Official Review · ICLR 2025

This paper proposes a novel framework for analyzing out-of-distribution (OOD) generalization, integrating a generalization ratio to quantify generalization capacity and a generalization decision process (GDP) to link losses across seen and unseen domains. This framework enables backpropagation-based training without model selection, using a combination of informativeness and variation in generalization to derive a generalization inequality, with experiments showcasing its effectiveness in improving generalization ability. 1. They designed an algorithm focused on enhancing the generalization ability of gradient descent, leveraging the concept of a generalization ratio and insights from reinforcement learning. 2. They experimentally demonstrate that the algorithm outperforms traditional gradient descent. 1. The title is intriguing. Would it be clearer to use 'Gradient Descent with the Generalization Ratio: Enhancing.......' instead? 2. In my view, GGD is proposed based on reinforcement learning and the generalization ratio. However, the motivation behind the algorithm design could be clarified further. 3. There is considerable overlap with the paper by Ye et al. (2021). It might help to elaborate more on the ideas behind the algorithm's design. 4. The writing could be improved, as there are several grammatical errors throughout. 1. You introduced the concept of the generalization ratio (GR) in this paper. Could you clarify the functional difference between V_KL for different e_i sets and GR? 2. As a novel and important criterion, could you elaborate on the properties of GR and why it serves as a strong evaluation metric? 3. Could you design some experiments that demonstrate the advantages of using GR instead of V_KL as Ye et al. (2021) also provided a model selection criterion?