Class-Incremental Learning (CIL) aims to continuously acquire new classes while preserving knowledge of previously learned ones. Recent CIL research has largely focused on single modalities such as images or text, leaving the audio‑visual intersection largely unexplored. Preliminary experiments reveal that frozen multimodal foundational model (MFM) embeddings perform poorly in incremental settings because they lack a stable cross‑task contextual signal.\ \ To address this gap, we propose a frozen‑MFM based framework for audio‑visual CIL. The framework consists of two key components:\
- Dense audio‑visual feature extraction: The MFM’s audio encoder and visual encoder produce high‑dimensional audio features $\mathbf{a}$ and visual features $\mathbf{v}$ respectively.\
- Attention‑driven contextual guidance module: Audio features are treated as a task‑stable context vector $\mathbf{c}$, which guides the visual features via a multi‑head attention operation, yielding context‑aware visual representations $\mathbf{\hat{v}} = \text{Attention}(\mathbf{v}, \mathbf{c})$. This context remains unchanged across incremental tasks, providing a consistent reference for the visual classifier.\ \ To further mitigate catastrophic forgetting, we employ both feature‑level and logit‑level distillation:\ $$\mathcal{L}_{\text{distill}} = \|\mathbf{v}^{t} - \mathbf{v}^{t-1}\|_2^2 + \text{KL}(\mathbf{z}^{t}\|\mathbf{z}^{t-1})$$\ where $\mathbf{v}^{t}$ and $\mathbf{z}^{t}$ denote the current task’s features and logits, and $t$ indexes the task. Extensive experiments on several public audio‑visual CIL benchmarks demonstrate consistent superiority over state‑of‑the‑art methods. The code is released at https://github.com/avigupta2798/MFM-CIL.\ \ Review: By leveraging audio as a stable cross‑task context and integrating it through attention, the proposed approach achieves robust visual incremental learning. Combined with feature and logit distillation, it effectively curbs forgetting, offering a versatile solution for multimodal incremental scenarios.