The rise of foundation models has transformed machine learning research, prompting efforts to uncover their inner workings and develop more efficient and reliable applications for better control. While significant progress has been made in interpreting Large Language Models (LLMs), multimodal foundation models (MMFMs) - such as contrastive vision-language models, generative vision-language models, and text-to-image models - pose unique interpretability challenges beyond unimodal frameworks. Despite initial studies, a substantial gap remains between the interpretability of LLMs and MMFMs. This survey explores two key aspects: (1) the adaptation of LLM interpretability methods to multimodal models and (2) understanding the mechanistic differences between unimodal language models and crossmodal systems. By systematically reviewing current MMFM analysis techniques, we propose a structured taxonomy of interpretability methods, compare insights across unimodal and multimodal architectures, and highlight critical research gaps.

Zeroth-order optimization (ZO) has demonstrated remarkable promise in efficient fine-tuning tasks for Large Language Models (LLMs). In particular, recent advances incorporate the low-rankness of gradients, introducing low-rank ZO estimators to further reduce GPU memory consumption. However, most existing works focus solely on the low-rankness of each individual gradient, overlooking a broader property shared by all gradients throughout the training, i.e., all gradients approximately reside within a similar subspace. In this paper, we consider two factors together and propose a novel low-rank ZO estimator, TeZO, which captures the low-rankness across both the model and temporal dimension. Specifically, we represent ZO perturbations along the temporal dimension as a 3D tensor and employ Canonical Polyadic Decomposition (CPD) to extract each low-rank 2D matrix, significantly reducing the training cost. TeZO can also be easily extended to the Adam variant while consuming less memory than MeZO-SGD, and requiring about only 35% memory of MeZO-Adam. Both comprehensive theoretical analysis and extensive experimental research have validated its efficiency, achieving SOTA-comparable results with lower overhead of time and memory.

Generative models have achieved remarkable success across numerous applications, showcasing their versatility and effectiveness in domains such as image synthesis, natural language processing, and scientific discovery (Achiam et al. 2023; Goodfellow et al. 2014; Karras et al. 2020; Van Den Oord et al. 2016). While extensive research has focused on developing and refining generative models, comparatively less attention has been given to evaluating these models. Evaluating generative models is essential for quantifying the quality of their outputs and identifying the best model when comparing multiple options. Evaluating a generative model is significantly more challenging than the evaluation of a predictor or a classifier. To evaluate the performance of prediction or classification, we can directly compare the model's output with the true label. In contrast, the quality of a generative model is determined by how closely the distribution of its generated data matches that of the input data, rather than the similarity between generated data points and input data points (also known as the reconstruction error).

Predicting molecular properties is essential for drug discovery, and computational methods can greatly enhance this process. Molecular graphs have become a focus for representation learning, with Graph Neural Networks (GNNs) widely used. However, GNNs often struggle with capturing long-range dependencies. To address this, we propose MolGraph-xLSTM, a novel graph-based xLSTM model that enhances feature extraction and effectively models molecule long-range interactions. Our approach processes molecular graphs at two scales: atom-level and motif-level. For atom-level graphs, a GNN-based xLSTM framework with jumping knowledge extracts local features and aggregates multilayer information to capture both local and global patterns effectively. Motif-level graphs provide complementary structural information for a broader molecular view. Embeddings from both scales are refined via a multi-head mixture of experts (MHMoE), further enhancing expressiveness and performance. We validate MolGraph-xLSTM on 10 molecular property prediction datasets, covering both classification and regression tasks. Our model demonstrates consistent performance across all datasets, with improvements of up to 7.03% on the BBBP dataset for classification and 7.54% on the ESOL dataset for regression compared to baselines. On average, MolGraph-xLSTM achieves an AUROC improvement of 3.18\% for classification tasks and an RMSE reduction of 3.83\% across regression datasets compared to the baseline methods. These results confirm the effectiveness of our model, offering a promising solution for molecular representation learning for drug discovery.

Large Language Models (LLMs) have revolutionized natural language processing (NLP) by delivering state-of-the-art performance across a variety of tasks. Among these, Transformer-based models like BERT and GPT rely on pooling layers to aggregate token-level embeddings into sentence-level representations. Common pooling mechanisms such as Mean, Max, and Weighted Sum play a pivotal role in this aggregation process. Despite their widespread use, the comparative performance of these strategies on different LLM architectures remains underexplored. To address this gap, this paper investigates the effects of these pooling mechanisms on two prominent LLM families -- BERT and GPT, in the context of sentence-level sentiment analysis. Comprehensive experiments reveal that each pooling mechanism exhibits unique strengths and weaknesses depending on the task's specific requirements. Our findings underline the importance of selecting pooling methods tailored to the demands of particular applications, prompting a re-evaluation of common assumptions regarding pooling operations. By offering actionable insights, this study contributes to the optimization of LLM-based models for downstream tasks.

