Official Review · ICLR 2026

The authors investigate the performance dependency of convolutional networks and Transformers in 3D segmentation tasks. They demonstrate that CNN–Transformer hybrid architectures rely heavily on the convolutional component for achieving strong segmentation performance, while the Transformer’s contribution remains minimal. To address this imbalance, the authors propose two variants of Primus that incorporate recent Transformer design advances—such as Rotary Positional Embeddings (RoPE) and EVA-style MLPs—to reduce reliance on CNNs and enhance the Transformer’s contribution in 3D segmentation. - The paper presents a investigation comparing CNN and Transformer architectures for 3D segmentation. - It provides many ablation studies on architectural design choices, offering valuable insights that can guide future research on Transformer-based 3D segmentation. - Unclear performance gap explanation: the authors investigate the issue that CNN–Transformer hybrid architectures heavily rely on the CNN module for 3D segmentation. However, it remains unclear why CNNs consistently outperform Transformers in this domain, or how the proposed design choices specifically address this performance gap. A more detailed analysis or visualization explaining the inductive bias difference would strengthen the argument. - Misleading problem statement: the paper argues that future 3D segmentation models should move toward pure Transformer architectures, citing advantages in self-supervised learning and multi-modal integration. However, this justification is not fully convincing. Hybrid CNN–Transformer architectures (e.g., KL-VAE, VQ-VAE, and their variants) have demonstrated strong self-supervised learning capabilities, and current research trends do not exclusively favor pure Transformers for such tasks. Moreover, hybrid models can effectively perform multimodal fusion through either Transformer or CNN blocks, provided feature dimensionalities are compatible. Thus, the necessity of abandoning CNN-based architectures remains questionable. - High parameter count and computational cost: the proposed model family includes four sizes—Small, Base, Medium, and Large. While the Large variant achieves competitive results with nnUnet, it requires around 300M parameters and 560B FLOPs, significantly higher than other more lightweight architectures. Although the paper introduces smaller variants (Small and Base), their performance gains over prior work appear marginal, suggesting that improvements may primarily stem from scaling up model size rather than architectural innovation. - Outdated baseline comparisons: the baseline comparisons in Table 6 seem outdated. Given the field's rapid pace, works from 2023-24 would also be acceptable. However, most of the comparison baselines are from 2021. - Pleas see weakness section.