mdlARC Scores 44% on ARC-AGI for 67 Cents

mdlARC is a compelling sample-efficiency result, but its benchmark specialization should not be confused with general reasoning.

• –Uses 3D RoPE, per-task embeddings, data augmentation, and test-time training to improve ARC performance.
• –The standard transformer architecture challenges the assumption that ARC requires elaborate recursive designs.
• –Ablations suggest representation choices matter more than raw scale, with several removals dropping performance sharply.
• –Its low compute cost makes rapid experimentation accessible to individual researchers.
• –The result remains limited by public-benchmark optimization and does not establish broad generalization.