Will any plain transformer model achieve 60% or more on ARC-AGI-2 by 2030?

The inference cost to achieve this result does not matter.

The model that achieves this result must use the same "transformer recipe" common between 2023-2025: techniques like RLHF/RLAIF/CoT/RAG/vision encoders are allowed, but any specialized components must also be made of vanilla transformer blocks; Any new inductive biases, such as tree-search, neurosymbolic logic, etc. would not qualify.

The result must be verified by at least one reputable, unaffiliated org (ARC, Epoch, OpenAI Evals, academic lab, etc.) or a publicly re-runnable result (notebook on Kaggle, etc.).

Resolution uses the ARC-AGI-2 evaluation set and scoring script as published on arcprize.org on the day this market opens. Later revisions are ignored.

• Update 2025-10-12 (PST) (AI summary of creator comment): PPO, GRPO, and RLVR are allowed training methods.

Generating synthetic data using other models to train the transformer is allowed, as long as the final model follows the common transformer recipe.