ARC-AGI-2 Technical Report

A collaborative research team has introduced ARC-AGI-2, a novel transformer-based system designed to tackle the challenging Abstraction and Reasoning Corpus (ARC) benchmark. The ARC is specifically crafted to assess an AI's ability to generalize beyond simple pattern matching, requiring models to infer abstract, symbolic rules from just a few demonstration examples—a core challenge for achieving human-like reasoning. The team's approach is built on four key innovations that work synergistically to push performance beyond prior neural solvers.

First, the system reformulates ARC reasoning as a sequence modeling problem using an extremely compact task encoding of only 125 tokens. This enables efficient long-context processing using a modified LongT5 architecture. Second, it introduces a principled data augmentation framework based on group symmetries, grid traversals, and automata perturbations, which enforces the model's invariance to changes in how a problem is represented.

Third, and crucially, the system employs test-time training (TTT) with lightweight LoRA (Low-Rank Adaptation) fine-tuning. This allows the model to specialize for each unique, unseen task during inference by learning its specific transformation logic directly from the provided demonstrations. Finally, a symmetry-aware decoding and scoring pipeline aggregates solution likelihoods across multiple augmented views of a task, performing what the authors call 'multi-perspective reasoning' to improve consistency.

The combined result is a system where augmentations expand the hypothesis space, TTT sharpens local reasoning on the fly, and symmetry-based scoring validates solutions. This architecture represents a significant step in moving AI from statistical pattern recognition toward more robust, abstract, and generalizable reasoning capabilities, as evidenced by its improved performance on the demanding ARC benchmark.