The Abstraction and Reasoning Corpus (ARC) is increasingly used to test whether neural networks can infer task‑specific transformations from only a few demonstrations. Recent ARC models differ not only in architecture but also in training and evaluation protocols such as pretraining data, data augmentation, test‑time setup, and optimization budgets, making it hard to attribute performance gains solely to architecture. In this work we adopt a common inductive protocol on ARC‑AGI‑1 to systematically study how the vision‑inspired VARC and the recursion‑based TRM respond to varying pretraining and test‑time training (TTT) budgets. Five pretraining budgets are crossed with four TTT budgets for each test task, where each budget is defined by the number of task variants and the total number of sample presentations, thus creating controlled training budgets. Results show that increasing the number of TTT sample presentations consistently improves the standard ARC pass@2 metric for both models. Raising the TTT task‑variant count benefits TRM across the board, while VARC often performs better with fewer variants. VARC remains competitive even without pretraining, whereas TRM gains much more from pretraining. These findings demonstrate that both pretraining and TTT budgets significantly affect ARC evaluation and motivate reporting results under comparable budgets.
Review