Multi-token prediction (MTP) is a prominent strategy to significantly speed up generation in large language models (LLMs), including byte-level LLMs, which are tokeniser-free but prohibitively slow. However, existing MTP methods often sacrifice expressiveness by assuming independence between future tokens. In this work, we investigate the trade-off between expressiveness and latency in MTP within the framework of probabilistic circuits (PCs). Our framework, named MTPC, allows one to explore different ways to encode the joint distributions over future tokens by selecting different circuit architectures, generalising classical models such as (hierarchical) mixture models, hidden Markov models and tensor networks. We show the efficacy of MTPC by retrofitting existing byte-level LLMs, such as EvaByte. Our experiments show that, when combined with speculative decoding, MTPC significantly speeds up generation compared to MTP with independence assumptions, while guaranteeing to retain the performance of the original verifier LLM. We also rigorously study the optimal trade-off between expressiveness and latency when exploring the possible parameterisations of MTPC, such as PC architectures and partial layer sharing between the verifier and draft LLMs.

In this paper, we study the role of initialization in Low Rank Adaptation (LoRA) as originally introduced in Hu et al. [19]. Essentially, to start from the pretrained model as initialization for finetuning, one can either initialize B to zero and A to random (default initialization in PEFT package), or vice-versa. In both cases, the product BA is equal to zero at initialization, which makes finetuning starts from the pretrained model. These two initialization schemes are seemingly similar. They should in-principle yield the same performance and share the same optimal learning rate. We demonstrate that this is an incorrect intuition and that the first scheme (initializing B to zero and A to random) on average yields better performance compared to the other scheme. Our theoretical analysis shows that the reason behind this might be that the first initialization allows the use of larger learning rates (without causing output instability) compared to the second initialization, resulting in more efficient learning of the first scheme.

Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enhances adaptability while reducing computational costs. However, fine-tuning can compromise safety alignment, even with benign data, increasing susceptibility to harmful outputs. Existing safety alignment methods struggle to capture complex parameter shifts, leading to suboptimal safety-utility trade-offs. To address this issue, we propose Safe Pruning LoRA (SPLoRA), a novel pruning-based approach that selectively removes LoRA layers that weaken safety alignment, improving safety while preserving performance. At its core, we introduce Empirical-DIEM (E-DIEM), a dimension-insensitive similarity metric that effectively detects safety misalignment in LoRA-adapted models. We conduct extensive experiments on LLMs fine-tuned with mixed of benign and malicious data, and purely benign datasets, evaluating SPLoRA across utility, safety, and reliability metrics. Results demonstrate that SPLoRA outperforms state-of-the-art safety alignment techniques, significantly reducing safety risks while maintaining or improving model performance and reliability. Additionally, SPLoRA reduces inference overhead, making it a scalable and efficient solution for deploying safer and more reliable LLMs. The code is available at https://github.com/AoShuang92/SPLoRA.

--Federated Learning (FL) has recently been utilized to collaboratively fine-tune foundation models (FMs) across multiple clients. Notably, federated low-rank adaptation (LoRA)- based fine-tuning methods have recently gained attention, which allows clients to fine-tune FMs with a small portion of train-able parameters locally. However, most existing methods do not account for the heterogeneous resources of clients or lack an effective local training strategy to maximize global fine-tuning performance under limited resources. In this work, we propose Fed-HeLLo, a novel federated LoRA-based fine-tuning framework that enables clients to collaboratively fine-tune an FM with different local trainable LoRA layers. T o ensure its effectiveness, we develop several heterogeneous LoRA allocation (HLA) strategies that adaptively allocate local trainable LoRA layers based on clients' resource capabilities and the layer importance. Specifically, based on the dynamic layer importance, we design a Fisher Information Matrix score-based HLA (FIM-HLA) that leverages dynamic gradient norm information. T o better stabilize the training process, we consider the intrinsic importance of LoRA layers and design a Geometrically-Defined HLA (GD-HLA) strategy. It shapes the collective distribution of trainable LoRA layers into specific geometric patterns, such as Triangle, Inverted Triangle, Bottleneck, and Uniform. Moreover, we extend GD-HLA into a randomized version, named Randomized Geometrically-Defined HLA (RGD-HLA), for enhanced model accuracy with randomness. By co-designing the proposed HLA strategies, we incorporate both the dynamic and intrinsic layer importance into the design of our HLA strategy. T o thoroughly evaluate our approach, we simulate various complex federated LoRA-based fine-tuning settings using five datasets and three levels of data distributions ranging from IID to extreme Non-IID. The experimental results demonstrate the effectiveness and efficiency of Fed-HeLLo with the proposed HLA strategies. OUNDA TION models (FMs) [13], [16], [36], [37], [68], characterized by their extensive parameter counts ranging into millions or billions, serve as robust initial weights for a variety of downstream tasks [47], [52] via fine-tuning. However, employing FMs presents substantial challenges, especially the high computational costs of fine-tuning the model. To mitigate the high computational requirement of fine-tuning FMs, researchers have developed a variety of parameter-efficient fine-tuning (PEFT) methods.

