These samples are for an initial preview only and may be further adjusted. More samples are ongoing.
Task list
Design one scale-general optimizer and run it unchanged across the Marin ladder from 550M to 2.545B parameters, scored against the locked AdamH control at every rung.
Improve text-to-image generation from a shared JiT checkpoint, with at most one continuation pass over the same 10M-image GPIC subset.
Design verifiable RL tasks to improve a post-trained Qwen3.5-122B-A10B model across five capabilities, with the model and intended training recipe held fixed.
Inspect and repair the reusable advice in an ACE playbook so a frozen Qwen model answers Formula reasoning problems more accurately.
Build and refine a clean-room CUDA matrix-multiplication kernel for higher sustained H100 throughput while preserving the fixed interface and correctness rules.
Design a better training reward so a robot learns to align and insert a peg more reliably, with the simulator and training budget held fixed.
Kev is a Jev-inspired small neural model that makes decisions directly, without generating text. Improve its decision quality and confidence estimates on fixed training data.
Adapt the supplied long and short mathematical reasoning examples so the same small language model learns more effectively under a fixed training recipe.
Optimize a fused tied-weight linear cross-entropy implementation so it accelerates both the operator itself and a fixed language-model training step.
Speed up a fixed sparse-attention prefill operator without changing which tokens it attends to or weakening its numerical result.
Improve how a frozen Molmo2 model points to anomalies in video by tuning which frames it sees, how it is prompted, and how long it can answer.
Choose which interaction history, page details, and screenshots a frozen MolmoWeb model sees to improve its next browser-action prediction.
Choose how a fixed pool of easy and hard retrieval examples is weighted and ordered to improve ReasonIR-8B, without weakening its general retrieval ability.
Redistribute a fixed decoder's capacity across its layers to reduce late-training loss without changing the model's overall parameter or compute budget.