The IOAI Field Manual
Every IOAI task from 2024 to the actual 2026 home tasks, decoded for someone starting from zero: what the task really asks, how the given baseline works, how the winning solution differs, and exactly how you'd get there driving Gemma 4 under a 2000-token limit.
Each page has the same six sections: 📜 the task in plain English (story stripped, metric extracted) · 🔧 the baseline you're given · 🚀 baseline vs. solution side by side · 🧒 a from-zero explanation of the concepts · 💬 a simulated Gemma 4 playthrough — the literal chats that would solve it under the token cap · 🎯 takeaways and likely Day-1 extensions.
- Metric first. Find the scoring cell before reading the story. The metric defines the game.
- Run the baseline untouched → submit it. Points on the board; format verified; scores are normalized against the baseline, so any improvement counts.
- Diagnose before improving. Confusion matrix, failed-episode replays, per-class accuracy — find where points are lost.
- Read the hints; rank fixes by points-per-hour. GAITE statements contain the intended solution.
- One change at a time, measure each. Keep a scrappy log: change → score.
- Name the lesson. Every task secretly teaches one failure mode (forgetting, drift, noisy oracle). Name it and you know the solution family.
- Guard the clock. Baseline submitted in 30 min; freeze experiments 45 min before the end.
- Never leave a task at zero.
Operation Night Watch
Teach an audio model 13 new sounds without it forgetting the 16 it already knows.
🤖Robot Delivery Academy
Teach a robot to deliver packages by copying an expert — and learn why copying is harder than it looks.
🗄️The Analytical Language of John Wilkins
20 Questions against an LLM oracle: find 1 animal among ~1,400 in 15 yes/no questions.
Chameleon
Guess the secret word from a sequence of icons — embeddings and semantic similarity.
📡Radar
Detect humans in radar heatmaps — image classification on data that isn't photos.
🌧️Weather
Predict rain from satellite images plus context features — fusing two data types in one model.
Concepts (Chameleon's Day-1 extension)
The at-home guessing game returns — now with an official LLM judge API and a call budget.
🏺Antique
500 rows, 5 features, some labels unknown — classical sklearn beats deep learning here.
🐔Chicken Counting
Count chickens via density maps — with a frozen encoder you're not allowed to touch.
🚻Restroom
Match male↔female restroom icons from the same restroom — embeddings for images.
🔲Pixel
Choose the most informative pixels under a budget — probe what models actually look at.
Word Segmentation
Split German compound words character by character — sequence labeling made friendly.
🎙️Synthetic Speech Detector
Spot AI-generated speech from spectrograms — the statement literally tells you the solution.
Help BOBAI (fix the model)
A model is broken or must grow new abilities — repair and extend instead of rebuilding.
🌌Lost in Hyperspace
Navigate embedding space itself — vectors, distances, and what 'nearby' means.
🐄Madarian Cow (edit the weights)
Make an image generator draw zebras when asked for giraffes — surgery inside the model.