Large Language Model (LLM) unlearning is essential for removing sensitive or copyrighted knowledge while preserving general utility. Existing methods often leave residual knowledge in intermediate representations, which can still be recovered. To address this, we propose Cascade, a hierarchical recoverability control framework that minimizes the internal identifiability of target knowledge. Cascade combines three complementary controls: path-level routing to suppress privacy-associated activation routes, representation-level compression to reduce geometric separability, and decoding-level intervention to limit residual recovery. Experiments on TOFU, MUSE-News, and WMDP, including robustness tests with query reformulation and extraction-style prompts, show that Cascade effectively reduces recoverability while maintaining stable model utility.

Client-side privacy rewriting is crucial for deploying LLMs in privacy-sensitive domains. However, existing approaches struggle to balance privacy and utility. Full-text methods often distort context, while span-level approaches rely on impractical manual masks or brittle static dictionaries. Attempts to automate localization via prompt-based LLMs prove unreliable, as they suffer from unstable instruction following that leads to privacy leakage and excessive context scrubbing. To address these limitations, we propose DAMPER (Domain-Aware Mask-free Privacy Extraction and Rewriting). DAMPER operationalizes latent privacy semantics into compact Domain Privacy Prototypes via contrastive learning, enabling precise, autonomous span localization. Furthermore, we introduce a Prototype-Guided Preference Alignment, which leverages learned prototypes as semantic anchors to construct preference pairs, optimizing a domain-compliant rewriting policy without human annotations. At inference time, DAMPER integrates a sampling-based Exponential Mechanism to provide rigorous span-level Differential Privacy (DP) guarantees. Extensive experiments demonstrate that DAMPER significantly outperforms existing baselines, achieving a superior privacy-utility trade-off.

Knowledge graph embedding (KGE) is an important task and it can benefit lots of downstream applications. General KGE can increase the embedding dimension to improve model performance. High-dimensional KGE will significantly increase the number of model parameters and training time. Therefore, knowledge distillation is proposed for learning a low-dimensional model from a pre-trained high-dimensional model. To avoid introducing a complex teacher model, we use self-knowledge distillation. However, there are still some issues with the self-knowledge distillation model we mentioned later. One of them is misdirection from incorrect predictions during model training. Another is the loss of discrimination information caused by excessive distillation temperature. To address these issues, we apply self-knowledge distillation, knowledge adjustment and dynamic temperature distillation to KGE. Self-knowledge distillation uses the information from the latest iteration to guide the training in the current iteration. Knowledge adjustment fixes the predictions of misjudged training samples. Dynamic temperature distillation designs dynamic sample-wise temperatures to compute soft targets. Our model can not only improve model performance but also achieve a lightweight model. Experimental results demonstrate the effectiveness and generalization ability of our model in link prediction. The lightweight model can maintain good model performance while reducing the number of model parameters and training time.