ARC-AGI without pretraining

By Isaac Liao and Albert Gu In this blog post, we aim to answer a simple yet fundamental question: Can lossless information compression by itself produce intelligent behavior? The idea that efficient compression by itself lies at the heart of intelligence is not new (see, e.g., Hernández-Orallo & Minaya-Collado, 1998; Mahoney, 1999; Hutter, 2005; Legg & Hutter, 2007). Rather than revisiting those theoretical discussions, we make a practical demonstration instead. In this work, we give evidence that lossless compression during inference time is sufficient to produce intelligent behavior, by developing a method purely based on compression that performs well on the ARC-AGI challenge, a dataset of IQ-test-like puzzles about inferring a procedure/rule from limited demonstrations. Crucially, our solution, which we name CompressARC, obeys the following three restrictions: No pretraining ; models are randomly initialized and trained during inference time. ; models are randomly initia ... Read full article.