Contest · 2026 · 8th edition
Euro-Information coding contest
“For a responsible AI”: a 52-hour optimisation challenge, solo, with AI tools allowed. The task was to choose which AI models a computing centre should train under an energy cap to maximise their total value, across seven input files of up to 50,000 models.
- 14thout of 177 entrants
- 52 hof competition
- 7input files
The problem, the seven files and the method
From 9 October (8 pm) to 11 October 2026 (11:59 pm). A single problem split into 7 input files: one answer file is submitted per input, and each is worth up to 1,000,000 points, pro rata to the best score in the contest.
Each model needs an amount of data of 4 types (numerical, text, images, code) within a given range. Each dataset can be used only once. A copyrighted dataset halves the model's value. A model that uses a single data type can act as a generator to cover another model's need, and that model then also loses half its value.
- “Example” (5 models, 5 datasets): a tiny test case, solved to optimality (12,946).
- “So it begins.” (3,000 models, 10,000 datasets): copyright-free data is scarce, and models that need only a sliver of data “waste” a whole dataset. Best score: 272,318.
- “Everything is free!” (5,000 models, 20,000 datasets): everything is copyright-free, so it is a pure knapsack problem. Proven optimum (523,758).
- “Those models are demanding!” (8,000 models, 40,000 datasets): very narrow data ranges (10 units at most) and no generators allowed. Each need has to be filled almost to the exact dataset. Best score: 1,031,941, at most 3 points from the optimum.
- “Looks like there's a discount on our electricity bill.” (11,000 models, 100,000 datasets): energy is plentiful; data is the limiting factor. Best score: 3,786,059.
- “Where are the files?” (15,000 models, only 5,000 datasets): most of the value comes from models fed solely by generators, and every dataset counts. Best score: 669,670.
- “We're in the endgame now.” (50,000 models, 250,000 datasets): the largest file, with every mechanism at once. Best score: 5,822,547.
Method. Integer linear programming (HiGHS, Gurobi) to choose the models, then a C++ allocator (local search and exact subset sums) to assign the datasets. Several Claude sessions worked in parallel, each on one file.
Loophole left alone. The official grader accepted negative indices, which counted some models twice. Some competitors took advantage of it; I did not use it, and the organiser later fixed it.