MambaVision + PoM: Polynomial Mixing as a drop-in attention replacement in a hybrid Mamba-Transformer vision backbone

Investigating whether the Polynomial Mixer (PoM), a linear-time token-mixing operator from CVPR 2026, can replace self-attention in MambaVision (CVPR 2025, NVIDIA) while preserving accuracy and improving efficiency at high resolution. The core finding: PoM delivers a theory-consistent throughput advantage in the long-sequence regime, while memory remains identical due to PyTorch's Flash-Attention-compatible SDPA kernel.