Missing data is a pervasive challenge in wireless networks and many other domains, often compromising the performance of machine learning and deep learning models. To address this, we propose a novel framework, FGATT, that combines the Fuzzy Graph Attention Network (FGAT) with the Transformer encoder to perform robust and accurate data imputation. FGAT leverages fuzzy rough sets and graph attention mechanisms to capture spatial dependencies dynamically, even in scenarios where predefined spatial information is unavailable. The Transformer encoder is employed to model temporal dependencies, utilizing its self-attention mechanism to focus on significant time-series patterns. A self-adaptive graph construction method is introduced to enable dynamic connectivity learning, ensuring the framework's applicability to a wide range of wireless datasets. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods in imputation accuracy and robustness, particularly in scenarios with substantial missing data. The proposed model is well-suited for applications in wireless sensor networks and IoT environments, where data integrity is critical.

Averaging iterations of Stochastic Gradient Descent (SGD) have achieved empirical success in training deep learning models, such as Stochastic Weight Averaging (SWA), Exponential Moving Average (EMA), and LAtest Weight Averaging (LAWA). Especially, with a finite weight averaging method, LAWA can attain faster convergence and better generalization. However, its theoretical explanation is still less explored since there are fundamental differences between finite and infinite settings. In this work, we first generalize SGD and LAWA as Finite Weight Averaging (FWA) and explain their advantages compared to SGD from the perspective of optimization and generalization. A key challenge is the inapplicability of traditional methods in the sense of expectation or optimal values for infinite-dimensional settings in analyzing FWA's convergence. Second, the cumulative gradients introduced by FWA introduce additional confusion to the generalization analysis, especially making it more difficult to discuss them under different assumptions. Extending the final iteration convergence analysis to the FWA, this paper, under a convexity assumption, establishes a convergence bound $\mathcal{O}(\log\left(\frac{T}{k}\right)/\sqrt{T})$, where $k\in[1, T/2]$ is a constant representing the last $k$ iterations. Compared to SGD with $\mathcal{O}(\log(T)/\sqrt{T})$, we prove theoretically that FWA has a faster convergence rate and explain the effect of the number of average points. In the generalization analysis, we find a recursive representation for bounding the cumulative gradient using mathematical induction. We provide bounds for constant and decay learning rates and the convex and non-convex cases to show the good generalization performance of FWA. Finally, experimental results on several benchmarks verify our theoretical results.

In recent years, Transformer-based models (Transformers) have achieved significant success in multivariate time series forecasting (MTSF). However, previous works focus on extracting features either from the time domain or the frequency domain, which inadequately captures the trends and periodic characteristics. To address this issue, we propose a wavelet learning framework to model complex temporal dependencies of the time series data. The wavelet domain integrates both time and frequency information, allowing for the analysis of local characteristics of signals at different scales. Additionally, the Softmax self-attention mechanism used by Transformers has quadratic complexity, which leads to excessive computational costs when capturing long-term dependencies. Therefore, we propose a novel attention mechanism: Rotary Route Attention (RoRA). Unlike Softmax attention, RoRA utilizes rotary position embeddings to inject relative positional information to sequence tokens and introduces a small number of routing tokens $r$ to aggregate information from the $KV$ matrices and redistribute it to the $Q$ matrix, offering linear complexity. We further propose WaveRoRA, which leverages RoRA to capture inter-series dependencies in the wavelet domain. We conduct extensive experiments on eight real-world datasets. The results indicate that WaveRoRA outperforms existing state-of-the-art models while maintaining lower computational costs. Our code is available at https://github.com/Leopold2333/WaveRoRA.

Minimax optimization is gaining increasing attention in modern machine learning applications. Driven by large-scale models and massive volumes of data collected from edge devices, as well as the concern to preserve client privacy, communication-efficient distributed minimax optimization algorithms become popular, such as Local Stochastic Gradient Descent Ascent (Local-SGDA), and Local Decentralized SGDA (Local-DSGDA). While most existing research on distributed minimax algorithms focuses on convergence rates, computation complexity, and communication efficiency, the generalization performance remains underdeveloped, whereas generalization ability is a pivotal indicator for evaluating the holistic performance of a model when fed with unknown data. In this paper, we propose the stability-based generalization analytical framework for Distributed-SGDA, which unifies two popular distributed minimax algorithms including Local-SGDA and Local-DSGDA, and conduct a comprehensive analysis of stability error, generalization gap, and population risk across different metrics under various settings, e.g., (S)C-(S)C, PL-SC, and NC-NC cases. Our theoretical results reveal the trade-off between the generalization gap and optimization error and suggest hyperparameters choice to obtain the optimal population risk.

Large pretrained transformer models have revolutionized modern AI applications with their state-of-the-art performance in natural language processing (NLP). However, their substantial parameter count poses challenges for real-world deployment. To address this, researchers often reduce model size by pruning parameters based on their magnitude or sensitivity. Previous research has demonstrated the limitations of magnitude pruning, especially in the context of transfer learning for modern NLP tasks. In this paper, we introduce a new magnitude-based pruning algorithm called mixture Gaussian prior pruning (MGPP), which employs a mixture Gaussian prior for regularization. MGPP prunes non-expressive weights under the guidance of the mixture Gaussian prior, aiming to retain the model's expressive capability. Extensive evaluations across various NLP tasks, including natural language understanding, question answering, and natural language generation, demonstrate the superiority of MGPP over existing pruning methods, particularly in high sparsity settings. Additionally, we provide a theoretical justification for the consistency of the sparse transformer, shedding light on the effectiveness of the proposed pruning method.