CompressARC, a method demonstrated by Isaac Liao and Albert Gu, shows that lossless information compression at inference time can result in intelligent behavior without pretraining, datasets, or exhaustive search. Achieving a score of 34.75% on training and 20% on evaluation sets of the ARC-AGI challenge, CompressARC utilizes neural networks trained during inference to solve puzzles by compressing them into compact representations. This approach challenges conventional AI methods that rely heavily on extensive pretraining and large datasets.
Sort: