ARC-AGI Without Pretraining (2025) (iliao2345.github.io)

A groundbreaking approach titled "CompressARC" demonstrates that lossless information compression alone can yield intelligent behavior, challenging traditional AI methodologies that rely heavily on extensive pretraining and large datasets. This technique is showcased by its performance on the ARC-AGI challenge—a benchmark for assessing the ability to infer and generalize abstract rules from minimal examples—where CompressARC achieved a score of 34.75% on the training set and 20% on the evaluation set, despite being trained solely during inference on the specific puzzle at hand. CompressARC operates under strict constraints: no pretraining, no extensive datasets, and minimal search beyond gradient descent. The neural network employed transforms incompletely defined puzzles into complete solutions through a compression framework during inference. This innovative approach not only streamlines the process of problem-solving but also suggests a future where tailored compressive objectives can efficiently harness minimal input to derive deeper intelligence, signifying potential shifts in AI development paradigms toward more resource-efficient models. This work could redefine foundational concepts in artificial general intelligence by highlighting the effectiveness of information compression as a primary mechanism for intelligence.