Multi-Stage Rule Chaining for Compositional and Interpretable Reasoning

Summary

This paper presents a multi-stage rule-chaining framework for the Abstraction and Reasoning Corpus, a benchmark of cognitive generalization from limited examples. The system combines three sequential solvers: a deterministic module that discovers atomic geometric, color, and object transformations; a pattern-composition engine that reconstructs outputs through block merging, repetition, and spatial heuristics; and a structural abstraction layer that identifies hierarchical and nested relationships between grids. The stages follow a progressive fallback hierarchy and reuse earlier reasoning traces, with the goal of making the system’s decisions more interpretable and improving generalization. On ARC-AGI-2, training completed for 995 of 1,000 tasks, while the system was evaluated on 105 of 120 tasks and solved 230 of 240 test tasks. The authors report overall accuracy above 95% and strong coverage of deterministic, compositional, and abstract task categories. They present the results as evidence that combining rule chaining with hierarchical composition can connect symbolic reasoning and pattern synthesis without task-specific tuning, although the abstract does not provide further breakdowns of errors or comparisons with other systems.