Abstract reasoning poses significant challenges to artificial intelligence algorithms, demanding a cognitive ability beyond that required for perceptual tasks. In this study, we introduce the Cross-Feature Network (CFN), a novel framework designed to separately extract concepts and features from images. This framework utilizes the responses of features to concepts as representations for reasoning, particularly in addressing the Bongard-Logo problem. By integrating an Expectation-Maximization process between the extracted concepts and features within the CFN, we have achieved notable results, albeit with certain limitations. To overcome these limitations, we propose the Triple-CFN, an efficient model that maximizes feature extraction from images and demonstrates effectiveness in both the Bongard-Logo and Raven's Progressive Matrices (RPM) problems. Furthermore, we introduce Meta Triple-CFN, an advanced version of Triple-CFN, which explicitly constructs a concept space tailored for RPM problems. This ensures high accuracy of reasoning and interpretability of the concepts involved. Overall, this work explores innovative network designs for abstract reasoning, thereby advancing the frontiers of machine intelligence.