ARC-AGI a secret sauce for reasoning LLMs?

Giotto.ai achieved remarkable results on the 2025 Abstract and Reasoning Corpus (ARC-AGI) challenge, a benchmark designed to measure true reasoning capabilities in AI systems. In this presentation, we will walk through the core innovations that enabled this success - including test-time training, symmetry-based scoring, and automata-driven data augmentation. We’ll explain how these components interact to improve generalization and adaptability in reasoning LLMs, and what this means for the future of machine intelligence.