Recent advancements in Artificial Intelligence have led to the development of Multimodal Large Language Models (MLLMs). However, adapting these pre-trained models to dynamic data distributions and various tasks efficiently remains a challenge. Fine-tuning MLLMs for specific tasks often causes performance degradation in the model's prior knowledge domain, a problem known as 'Catastrophic Forgetting'. While this issue has been well-studied in the Continual Learning (CL) community, it presents new challenges for MLLMs. This review paper, the first of its kind in MLLM continual learning, presents an overview and analysis of 440 research papers in this area.The review is structured into four sections. First, it discusses the latest research on MLLMs, covering model innovations, benchmarks, and applications in various fields. Second, it categorizes and overviews the latest studies on continual learning, divided into three parts: non-large language models unimodal continual learning (Non-LLM Unimodal CL), non-large language models multimodal continual learning (Non-LLM Multimodal CL), and continual learning in large language models (CL in LLM). The third section provides a detailed analysis of the current state of MLLM continual learning research, including benchmark evaluations, architectural innovations, and a summary of theoretical and empirical studies.Finally, the paper discusses the challenges and future directions of continual learning in MLLMs, aiming to inspire future research and development in the field. This review connects the foundational concepts, theoretical insights, method innovations, and practical applications of continual learning for multimodal large models, providing a comprehensive understanding of the research progress and challenges in this field, aiming to inspire researchers in the field and promote the advancement of related technologies.

Speech representation models are highly effective at extracting general features for various tasks. While fine-tuning can enhance these representations for specific applications, it often compromises their generalization ability. To address this challenge, we propose Speech-FT, a fine-tuning strategy for speech representation models that leverages model merging to preserve generalization ability while still benefiting from fine-tuning. Speech-FT is effective across different fine-tuning scenarios and is compatible with various types of speech representation models, providing a versatile solution. Speech-FT offers an efficient and practical approach to further improving general speech representations after pre-training.

Low-rank Adaptation (LoRA) has demonstrated remarkable capabilities for task specific fine-tuning. However, in scenarios that involve multiple tasks, training a separate LoRA model for each one results in considerable inefficiency in terms of storage and inference. Moreover, existing parameter generation methods fail to capture the correlations among these tasks, making multi-task LoRA parameter generation challenging. To address these limitations, we propose In-Context Meta LoRA (ICM-LoRA), a novel approach that efficiently achieves task-specific customization of large language models (LLMs). Specifically, we use training data from all tasks to train a tailored generator, Conditional Variational Autoencoder (CVAE). CVAE takes task descriptions as inputs and produces task-aware LoRA weights as outputs. These LoRA weights are then merged with LLMs to create task-specialized models without the need for additional fine-tuning. Furthermore, we utilize in-context meta-learning for knowledge enhancement and task mapping, to capture the relationship between tasks and parameter distributions. As a result, our method achieves more accurate LoRA parameter generation for diverse tasks using CVAE. ICM-LoRA enables more accurate LoRA parameter reconstruction than current parameter reconstruction methods and is useful for implementing task-specific enhancements of LoRA parameters. At the same time, our method occupies 283MB, only 1\% storage compared with the original LoRA.

Quantum computing is a transformative technology with wide-ranging applications, and efficient quantum circuit generation is crucial for unlocking its full potential. Current diffusion model approaches based on U-Net architectures, while promising, encounter challenges related to computational efficiency and modeling global context. To address these issues, we propose UDiT,a novel U-Net-style Diffusion Transformer architecture, which combines U-Net's strengths in multi-scale feature extraction with the Transformer's ability to model global context. We demonstrate the framework's effectiveness on two tasks: entanglement generation and unitary compilation, where UDiTQC consistently outperforms existing methods. Additionally, our framework supports tasks such as masking and editing circuits to meet specific physical property requirements. This dual advancement, improving quantum circuit synthesis and refining generative model architectures, marks a significant milestone in the convergence of quantum computing and machine learning research.

Multi-view 3D reconstruction remains a core challenge in computer vision, particularly in applications requiring accurate and scalable representations across diverse perspectives. Current leading methods such as DUSt3R employ a fundamentally pairwise approach, processing images in pairs and necessitating costly global alignment procedures to reconstruct from multiple views. In this work, we propose Fast 3D Reconstruction (Fast3R), a novel multi-view generalization to DUSt3R that achieves efficient and scalable 3D reconstruction by processing many views in parallel. Fast3R's Transformer-based architecture forwards N images in a single forward pass, bypassing the need for iterative alignment. Through extensive experiments on camera pose estimation and 3D reconstruction, Fast3R demonstrates state-of-the-art performance, with significant improvements in inference speed and reduced error accumulation. These results establish Fast3R as a robust alternative for multi-view applications, offering enhanced scalability without compromising reconstruction accuracy.

