Épreuves d'olympiades

One-hour record-linkage task: match GPS records to traffic-log entries one-to-one within blocks without a shared key.

Three-hour paper of 22 multiple-choice and written problems on search, supervised and unsupervised learning, recommenders, RL, evaluation and kernel methods.

Six subtasks from a primality check to embedding clustering, semantic search and two-hop evidence retrieval over a fictional corpus, without LLMs.

On 4-channel machine sensor recordings, detect faults, classify four fault types and localise the fault interval, scored by macro F1 and mean IoU.

Answer 10,000 geometry queries (circumcircle area, convex hull area, shape type from distances, triangle angles from cevians) with no training data.

Predict the finishing order of spaceships racing across alien biomes whose names in the test set never appear in training.

Repair a Vision Transformer for 15 animals that breaks on new test images, with only 15 clean training photos and no other pretrained model.

Tag the names in 4,000 news sentences with only 50 fully labelled examples, plus 8,000 sentences where you only know how many names of each type they contain.

Find the 10 impostor pages hidden in a 3,000-page cartoon diary without any labelled training data.

Fifteen multiple-choice questions on ML basics such as k-NN, bias in hiring models and evaluation.

Search for alien-token inputs that make a given translator model emit prescribed English sentences.

Choose which of four sentence sequences describes the same action as a target, using fixed word embeddings.

Predict one crop per image that keeps listed wanted objects and excludes unwanted ones, from 15 examples.

Craft small pixel perturbations that make a ResNet18 and a ViT-Tiny disagree in prescribed ways on images both classify correctly.

Locate the character position where a machine-altered continuation begins inside passages from books.

Train from scratch a network under 20,260 parameters that reproduces a parameterised 2D field drawing the IOAI logo.

Recover the original turn order of two-speaker dialogues whose per-turn audio clips have been shuffled.

Find a hidden word within 30 turns using only feedback on which of two guesses is semantically closer to it.

Predict the next action of robots in 6x6 gridworlds whose behaviour is described by natural-language instructions.

Teams train policies for a simulated Galbot G1 to pick products, navigate and place them on shelves in IsaacLab.

Clustering task (details behind a Kaggle invite link).

Detect defects in shape images.

Match fragments of poems that belong together.

Classify fault type in multichannel vibration recordings and localise the fault in time with only recording-level labels.

Recover the original order of row-shuffled positional-embedding matrices of a CLIP model.

Segment unspaced Chinese sentences into words without pretrained models or labelled data.

Computer-vision steering task (details behind a Kaggle invite link).

Classify how much audio clips were sped up.

Decode intercepted encrypted text messages.

Remove watermarks from images.

Separate two sounds mixed into a single waveform.

CLIP-based image-text competition (no public description).

Classify Romanian historical texts by region and tag token-level entities in old letters.

Detect defective hazelnuts in images.

Computer-vision detection/matching task described by organisers as a simple detection.

Image segmentation practice competition.

Predict four levels of rain intensity from weather data.

Given a token sequence formed by randomly interleaving two paragraphs, label which paragraph each token came from.

For triples of AI-generated images in different art styles, pick the one lacking the object type shared by the other two.

Hide 256-bit messages in the weights of a small MNIST CNN and recover hidden bits, under constraints from three different reviewers.

Train digit classifiers that stay accurate on images from unseen, faulty scanners (domain generalisation).

Score pairs of system-call trace windows by how likely they come from the same program, including a part with disguised command names.

From noisy coordinates of bodies in synthetic planetary systems, infer which body each one orbits and forecast future positions.

Find small token rationales for a frozen DistilBERT sentiment classifier that pass sufficiency and comprehensiveness probes, in three differently scored variants.

Read short uppercase alphanumeric strings from images and output their length, character-type pattern and normalised text.

From image slices, wave signals and symbol sequences, predict warping-spline coefficients, an open-set topological class and a stability index.

Locate the single star in each telescope image and regress its true flux using the image plus atmospheric metadata.

Estimate optimal state values of an unknown stochastic MDP from a fixed log of shuffled transitions.

Extend a deployed audio spectrogram transformer from 16 to 29 sound classes without forgetting the original ones, within a short GPU budget.

Learn by imitation from a small set of demonstrations a policy that picks up and delivers a package on 8x8 grid maps.