We introduce Vision as LoRA (VoRA), a novel paradigm for transforming an LLM into an MLLM. Unlike prevalent MLLM architectures that rely on external vision modules for vision encoding, VoRA internalizes visual capabilities by integrating vision-specific LoRA layers directly into the LLM. This design allows the added parameters to be seamlessly merged into the LLM during inference, eliminating structural complexity and minimizing computational overhead. Moreover, inheriting the LLM's ability of handling flexible context, VoRA can process inputs at arbitrary resolutions. To further strengthen VoRA's visual capabilities, we introduce a block-wise distillation method that transfers visual priors from a pre-trained ViT into the LoRA layers, effectively accelerating training by injecting visual knowledge. Additionally, we apply bi-directional attention masks to better capture the context information of an image. We successfully demonstrate that with additional pre-training data, VoRA can perform comparably with conventional encode-based MLLMs. All training data, codes, and model weights will be released at https://github.com/Hon-Wong/VoRA.

Federated fine-tuning (FedFT) has been proposed to fine-tune the pre-trained language models in a distributed manner. However, there are two critical challenges for efficient FedFT in practical applications, i.e., resource constraints and system heterogeneity. Existing works rely on parameter-efficient fine-tuning methods, e.g., low-rank adaptation (LoRA), but with major limitations. Herein, based on the inherent characteristics of FedFT, we observe that LoRA layers with higher ranks added close to the output help to save resource consumption while achieving comparable fine-tuning performance. Then we propose a novel LoRA-based FedFT framework, termed LEGEND, which faces the difficulty of determining the number of LoRA layers (called, LoRA depth) and the rank of each LoRA layer (called, rank distribution). We analyze the coupled relationship between LoRA depth and rank distribution, and design an efficient LoRA configuration algorithm for heterogeneous devices, thereby promoting fine-tuning efficiency. Extensive experiments are conducted on a physical platform with 80 commercial devices. The results show that LEGEND can achieve a speedup of 1.5-2.8$\times$ and save communication costs by about 42.3% when achieving the target accuracy, compared to the advanced solutions.

Low-Rank Adaptation (LoRA) layers have emerged as a promising approach for efficient model fine-tuning, but their capabilities and limitations have not been fully explored. This paper: 1) Investigates the fundamental question of whether LoRA layers are effective at increasing reasoning + planning abilities 2) We introduce HashChain Reasoning, a novel evaluation dataset that deterministically tests reasoning capabilities. Through systematic ablation studies on GPT-2, we demonstrate that reasoning capabilities appear to exist primarily in low-rank spaces and can be effectively enhanced using LoRA layers. The effective rank analysis of trained LoRA matrices reveals a 2-3x lower rank requirement for reasoning tasks compared to planning tasks, giving context on where LoRA layers would be effective. This also provides evidence for reasoning fundamentally preferring low-parameter spaces for generalization.

Low-Rank Adaptation (LoRA) is a parameter-efficient technique for rapidly fine-tuning foundation models. In standard LoRA training dynamics, models tend to quickly converge to a local optimum near the initialization. However, this local optimum may not be ideal for out-of-distribution data or tasks such as merging and pruning. In this work, we propose a novel progressive training strategy for LoRA with random layer dropping. This strategy also optimizes the Shapley value of LoRA parameters in each layer, treating each layer as a player in a cooperative game. We refer to this method as Cooperative LoRA (CopRA). Our experimental results demonstrate that parameters trained with CopRA exhibit linear mode connectivity, which enables efficient model merging. This also paves the way for federated learning and multi-task learning via LoRA merging. Additionally, by optimizing the Shapley value, CopRA shows superior performance in pruning tasks.

Transformer-based Large Language Models (LLMs) traditionally rely on final-layer loss for training and final-layer representations for predictions, potentially overlooking the predictive power embedded in intermediate layers. Surprisingly, we find that focusing training efforts on these intermediate layers can yield training losses comparable to those of final layers, with complementary test-time performance. We introduce a novel tuning framework, Mixture-of-Depths (MoD), which trains late layers as ensembles contributing to the final logits through learned routing weights. With the auxiliary distillation loss and additional normalization modules, we ensure that the outputs of the late layers adapt to language modeling. Our MoD framework, which can be integrated with any existing tuning method, shows consistent improvement on various language modelling tasks. Furthermore, by replacing traditional trainable modules with MoD, our approach achieves similar performance with significantly fewer trainable parameters, demonstrating the potential of leveraging predictive power from intermediate representations during training.