Fine-tuning helps large language models (LLM) recover degraded information and enhance task performance. Although Low-Rank Adaptation (LoRA) is widely used and effective for fine-tuning, we have observed that its scaling factor can limit or even reduce performance as the rank size increases. To address this issue, we propose RoRA (Rank-adaptive Reliability Optimization), a simple yet effective method for optimizing LoRA's scaling factor. By replacing $\alpha/r$ with $\alpha/\sqrt{r}$, RoRA ensures improved performance as rank size increases. Moreover, RoRA enhances low-rank adaptation in fine-tuning uncompressed models and excels in the more challenging task of accuracy recovery when fine-tuning pruned models. Extensive experiments demonstrate the effectiveness of RoRA in fine-tuning both uncompressed and pruned models. RoRA surpasses the state-of-the-art (SOTA) in average accuracy and robustness on LLaMA-7B/13B, LLaMA2-7B, and LLaMA3-8B, specifically outperforming LoRA and DoRA by 6.5% and 2.9% on LLaMA-7B, respectively. In pruned model fine-tuning, RoRA shows significant advantages; for SHEARED-LLAMA-1.3, a LLaMA-7B with 81.4% pruning, RoRA achieves 5.7% higher average accuracy than LoRA and 3.9% higher than DoRA.

Training transformer-based encoder-decoder models for long document summarization poses a significant challenge due to the quadratic memory consumption during training. Several approaches have been proposed to extend the input length at test time, but training with these approaches is still difficult, requiring truncation of input documents and causing a mismatch between training and test conditions. In this work, we propose CachED (Gradient $\textbf{Cach}$ing for $\textbf{E}$ncoder-$\textbf{D}$ecoder models), an approach that enables end-to-end training of existing transformer-based encoder-decoder models, using the entire document without truncation. Specifically, we apply non-overlapping sliding windows to input documents, followed by fusion in decoder. During backpropagation, the gradients are cached at the decoder and are passed through the encoder in chunks by re-computing the hidden vectors, similar to gradient checkpointing. In the experiments on long document summarization, we extend BART to CachED BART, processing more than 500K tokens during training and achieving superior performance without using any additional parameters.

Large language models (LLMs) show impressive performance in solving complex language tasks. However, its large number of parameters present significant challenges for the deployment and application of the model on edge devices. Compressing large language models to low bits can enable them to run on resource-constrained devices, often leading to performance degradation. To address this problem, we propose gradient-aware weight quantization (GWQ), the first quantization approach for low-bit weight quantization that leverages gradients to localize outliers, requiring only a minimal amount of calibration data for outlier detection. GWQ retains the weights corresponding to the top 1% outliers preferentially at FP16 precision, while the remaining non-outlier weights are stored in a low-bit format. GWQ found experimentally that utilizing the sensitive weights in the gradient localization model is more scientific compared to utilizing the sensitive weights in the Hessian matrix localization model. Compared to current quantization methods, GWQ can be applied to multiple language models and achieves lower PPL on the WikiText2 and C4 dataset. In the zero-shot task, GWQ quantized models have higher accuracy compared to other quantization methods. GWQ is also suitable for multimodal model quantization, and the quantized Qwen-VL family model is more accurate than other methods. Zero-shot target detection task dataset RefCOCO outperforms the current stat-of-the-arts method SPQR. GWQ achieves 1.2 times inference speedup in comparison to the original model, and effectively reduces the inference memory.

When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for induction (inferring latent functions) and transduction (directly predicting the test output for a given test input). We train on synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.

Neural networks have emerged as powerful tools across various applications, yet their decision-making process often remains opaque, leading to them being perceived as "black boxes." This opacity raises concerns about their interpretability and reliability, especially in safety-critical scenarios. Network inversion techniques offer a solution by allowing us to peek inside these black boxes, revealing the features and patterns learned by the networks behind their decision-making processes and thereby provide valuable insights into how neural networks arrive at their conclusions, making them more interpretable and trustworthy. This paper presents a simple yet effective approach to network inversion using a meticulously conditioned generator that learns the data distribution in the input space of the trained neural network, enabling the reconstruction of inputs that would most likely lead to the desired outputs. To capture the diversity in the input space for a given output, instead of simply revealing the conditioning labels to the generator, we encode the conditioning label information into vectors and intermediate matrices and further minimize the cosine similarity between features of the generated images. Additionally, we incorporate feature orthogonality as a regularization term to boost image diversity which penalises the deviations of the Gram matrix of the features from the identity matrix, ensuring orthogonality and promoting distinct, non-redundant representations for each label. The paper concludes by exploring immediate applications of the proposed network inversion approach in interpretability, out-of-distribution detection, and training data reconstruction.