Advancing Abstract Reasoning for RPMs with a Path Aggregation Network and Deep Predictive Reasoning

Diagrammatic reasoning requires deciphering visual patterns, uncovering abstract relationships, and applying logical rules to reach a solution. Although humans excel at spotting analogies and regularities, machines still find it difficult to capture abstract relations, handle multi‑scale dependencies, and generalize across visual tasks. Raven’s Progressive Matrices (RPMs) are a standard benchmark for such reasoning. We introduce PAtNet, a new neural architecture that combines a Path Aggregation Network (PAN) with an attention module to strengthen hierarchical feature extraction and relational reasoning. By jointly modeling global context and fine‑grained details, our method generalizes well across several RPM benchmarks. Tested on the RAVEN, I‑RAVEN, and RAVEN‑FAIR datasets, PAtNet surpasses previous state‑of‑the‑art results, boosting accuracy by 1.3%, 0.1% and 1.1%, respectively. Ablation studies reveal that multi‑scale feature representations from PAN and attention‑based relational reasoning are complementary: the former emphasizes high‑level semantics, while the latter sharpens fine‑detail comparisons. Their synergy yields a more robust and interpretable grasp of diagrammatic reasoning, pointing toward deep learning approaches grounded in cognitive insights.