Multimodal large language models (MLLMs) trained on image-text data may retain sensitive information in textual memories, visual representations, and cross-modal associations, creating a need for reliable unlearning. Although existing unlearning methods can reduce target disclosure under clean queries, suppressed knowledge may remain recoverable under adversarial conditions. We introduce a new perspective on this vulnerability by assessing post-unlearning recovery through conditional perturbation generation, where a learned generator maps each image-text input to perturbations that promote recovery of the suppressed response. This framework is instantiated with two generators: CEPG injects conditional soft tokens into the multimodal input embedding sequence, while CPPG generates pixel-space perturbations under CLIP-based semantic conditions. Experiments across five unlearning methods show strong recovery risks. CEPG achieves average recovery ratio (Rec) scores of 60.19% on MLLMU-Bench and 46.74% on CLEAR, while CPPG attains 34.28% and 29.47%, respectively. Both attacks remain effective with limited or no direct access to forgotten samples, highlighting that current unlearning offers insufficient robustness guarantees. To mitigate this vulnerability, we introduce Adversarially Robust Unlearning (ARU), which integrates conditional recovery attacks into unlearning with attribution-guided parameter selection, significantly reducing recovery rates while preserving utility on retain sets.

Large language models (LLMs) face a significant challenge due to the excessive computational and memory requirements of the commonly used Transformer architecture. While state space model (SSM) is a new type of foundational network architecture offering lower computational complexity, their performance has yet to fully rival that of Transformers. This paper introduces DenseSSM, a novel approach to enhance the flow of hidden information between layers in SSMs. By selectively integrating shallow-layer hidden states into deeper layers, DenseSSM retains fine-grained information crucial for the final output. This incremental improvement maintains the training parallelizability and inference efficiency of SSMs while significantly boosting performance. The proposed method is broadly applicable to various SSM types, including RetNet and Mamba, and DenseSSM achieves significant performance improvements on public benchmarks, demonstrating its effectiveness and versatility.