K-Inverse-RFM: Bridging Recursive Feature Machines and Neural Networks on Data-Corrupted Mathematical Tasks

This paper studies why Recursive Feature Machines (RFMs), kernel machines that use the Average Gradient Outer Product (AGOP) for feature learning, can underperform feedforward neural networks on corrupted mathematical tasks. It introduces K-Inverse-RFM, a transformation applied to training labels. According to the abstract, the transformation is designed to improve learning under noisy data, complex representations, and class imbalance. The authors report that it closes the performance gap between RFMs and FNNs in these settings and can outperform neural networks in some cases. The available summary does not provide benchmark names, dataset sizes, numerical gains, or ablation details.