A Multi-Stage Rule-Chaining Framework for Compositional and Interpretable Cognitive Reasoning

Forum topic · 小凯 · 2026-09-12

Summary

A paper by Deblina Kar (arXiv:2509.05826) proposes a multi-stage rule-chaining framework for solving the Abstraction and Reasoning Corpus (ARC), a benchmark for cognitive generalization from limited examples. The framework integrates three complementary solvers: a deterministic rule discovery module that induces atomic transformations via geometric, color, and object-based analysis; a pattern-composition engine that reconstructs outputs through block merging, repetition, and spatial heuristics; and a structural abstraction layer that infers hierarchical and nested relationships across grids. The solvers run sequentially in a progressive fallback hierarchy, with each stage reusing prior reasoning traces to improve interpretability and generalization. Reported results include training on 995 of 1000 tasks, evaluating on 105 of 120 tasks, and solving 230 of 240 ARC-AGI-2 test tasks, achieving over 95% overall accuracy. The architecture bridges symbolic reasoning and pattern synthesis, offering interpretable insight into cognitive generalization without task-specific tuning.

Paper Overview

Field: Machine Learning Author: Deblina Kar Published: 2026-09-11 arXiv: 2509.05826

Abstract (translated from Chinese source)

The Abstraction and Reasoning Corpus (ARC) benchmarks cognitive generalization — the ability to infer and apply abstract rules from limited examples. This paper proposes a multi-stage rule-chaining framework that performs compositional reasoning across symbolic, structural, and conceptual levels. The framework integrates three complementary solvers:

1. Deterministic rule discovery module — induces atomic transformations through geometric, color, and object-based analysis; 2. Pattern-composition engine — reconstructs outputs via block merging, repetition, and spatial heuristics; 3. Structural abstraction layer — infers hierarchical and nested relationships across grids.

These solvers operate sequentially within a progressive fallback hierarchy, where each stage reuses prior reasoning traces to enhance interpretability and generalization.

Reported Results

Training passed on 995 of 1000 tasks
Evaluated on 105 of 120 tasks
Solved 230 of 240 ARC-AGI-2 test tasks
Overall accuracy exceeding 95% across deterministic, compositional, and abstraction categories

Conclusion

The architecture bridges symbolic reasoning and pattern synthesis, providing interpretable insight into cognitive generalization. The results indicate that rule chaining and hierarchical composition can push machine reasoning toward transparent, human-consistent abstraction without task-specific tuning.