The decoder plays an important role in semantic segmentation networks, which determines the effectiveness of predicting the target from the abstract features. However, existing works mostly use an upsampling subnetwork to gradually reconstruct high-resolution outputs. This tends to affect the segmentation accuracy of inter-class boundaries and blurs the edges of the parsed results. To address this problem, we propose an Implicit Representation Querying Network (IRQNet) to query the decoding results with known information. The prediction process corresponds to a continuous coordinate-based function mapping information to semantic labels, which is used to directly query the high-resolution results instead of recovering features. To further enhance contextual understanding and attention guidance, we introduce an Information Enhancement Module, comprising Deformable Fusion and Position Encoding. Our approach achieves significant gains on two popular scene segmentation datasets, demonstrating its effectiveness. The code is available at https://github.com/nchennnn/IRQNet .