Large Language Models (LLMs) offer a transparent brain with accessible parameters that encode extensive knowledge, which can be analyzed, located and transferred. Consequently, a key research challenge is to transcend traditional knowledge transfer paradigms rooted in symbolic language and achieve genuine Parametric Knowledge Transfer (PKT). Significantly, exploring effective methods for transferring knowledge across LLMs of different scales through parameters presents an intriguing and valuable research direction. In this paper, we first demonstrate $\textbf{Alignment}$ in parametric space is the fundamental prerequisite to achieve successful cross-scale PKT. We redefine the previously explored knowledge transfer as Post-Align PKT (PostPKT), which utilizes extracted parameters for LoRA initialization and requires subsequent fine-tune for alignment. Hence, to reduce cost for further fine-tuning, we introduce a novel Pre-Align PKT (PrePKT) paradigm and propose a solution called $\textbf{LaTen}$ ($\textbf{L}$oc$\textbf{a}$te-$\textbf{T}$h$\textbf{e}$n-Alig$\textbf{n}$) that aligns the parametric spaces of LLMs across scales only using several training steps without following training. Comprehensive experiments on four benchmarks demonstrate that both PostPKT and PrePKT face challenges in achieving consistently stable transfer. Through in-depth analysis, we identify $\textbf{Neural Incompatibility}$ as the ethological and parametric structural differences between LLMs of varying scales, presenting fundamental challenges to achieving effective PKT. These findings provide fresh insights into the parametric architectures of LLMs and highlight promising directions for future research on efficient PKT. Our code is available at https://github.com/Trae1ounG/Neural_Incompatibility.

Large language models (LLMs) have shown great potential as general-purpose AI assistants across various domains. To fully leverage this potential in specific applications, many companies provide fine-tuning API services, enabling users to upload their own data for LLM customization. However, fine-tuning services introduce a new safety threat: user-uploaded data, whether harmful or benign, can break the model's alignment, leading to unsafe outputs. Moreover, existing defense methods struggle to address the diversity of fine-tuning datasets (e.g., varying sizes, tasks), often sacrificing utility for safety or vice versa. To address this issue, we propose Safe Delta, a safety-aware post-training defense method that adjusts the delta parameters (i.e., the parameter change before and after fine-tuning). Specifically, Safe Delta estimates the safety degradation, selects delta parameters to maximize utility while limiting overall safety loss, and applies a safety compensation vector to mitigate residual safety loss. Through extensive experiments on four diverse datasets with varying settings, our approach consistently preserves safety while ensuring that the utility gain from benign datasets remains unaffected.

Transformer-based models with the pretrain-finetune paradigm bring about significant progress, along with the heavy storage and deployment costs of finetuned models on multiple tasks. Delta compression attempts to lower the costs by reducing the redundancy of delta parameters (i.e., the difference between the finetuned and pre-trained model weights) through pruning or quantization. However, existing methods by default employ the pretrained model as the base model and compress the delta parameters for every task, which may causes significant performance degradation, especially when the compression rate is extremely high. To tackle this issue, we investigate the impact of different base models on the performance of delta compression and find that the pre-trained base model can hardly be optimal. To this end, we propose Dynamic Base Model Shift (DBMS), which dynamically adapts the base model to the target task before performing delta compression. Specifically, we adjust two parameters, which respectively determine the magnitude of the base model shift and the overall scale of delta compression, to boost the compression performance on each task. Through low-cost learning of these two parameters, our DBMS can maintain most of the finetuned model's performance even under an extremely high compression ratio setting, significantly surpassing existing methods. Moreover, our DBMS is orthogonal and can be integrated with a variety of other methods, and it has been evaluated across different types of models including language, vision transformer, and multi-modal models.

With the proliferation of task-specific large language models, delta compression has emerged as a method to mitigate the resource challenges of deploying numerous such models by effectively compressing the delta model parameters. Previous delta-sparsification methods either remove parameters randomly or truncate singular vectors directly after singular value decomposition (SVD). However, these methods either disregard parameter importance entirely or evaluate it with too coarse a granularity. In this work, we introduce ImPart, a novel importance-aware delta sparsification approach. Leveraging SVD, it dynamically adjusts sparsity ratios of different singular vectors based on their importance, effectively retaining crucial task-specific knowledge even at high sparsity ratios. Experiments show that ImPart achieves state-of-the-art delta sparsification performance, demonstrating $2\times$ higher compression ratio than baselines at the same performance level. When integrated with existing methods, ImPart sets a new state-of-the-art on delta quantization and model merging.

Model merging has achieved significant success, with numerous innovative methods proposed to enhance capabilities by combining multiple models. However, challenges persist due to the lack of a unified framework for classification and systematic comparative analysis, leading to inconsistencies in terminologies and categorizations. Meanwhile, as an increasing number of fine-tuned models are publicly available, their original training data often remain inaccessible due to privacy concerns or intellectual property restrictions. This makes traditional multi-task learning based on shared training data impractical. In scenarios where direct access to training data is infeasible, merging model parameters to create a unified model with broad generalization across multiple domains becomes crucial, further underscoring the importance of model merging techniques. Despite the rapid progress in this field, a comprehensive taxonomy and survey summarizing recent advances and predicting future directions are still lacking. This paper addresses these gaps by establishing a new taxonomy of model merging methods, systematically comparing different approaches, and providing an overview of key developments. By offering a structured perspective on this evolving area, we aim to help newcomers quickly grasp the field's landscape and inspire further innovations.

