem on the understanding of diagrams for geometry and physics problems. Describing diagrams for STEM problems is still a major challenge even for the largest language models. In this problem, you will need to deal with resource limitations. You will also not have access to publicly available test data.
Qualification setup and infrastructure
All participants will submit their solution code, not answer sheets, almost like in algorithmic competitions! However, you'll have to submit files with large 10-gigabyte models and their inference code, which we'll run on our GPU servers using private data and show your score. You won’t be able to run the public leaderboard test data through ChatGPT 5 ;)
Of course, all this required us to put a lot of effort into the platform infrastructure: Yandex Cup will be holding a competition in this format for the first time. Just imagine running inference of models with billions of parameters on GPU servers, which takes hours. Our developers have done a great job to launch such a platform! Some of you may know that kaggle and other competitive ML platforms already do such formats for some time. For us it is better late than never :)
Final
We will invite the best participants in each problem to the final in Istanbul on December 6! The final won’t be easy. We will hold a 5-hour ML contest, and the winners of the ML track will be determined based on its results!
From me
I recommend all of you to read and try to solve all the qualification problems! We’ve tried to make them interesting and relevant, so I think you’ll have a great time. I also invite school students to participate — they will have a separate leaderboard!
Good luck!
