ARC-AGI-2 技术报告

The Abstraction and Reasoning Corpus (ARC) is designed to assess generalization beyond pattern matching, requiring models to infer symbolic rules from very few examples. A transformer-based system is proposed that enhances ARC performance by integrating neural inference with structure-aware priors and online task adaptation. The approach reformulates ARC reasoning as a sequence modeling problem using a compact task encoding of only 125 tokens, enabling efficient long-context processing with a modified LongT5 architecture. Additionally, a principled augmentation framework is introduced based on group symmetries, grid traversals, and automata perturbations, enforcing invariance to representation change. These innovations demonstrate improved adaptability and generalization capabilities in tackling complex reasoning tasks.