With transformer-based models and the pretrain-finetune paradigm becoming mainstream, the high storage and deployment costs of individual finetuned models on multiple tasks pose critical challenges. Delta compression attempts to lower the costs by reducing the redundancy of delta parameters (i.e., the difference between the finetuned and pre-trained model weights). However, existing methods usually face problems including data accessibility and training requirements. To tackle this issue, we introduce Delta-DCT, the first data-free delta compression method inspired by classic JPEG image compression, leveraging the Discrete Cosine Transform (DCT). We first (a) group delta parameters within a layer into patches. Then we (b) assess the importance of each patch and allocate them with different quantization bit-widths. Afterwards, we (c) convert these patches to the DCT domain and conduct quantization to each patch based on the allocated bit-width. The proposed Delta-DCT does not require any training or data calibration, while achieving performance comparable to or even surpassing original finetuned models under 1-bit equivalent delta compression ratios on different kinds of models including: (1) recently-released LLMs of different sizes from 7B to 13B, (2) relatively smaller language models including RoBERTa and T5 models, (3) variants of vision transformer models, and (4) multi-modal BEiT-3 models.

Lately, the practice of utilizing task-specific fine-tuning has been implemented to improve the performance of large language models (LLM) in subsequent tasks. Through the integration of diverse LLMs, the overall competency of LLMs is significantly boosted. Nevertheless, traditional ensemble methods are notably memory-intensive, necessitating the simultaneous loading of all specialized models into GPU memory. To address the inefficiency, model merging strategies have emerged, merging all LLMs into one model to reduce the memory footprint during inference. Despite these advances, model merging often leads to parameter conflicts and performance decline as the number of experts increases. Previous methods to mitigate these conflicts include post-pruning and partial merging. However, both approaches have limitations, particularly in terms of performance and storage efficiency when merged experts increase. To address these challenges, we introduce Channel Merging, a novel strategy designed to minimize parameter conflicts while enhancing storage efficiency. This method clusters and merges channel parameters based on their similarity to form several groups offline. By ensuring that only highly similar parameters are merged within each group, it significantly reduces parameter conflicts. During inference, we can instantly look up the expert parameters from the merged groups, preserving specialized knowledge. Our experiments demonstrate that Channel Merging consistently delivers high performance, matching unmerged models in tasks like English and Chinese reasoning, mathematical reasoning, and code generation. Moreover, it obtains results comparable to model ensemble with just 53% parameters when used with a task-specific router.

Although large language models (LLMs) achieve effective safety alignment at the time of release, they still face various safety challenges. A key issue is that fine-tuning often compromises the safety alignment of LLMs. To address this issue, we propose a method named \textbf{IRR} (\textbf{I}dentify, \textbf{R}emove, and \textbf{R}ecalibrate for Safety Realignment) that performs safety realignment for LLMs. The core of IRR is to identify and remove unsafe delta parameters from the fine-tuned models, while recalibrating the retained ones. We evaluate the effectiveness of IRR across various datasets, including both full fine-tuning and LoRA methods. Our results demonstrate that IRR significantly enhances the safety performance of fine-tuned models on safety benchmarks, such as harmful queries and jailbreak attacks, while maintaining their performance on downstream tasks. The source code is available at: \url{https://anonymous.4open.science/r/IRR-BD4F}.

Multimodal LLMs (MLLMs) equip language models with visual capabilities by aligning vision encoders with language models. Existing methods to enhance the visual perception of MLLMs often involve designing more powerful vision encoders, which requires exploring a vast design space and re-aligning each potential encoder with the language model, resulting in prohibitively high training costs. In this paper, we introduce VisionFuse, a novel integration framework that efficiently utilizes multiple vision encoders from off-the-shelf MLLMs to enhance visual perception without requiring additional training. Our approach is motivated by the observation that different MLLMs tend to focus on distinct regions given the same query and image. Moreover, we find that the feature distributions of vision encoders within an MLLM family, a group of MLLMs sharing the same pretrained LLM, are highly aligned. Building on these insights, VisionFuse enriches the visual context by concatenating the tokens generated by the vision encoders of selected MLLMs within a family. By merging the parameters of language models from these MLLMs, VisionFuse allows a single language model to align with various vision encoders, significantly reducing deployment overhead. We conduct comprehensive evaluations across multiple multimodal benchmarks using various MLLM combinations, demonstrating substantial improvements in multimodal tasks. Notably, when integrating MiniGemini-8B and SLIME-8B, VisionFuse achieves an average performance increase of over 4%.

Continual learning methods based on pre-trained models (PTM) have recently gained attention which adapt to successive downstream tasks without catastrophic forgetting. These methods typically refrain from updating the pre-trained parameters and instead employ additional adapters, prompts, and classifiers. In this paper, we from a novel perspective investigate the benefit of sparse orthogonal parameters for continual learning. We found that merging sparse orthogonality of models learned from multiple streaming tasks has great potential in addressing catastrophic forgetting. Leveraging this insight, we propose a novel yet effective method called SoTU (Sparse Orthogonal Parameters TUning). We hypothesize that the effectiveness of SoTU lies in the transformation of knowledge learned from multiple domains into the fusion of orthogonal delta parameters. Experimental evaluations on diverse CL benchmarks demonstrate the effectiveness of the proposed approach. Notably, SoTU achieves optimal feature representation for streaming data without necessitating complex classifier designs, making it a Plug-and-Play solution.