<p>Reinforcement learning (RL)-based multi-hop knowledge graph reasoning has achieved remarkable success in real-world applications and can effectively handle knowledge graph completion tasks. This approach involves a policy-based agent navigating the graph environment to extend reasoning paths and identify the target entity. However, most existing multi-hop reasoning models are typically constrained to stepwise inference, which inherently disrupts the global information of multi-hop paths. To overcome this limitation, we introduce discriminative features between valid and invalid paths as global information. Here, we propose a multi-hop path encoder specifically designed to extract these discriminative features. Firstly, a multi-hop path encoding module is employed to derive each path’s hidden features, using cross-attention mechanisms to strengthen the interaction between triple context and path features. Secondly, a discriminative feature extraction module is used to capture the differences between valid and invalid paths. Thirdly, an attention-enhanced gated fusion mechanism is implemented to integrate these discriminative features into the multi-hop inference decoder. We further evaluate our method on five standard datasets. Our method outperforms the baseline models, demonstrating the effectiveness of discriminative features in improving prediction performance, learning speed, and path interpretability.</p>