The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored. Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures. To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mistral-7B). Results show MoE models achieve 31% higher per-layer efficiency via a “mid-activation, late-amplification” pattern: early layers screen experts, while late layers refine knowledge collaboratively. Ablation studies reveal a “basic-refinement” framework—shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes). Semantic-driven routing is evidenced by strong correlations between attention heads and experts (r=0.68), enabling task-aware coordination. Notably, architectural depth dictates robustness: deep Qwen-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow Olmoe suffers severe degradation (76% drop). Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation. These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.

As Multimodal Large Language Models (MLLMs) develop, their potential security issues have become increasingly prominent. **Machine Unlearning (MU)**, as an effective strategy for forgetting specific knowledge in training data, has been widely used in privacy protection. However, *MU for safety in MLLM has yet to be fully explored*. To address this issue, we propose , a safety unlearning benchmark for MLLMs, consisting of 3,000 images and 28.8K VQA pairs. We comprehensively evaluate unlearning methods from two perspectives: **_forget quality_** and **_model utility_**. Our findings show that existing MU methods struggle to maintain model performance while implementing the forget operation and often suffer from **_over-forgetting_**. Hence, we introduce **Prompt Decouple (PD) Loss** to alleviate over-forgetting through decouple prompt during unlearning process. To quantitatively measure over-forgetting mitigated by PD Loss, we propose a new metric called **Safe Answer Refusal Rate (SARR)**. Experimental results demonstrate that combining PD Loss with existing unlearning methods can effectively prevent over-forgetting and achieve a decrease of 79.5% in the SARR metric of LLaVA-7B and LLaVA-13B, while maintaining forget quality and model utility. Our code and dataset will be released upon acceptance. **Warning: This paper contains examples of harmful language and images, and reader discretion is recommended.**

The widespread deployment of large language models (LLMs) across critical domains has amplified the societal risks posed by algorithmically generated misinformation. Unlike traditional false content, LLM-generated misinformation can be self-reinforcing, highly plausible, and capable of rapid propagation across multiple languages, which traditional detection methods fail to mitigate effectively. This paper introduces a proactive defense paradigm, shifting from passive post hoc detection to anticipatory mitigation strategies. We propose a Three Pillars framework: (1) Knowledge Credibility, fortifying the integrity of training and deployed data; (2) Inference Reliability, embedding self-corrective mechanisms during reasoning; and (3) Input Robustness, enhancing the resilience of model interfaces against adversarial attacks. Through a comprehensive survey of existing techniques and a comparative meta-analysis, we demonstrate that proactive defense strategies offer up to 63% improvement over conventional methods in misinformation prevention, despite non-trivial computational overhead and generalization challenges. We argue that future research should focus on co-designing robust knowledge foundations, reasoning certification, and attack-resistant interfaces to ensure LLMs can effectively counter misinformation across varied domains.

Boundary information is critical for various Chinese language processing tasks, such as word segmentation, part-of-speech tagging, and named entity recognition. Previous studies usually resorted to the use of a high-quality external lexicon, where lexicon items can offer explicit boundary information. However, to ensure the quality of the lexicon, great human effort is always necessary, which has been generally ignored. In this work, we suggest unsupervised statistical boundary information instead, and propose an architecture to encode the information directly into pre-trained language models, resulting in Boundary-Aware BERT (BABERT). We apply BABERT for feature induction of Chinese sequence labeling tasks. Experimental results on ten benchmarks of Chinese sequence labeling demonstrate that BABERT can provide consistent improvements on all datasets. In addition, our method can complement previous supervised lexicon exploration, where further improvements can be achieved when integrated with external lexicon information.

Domain adaption for word segmentation and POS tagging is a challenging problem for Chinese lexical processing. Self-training is one promising solution for it, which struggles to construct a set of high-quality pseudo training instances for the target domain. Previous work usually assumes a universal source-to-target adaption to collect such pseudo corpus, ignoring the different gaps from the target sentences to the source domain. In this work, we start from joint word segmentation and POS tagging, presenting a fine-grained domain adaption method to model the gaps accurately. We measure the gaps by one simple and intuitive metric, and adopt it to develop a pseudo target domain corpus based on fine-grained subdomains incrementally. A novel domain-mixed representation learning model is proposed accordingly to encode the multiple subdomains effectively. The whole process is performed progressively for both corpus construction and model training. Experimental results on a benchmark dataset show that our method can gain significant improvements over a vary of baselines. Extensive analyses are performed to show the advantages of our final domain adaption model as well.