Hidden Forgetting in Continual Multimodal Learning: When Accuracy Survives but Grounding Fails

Multimodal continual learning (MMCL) aims to learn emerging knowledge from multimodal data while preserving knowledge. To mitigate forgetting, current MMCL methods usually focus on cross-modal representation alignment or semantic similarity, but they overlook whether the relative contributions of individual modalities and their interactions remain stable across incremental tasks. We term this decision-level shift Modality Contribution Drift (MCD) and quantify it with the MCD score, which combines contribution-strength and relative-reliance changes under controlled interventions on modality subsets. Theoretical and empirical analyses further explain why current MMCL methods cannot reliably mitigate this drift. To this end, we propose Continual Modality Contribution Drift Regularization (CMCDR), which preserves the modality contribution structure of previously learned tasks. Since MMCL settings differ in whether old exemplars are available, CMCDR includes both replay-based and replay-free versions. The replay-based version uses modality-subset interventions as diagnostic probes on stored old samples, compares their contribution profiles between the current model and a frozen previous model, and constrains changes in old-sample modality-specific and interaction contributions. The replay-free version uses current-task samples as probes and distills the frozen model's old-task contribution responses, thereby regularizing the observed contribution profile without exemplars. Experiments on multimodal class-incremental learning and continual visual question answering validate the generality and effectiveness of CMCDR.

Continual adaptation is essential for multimodal large language models (MLLMs) deployed across evolving domains, but the state-of-the-art MR-LoRA method highly relies on the assumption that a MLLM-based router is necessary to process complex multimodal inputs. This paper revisits this claim on the MLLM-CL benchmark and argues for two claims. \textbf{First}, routing does not require an MLLM: a simple training-free, replay-free ptotypical routing method (\textsc{RePRo}), uses frozen pretrained features and task prototypes to match the MLLM-based router of MR-LoRA at far lower computational cost. \textbf{Second}, shared experts do not improve continual learning for MLLMs, despite their theoretical appeal. We show that these findings arise from two structural limitations of MLLM-CL: (1) its tasks are \textbf{highly separable} in representation space, and (2) its fixed task order makes conclusions \textbf{sensitive to a single curriculum} rather than robust across diverse continual-learning trajectories. As a result, the benchmark primarily rewards learning in isolation rather than genuine continual transfer. This motivates a new design for future benchmarks of continual MLLM learning, with overlapping task manifolds, multiple task orders, fine-grained domain shifts, and evaluation protocols that reward forward transfer as well as retention.

Multimodal Large Language Models (MLLMs) are often constrained by a language-space bottleneck, forcing complex visual reasoning into discrete tokens which can lose perceptual nuance. A promising alternative is continuous latent reasoning, where the goal is to discover implicit reasoning pathways that bridge the multimodal query and the final answer. However, this introduces a severe train-inference mismatch: a training-time posterior, conditioned on the ground-truth answer, can exploit answer-dependent shortcuts. Standard variational training then forces the inference-time prior to mimic a posterior that has access to information unavailable at test time, leading to poor performance. To address this, we propose Asymmetric Mutual Variational Learning (AMVL), a framework that resolves this mismatch via a bidirectional calibration objective. A forward KL divergence trains the target-agnostic prior to match the posterior, while a novel reverse KL divergence simultaneously regularizes the posterior, preventing it from collapsing into inference-incompatible regions and mitigating this ``answer leakage''. We provide theoretical analysis formalizing this leakage as prior contamination and prove that our dual-KL objective reduces it. We instantiate AMVL in a latent-integrated MLLM and show that it consistently outperforms strong discrete and latent-reasoning baselines, improving the average score on the complex BLINK benchmark by +10.83 and achieving gains of up to +32.00 on individual reasoning tasks, with analyses confirming improved latent-space stability.