Auditory attention decoding (AAD) identifies the speaker a listener focuses on from physiological signals, enabling neuro‑steered hearing aids and natural human‑machine interaction. EEG is the dominant modality for AAD, yet it provides incomplete evidence in realistic audio‑visual scenes, motivating fusion with electrooculography (EOG). Existing approaches suffer from weak cross‑modal interaction, inefficient temporal modeling, and low robustness to sample variations.\ \ To overcome these limitations, we propose RAMamba-Net, a reliability‑aware Mamba‑based multimodal fusion network. The architecture comprises:\
- A Mamba‑enhanced band‑aware convolutional Transformer that captures band‑specific EEG patterns and long‑range temporal dynamics;\
- A dual‑branch temporal‑spatial encoder that models EOG temporal sequences and inter‑channel spatial dependencies;\
- Cross‑modal attention for explicit modality interaction;\
- A reliability‑aware module that estimates sample‑wise weights for EEG and EOG, enforcing feature and prediction consistency to strengthen fusion.\ \ Experiments on two public AAD benchmarks demonstrate that RAMamba-Net effectively exploits complementary EEG‑EOG information, achieving a 5.76% accuracy gain over unimodal baselines and delivering more robust decoding and discriminative representations. Further analysis shows that cross‑modal attention improves multimodal alignment, while the reliability‑aware module suppresses unreliable modality evidence and remains robust to signal perturbations and parameter variations.\ \ Review