As the size of language models notably grows, fine-tuning the models becomes more challenging: fine-tuning with first-order optimizers (e.g., SGD and Adam) requires high memory consumption, while fine-tuning with a memory-efficient zeroth-order optimizer (MeZO) has a significant accuracy drop and slower convergence rate. In this work, we propose a Low order Hybrid Optimizer (LoHO) which merges zeroth-order (ZO) and first-order (FO) optimizers for fine-tuning. LoHO is empowered with inter-layer hybrid optimization and intra-layer hybrid optimization, which boosts the accuracy of MeZO while keeping memory usage within a budget. The inter-layer hybrid optimization exploits the FO optimizer in deep layers and the ZO optimizer in shallow ones, therefore avoiding unnecessary gradient propagation to improve memory efficiency. The intra-layer hybrid optimization updates a proportion of parameters in a layer by the ZO optimizer, and the rest by the FO optimizer, taking advantage of gradient sparsity for high efficiency implementation. Our experimental results across common datasets on different pre-trained backbones (i.e., RoBERTa-large, OPT-13B and OPT-30B) demonstrate that LoHO can significantly improve the predictive accuracy and convergence rate of MeZO, while controlling the memory footprint during fine-tuning. Moreover, LoHO can achieve comparable performance with first-order fine-tuning using substantially fewer memory resources.

Large language models (LLMs) have demonstrated impressive capabilities across various tasks, but the billion-scale parameters pose deployment challenges. Although existing methods attempt to reduce the scale of LLMs, they require either special hardware support or expensive post-training to maintain model quality. To facilitate efficient and affordable model slimming, we propose a novel training-free compression method for LLMs, named “SoLA”, which leverages Soft activation sparsity and Low-rAnk decomposition. SoLA can identify and retain a minority of components significantly contributing to inference, while compressing the majority through low-rank decomposition, based on our analysis of the activation pattern in the feed-forward network (FFN) of modern LLMs. To alleviate the decomposition loss, SoLA is equipped with an adaptive component-wise low-rank allocation strategy to assign appropriate truncation positions for different weight matrices. We conduct extensive experiments on LLaMA-2-7B/13B/70B and Mistral-7B models across a variety of benchmarks. SoLA exhibits remarkable improvement in both language modeling and downstream task accuracy without post-training. For example, with a 30% compression rate on the LLaMA-2-70B model, SoLA surpasses the state-of-the-art method by reducing perplexity from 6.95 to 4.44 and enhancing downstream task accuracy by 10%.

Mixture of Experts (MoE) architecture improves Large Language Models (LLMs) with better scaling, but its higher parameter counts and memory demands create challenges for deployment. In this paper, we present MoE-SVD, a new decomposition-based compression framework tailored for MoE LLMs without any extra training. By harnessing the power of Singular Value Decomposition (SVD), MoE-SVD addresses the critical issues of decomposition collapse and matrix redundancy in MoE architectures. Specifically, we first decompose experts into compact low-rank matrices, resulting in accelerated inference and memory optimization. In particular, we propose selective decomposition strategy by measuring sensitivity metrics based on weight singular values and activation statistics to automatically identify decomposable expert layers. Then, we share a single V-matrix across all experts and employ a top-k selection for U-matrices. This low-rank matrix sharing and trimming scheme allows for significant parameter reduction while preserving diversity among experts. Comprehensive experiments on Mixtral, Phi-3.5, DeepSeek, and Qwen2 MoE LLMs show MoE-SVD outperforms other compression methods, achieving a 60% compression ratio and 1.5$\times$ faster inference with minimal performance loss.

As multimedia content continues to grow on the web, the integration of visual and textual data has become a crucial challenge for web applications, particularly in recommendation systems. Large Vision Language Models (LVLMs) have demonstrated considerable potential in addressing this challenge across various tasks that require such multimodal integration. However, their application in multimodal sequential recommendation (MSR) has not been extensively studied. To bridge this gap, we introduce MSRBench, the first comprehensive benchmark designed to systematically evaluate different LVLM integration strategies in web-based recommendation scenarios. We benchmark three state-of-the-art LVLMs, i.e., GPT-4 Vision, GPT-4o, and Claude-3-Opus, on the next item prediction task using the constructed Amazon Review Plus dataset, which includes additional item descriptions generated by LVLMs. Our evaluation examines five integration strategies: using LVLMs as recommender, item enhancer, reranker, and various combinations of these roles. The benchmark results reveal that 1) using LVLMs as rerankers is the most effective strategy, significantly outperforming others that rely on LVLMs to directly generate recommendations or only enhance items; 2) GPT-4o consistently achieves the best performance across most scenarios, particularly when employed as a reranker; 3) the computational inefficiency of LVLMs presents a major barrier to their widespread adoption in real-time multimodal recommendation systems. Our code and datasets are available at https://github.com/PALIN2018/MSRBench.

Sequential recommender systems (SRS) are designed to predict users' future behaviors based on their historical interaction data. Recent research has increasingly utilized contrastive learning (CL) to leverage unsupervised signals to alleviate the data sparsity issue in SRS. In general, CL-based SRS first augments the raw sequential interaction data by using data augmentation strategies and employs a contrastive training scheme to enforce the representations of those sequences from the same raw interaction data to be similar. Despite the growing popularity of CL, data augmentation, as a basic component of CL, has not received sufficient attention. This raises the question: Is it possible to achieve superior recommendation results solely through data augmentation? To answer this question, we benchmark eight widely used data augmentation strategies, as well as state-of-the-art CL-based SRS methods, on four real-world datasets under both warm- and cold-start settings. Intriguingly, the conclusion drawn from our study is that, certain data augmentation strategies can achieve similar or even superior performance compared with some CL-based methods, demonstrating the potential to significantly alleviate the data sparsity issue with fewer computational overhead. We hope that our study can further inspire more fundamental studies on the key functional components of complex CL techniques. Our processed datasets and codes are available at https://github.com/AIM-SE/DA4Rec.