The research presented a multi-stage rule-chaining framework designed for cognitive reasoning. The framework incorporates three solvers: a deterministic rule discovery module, a pattern-composition engine, and a structural abstraction layer. These solvers operate sequentially, utilizing prior reasoning traces for enhanced interpretability and generalization. The system was trained on 995 of 1000 ARC tasks and subsequently evaluated on 105 of 120 tasks, successfully solving 230 out of 240 ARC-AGI-2 tasks. Overall accuracy exceeded 95 percent across deterministic, compositional, and abstract categories.
Source: https://arxiv.org/abs/2609.10654