Based on the foundation of Large Language Models (LLMs), Multilingual Large Language Models (MLLMs) have been developed to address the challenges of multilingual natural language processing tasks, hoping to achieve knowledge transfer from high-resource to low-resource languages. However, significant limitations and challenges still exist, such as language imbalance, multilingual alignment, and inherent bias. In this paper, we aim to provide a comprehensive analysis of MLLMs, delving deeply into discussions surrounding these critical issues. First of all, we start by presenting an overview of MLLMs, covering their evolution, key techniques, and multilingual capacities. Secondly, we explore widely utilized multilingual corpora for MLLMs' training and multilingual datasets oriented for downstream tasks that are crucial for enhancing the cross-lingual capability of MLLMs. Thirdly, we survey the existing studies on multilingual representations and investigate whether the current MLLMs can learn a universal language representation. Fourthly, we discuss bias on MLLMs including its category and evaluation metrics, and summarize the existing debiasing techniques. Finally, we discuss existing challenges and point out promising research directions. By demonstrating these aspects, this paper aims to facilitate a deeper understanding of MLLMs and their potentiality in various domains.
The Mixture of Experts (MoE) for language models has been proven effective in augmenting the capacity of models by dynamically routing each input token to a specific subset of experts for processing. Despite the success, most existing methods face a challenge for balance between sparsity and the availability of expert knowledge: enhancing performance through increased use of expert knowledge often results in diminishing sparsity during expert selection. To mitigate this contradiction, we propose HyperMoE, a novel MoE framework built upon Hypernetworks. This framework integrates the computational processes of MoE with the concept of knowledge transferring in multi-task learning. Specific modules generated based on the information of unselected experts serve as supplementary information, which allows the knowledge of experts not selected to be used while maintaining selection sparsity. Our comprehensive empirical evaluations across multiple datasets and backbones establish that HyperMoE significantly outperforms existing MoE methods under identical conditions concerning the number of experts.
The feed-forward networks (FFNs) in transformers are recognized as a group of key-value neural memories to restore abstract high-level knowledge. In this work, we conduct an empirical ablation study on updating keys (the 1st layer in the FFNs layer) or values (the 2nd layer in the FFNs layer). We compare those two methods in various knowledge editing and fine-tuning tasks of large language models to draw insights to understand FFNs further. Code is available at $\href{https://github.com/qiuzh20/Tuning-keys-v.s.-values}{this\,repo}$.
It has been shown that semi-parametric methods, which combine standard neural networks with non-parametric components such as external memory modules and data retrieval, are particularly helpful in data scarcity and out-of-distribution (OOD) scenarios. However, existing semi-parametric methods mostly depend on independent raw data points - this strategy is difficult to scale up due to both high computational costs and the incapacity of current attention mechanisms with a large number of tokens. In this paper, we introduce a novel heterogeneous memory augmentation approach for neural networks which, by introducing learnable memory tokens with attention mechanism, can effectively boost performance without huge computational overhead. Our general-purpose method can be seamlessly combined with various backbones (MLP, CNN, GNN, and Transformer) in a plug-and-play manner. We extensively evaluate our approach on various image and graph-based tasks under both in-distribution (ID) and OOD conditions and show its competitive performance against task-specific state-of-the-art methods. Code is available at \url{https://github.com/qiuzh20/HMA}.
Incorporating modular designs into neural networks demonstrates superior out-of-generalization, learning efficiency, etc. Existing modular neural networks are generally $\textit{explicit}$ because their modular architectures are pre-defined, and individual modules are expected to implement distinct functions. Conversely, recent works reveal that there exist $\textit{implicit}$ modular structures in standard pre-trained transformers, namely $\textit{Emergent Modularity}$. They indicate that such modular structures exhibit during the early pre-training phase and are totally spontaneous. However, most transformers are still treated as monolithic models with their modular natures underutilized. Therefore, given the excellent properties of explicit modular architecture, we explore $\textit{whether and how dense pre-trained transformers can benefit from emergent modular structures.}$ To study this question, we construct \textbf{E}mergent $\textbf{M}$ixture-$\textbf{o}$f-$\textbf{E}$xperts (EMoE). Without introducing additional parameters, EMoE can be seen as the modular counterpart of the original model and can be effortlessly incorporated into downstream tuning. Extensive experiments (we tune 1785 models) on various downstream tasks (vision and language) and models (22M to1.5B) demonstrate that EMoE effectively boosts in-domain and out-of-domain generalization abilities. Further analysis and ablation study suggest that EMoE mitigates negative knowledge transfer and is robust to various configurations. Code is available at \url{https://github.com/qiuzh20/EMoE}
Policy constraint methods to offline reinforcement learning (RL) typically utilize parameterization or regularization that constrains the policy to perform actions within the support set of the behavior policy. The elaborative designs of parameterization methods usually intrude into the policy networks, which may bring extra inference cost and cannot take full advantage of well-established online methods. Regularization methods reduce the divergence between the learned policy and the behavior policy, which may mismatch the inherent density-based definition of support set thereby failing to avoid the out-of-distribution actions effectively. This paper presents Supported Policy OpTimization (SPOT), which is directly derived from the theoretical formalization of the density-based support constraint. SPOT adopts a VAE-based density estimator to explicitly model the support set of behavior policy and presents a simple but effective density-based regularization term, which can be plugged non-intrusively into off-the-shelf off-policy RL algorithms. On the standard benchmarks for offline RL, SPOT substantially outperforms state-of-the-art offline RL methods. Benefiting from the pluggable design, the offline pretrained models from SPOT can also be applied to perform online fine-tuning seamlessly.