AIBuildAI achieves first place at the International Olympiad in AI (IOAI)
Across all six IOAI 2025 individual-contest tasks, AIBuildAI achieved a total score of 568.49 out of 600, ranking first on the competition leaderboard. Given only the problem statements and provided data, AIBuildAI autonomously designed, implemented, and optimized a solution for every task within a six-hour budget, with no human intervention.
Recent progress has shown that advanced AI systems can reach gold-medal-level performance on Olympiad-style reasoning tasks. This raises a related question for autonomous AI systems: can an agent also perform well on tasks that require building working AI solutions, rather than only producing final answers?
We evaluate AIBuildAI on the individual contest from IOAI 2025, an Olympiad-style AI competition for pre-university students. The contest contains six AI tasks, with a maximum score of 100 points per task and 600 points in total. The 2025 edition drew 284 competitors; the winning human participant scored 542.05, and the gold-medal threshold was 351.41.
For each task, AIBuildAI received only the problem statement and provided data, then ran autonomously for six hours with no human intervention. We graded its outputs using the organizers' scoring code whenever available. The resulting total of 568.49 places AIBuildAI first on the IOAI 2025 individual-contest leaderboard.
The result across six IOAI tasks
The individual contest spans six substantially different forms of applied AI work. The same autonomous system handled all six, without task-specific tuning or prior exposure to the tasks.
- Radar. Turn radar heatmaps into a map of what is present in each location.
- Chicken Counting. Count animals in a photograph by estimating density.
- Antique. Detect forged paintings from five numerical features, given only four labelled examples.
- Restroom. Match a cropped sign to the corresponding sign from the same restroom.
- Pixel. Keep only 6% of an image while preserving a model's ability to recognize it.
- Concepts. Invent a code so a second AI can infer a secret word.
Figure 1. Agent versus the human field, by task. The chart compares AIBuildAI with the first-place human contestant and the median gold medalist.
First place on the IOAI leaderboard
AIBuildAI produced high-scoring submissions across all six individual tasks. Its strongest result was Restroom, where it matched all thirty held-out queries for a perfect score. It also performed strongly on Radar, Chicken Counting, Antique, and Pixel, and remained competitive on Concepts under the competition constraint.
The important point is not a single peak result, but consistency across very different task formats: perception, counting, visual retrieval, image-crop search, graph inference, and LLM-based clue generation.
| Rank | Participant | Radar | Chicken | Concepts | Restroom | Antique | Pixel | Total |
|---|---|---|---|---|---|---|---|---|
| 1 | AIBuildAI | 98.15 | 93.98 | 85.07 | 100.00 | 97.60 | 93.70 | 568.49 |
| 2 | KR | 98.11 | 84.29 | 65.92 | 100.00 | 98.88 | 94.85 | 542.05 |
| 3 | KCKR | 98.42 | 87.81 | 92.17 | 27.52 | 97.01 | 62.78 | 465.71 |
| 4 | WJ | 98.11 | 87.46 | 80.66 | 27.52 | 97.76 | 71.66 | 463.17 |
| 5 | VM | 97.57 | 98.81 | 79.86 | 42.00 | 96.27 | 46.18 | 460.69 |
| 6 | BM | 99.56 | 91.86 | 69.44 | 0.00 | 99.63 | 82.26 | 442.75 |
| 7 | GT | 93.19 | 16.86 | 88.80 | 27.52 | 96.64 | 100.00 | 423.01 |
| 8 | TS | 98.68 | 99.91 | 10.87 | 27.52 | 99.25 | 77.11 | 413.34 |
| 9 | MT | 93.00 | 70.95 | 23.93 | 27.52 | 100.00 | 88.27 | 403.67 |
| 10 | KA | 97.95 | 99.21 | 63.75 | 0.00 | 97.01 | 42.17 | 400.09 |
| 11 | NT | 87.00 | 72.84 | 26.44 | 13.04 | 100.00 | 99.72 | 399.04 |
The leaderboard contains 284 competitors, with gold medals awarded to the top 24. AIBuildAI's total is the sum of its six independently graded task scores.
How AIBuildAI solved the tasks
AIBuildAI was given only the problem statement and a folder of data. It then worked autonomously: choosing an approach, writing code, training or searching as needed, evaluating models, diagnosing failures, and producing a submission. The same system ran all six problems, with no task-specific tuning and no prior exposure to the tasks.
- Radar. A three-branch network, with one encoder for each radar view, fused into a U-Net that labels every cell. The loss up-weights sparse foreground classes.
- Chicken Counting. A CSRNet density-estimation network on an ImageNet-pretrained backbone. The total of the predicted density map gives the count.
- Antique. A mixture model over five features combined with graph label propagation from only four labelled points.
- Restroom. Four frozen vision backbones (CLIP ViT-B/32, CLIP ViT-L/14, SigLIP, and DINOv2), followed by a Hungarian one-to-one assignment over cosine similarities.
- Pixel. A brute-force search over image crops, retaining the crop whose frozen-CLIP prediction still agrees with the full image.
- Concepts. A budgeted offline search that designs marker clues accepted by the judge. The submitted program reads the precomputed clues.
The variation is important. Radar and Chicken Counting trained neural networks from data. Restroom, Pixel, and Concepts converted frozen models into search or matching objectives. Antique used graph label propagation from only four labelled points, with no neural network. The result is task-adaptive model building under a fixed budget, rather than one fixed recipe applied everywhere.
In Pixel, the agent had to black out 94% of a photograph and choose the surviving 6% so a fixed image model still named the animal. Its selected crops kept the model correct on 654 of 698 images.
Figure 2. The complete pipelines AIBuildAI designed and built for the six tasks. Green blocks use frozen or pretrained components, purple blocks are trained from data, and blue blocks are key stages or outputs.
A closer look at three tasks
Radar. Six sparse, noisy radar heatmaps must be converted into a dense semantic map, with each cell labelled as background or one of four objects. The rare foreground classes are weighted fifty times more than background. The champion and median gold medalist both scored about 98 out of 100. AIBuildAI's from-scratch U-Net ensemble reached 0.9815 weighted pixel accuracy, placing it alongside the strongest contestants.
Chicken Counting. The task counts free-range chickens by regressing a density map whose sum is the count. AIBuildAI scored 0.9398, compared with 84.3 for the champion and 75.5 for the median gold medalist.
Restroom. A cropped restroom pictogram must be matched to the corresponding opposite-gender sign from the same building among sixty candidates. The views differ in angle and are surrounded by near-identical signs. AIBuildAI solved all thirty held-out queries for a perfect score.
Restroom separated the human field sharply. The champion also solved it perfectly, but among the next four finishers, two scored 27.5, one scored 42, and one scored zero. The median gold medalist scored 13.8 out of 100.
Why it matters
Gold-medal-standard performance at the IMO showed AI matching strong human reasoning in the abstract. IOAI carries that capability into practical model building, where solving a problem means constructing a working AI system rather than writing a final answer.
Doing this across all six tasks, end to end and without a weak result, marks a further step for autonomous agents: from competition-grade reasoners toward competent builders of applied AI solutions.
What we released
We have open-sourced the complete AIBuildAI solutions for all six tasks, including the full source code, model configurations, and task outputs.
Code and artifacts: github.com/aibuildai/AI-Build-AI/tree/main/tasks/ioai2025
The six contest tasks and their scoring code are available at: github.com/IOAI-official/IOAI-2025
AIBuildAI ran each task on a single NVIDIA A100 GPU. For Concepts, Gemini 2.5 Flash was used as the judge LLM because the officially recommended model had been retired.