Despite advances using low-rank adapters and quantization, pretraining of large models on consumer hardware has not been possible without model sharding, offloading during training, or per-layer gradient updates. To address these limitations, we propose Low-Rank Adapters for Quantized Training (LoQT), a method for efficiently training quantized models. LoQT uses gradient-based tensor factorization to initialize low-rank trainable weight matrices that are periodically merged into quantized full-rank weight matrices. Our approach is suitable for both pretraining and fine-tuning models. We demonstrate this for language modeling and downstream task adaptation, finding that LoQT enables efficient training of models up to 7B parameters on a 24GB GPU. We also demonstrate the feasibility of training a 13B model using per-layer gradient updates on the same hardware.

Resource-constrained weight deployment is a task of immense practical importance. Recently, there has been interest in the specific task of \textit{Delta Compression}, where parties each hold a common base model and only communicate compressed weight updates. However, popular parameter efficient updates such as Low Rank Adaptation (LoRA) face inherent representation limitations - which are especially pronounced when combined with aggressive quantization. To overcome this, we build on recent work that improves LoRA representation capacity by using fixed-frequency sinusoidal functions to increase stable rank without adding additional parameters. We extend this to the quantized setting and present the first theoretical analysis showing how stable rank evolves under quantization. From this, we introduce SineLoRA$Δ$, a principled and effective method for delta compression that improves the expressivity of quantized low-rank adapters by applying a sinusoidal activation. We validate SineLoRA$Δ$ across a diverse variety of domains - including language modeling, vision-language tasks, and text-to-image generation - achieving up to 66% memory reduction with similar performance. We additionally provide a novel application of the canonical Bjøntegaard Delta metric to consistently compare adapter compression changes across the rate-distortion curve.

Low-rank adaptation (LoRA) has become a widely used paradigm for parameter-efficient fine-tuning of large language models, yet its representational capacity often lags behind full fine-tuning. Within the context of LoRA, a key open question is how to obtain expressive low-rank adapters from over-parameterized spaces. We propose \textit{PrunedLoRA}, a new framework that leverages structured pruning to obtain highly representative low-rank adapters from an over-parameterized initialization. Unlike prior approaches that impose a fixed low-rank budget, PrunedLoRA dynamically prunes less important components during fine-tuning and prevents their reactivation, enabling flexible and adaptive rank allocation. For structured pruning, by minimizing the pruning error for overall loss, we provide fine-grained pruning and recovery updates in a gradient-based pruning strategy with grounded interpretation. We provide the first theoretical analysis of the robustness of structured pruning and provably show that under the impact of weight perturbation, gradient-based pruning is more robust than activation-based pruning with respect to overall loss. Empirically, PrunedLoRA consistently outperforms LoRA and its variants across supervised fine-tuning tasks in mathematical reasoning, code generation, and natural language understanding, and it also demonstrates advantages over existing structured pruning methods across diverse sparsity levels.

Large language models (LLMs) are increasingly deployed with task-specific adapters catering to multiple downstream applications. In such a scenario, the additional compute associated with these apparently insignificant number of adapter parameters (typically less than 1% of the base model) turns out to be disproportionately significant during inference time (upto 2.5x times that of the base model). In this paper, we propose a new zero-latency fused low-rank adapter (zFLoRA) that introduces zero or negligible latency overhead on top of the base model. Experimental results on LLMs of size 1B, 3B and 7B show that zFLoRA compares favorably against the popular supervised fine-tuning benchmarks including low-rank adapters (LoRA) as well as full fine-tuning (FFT). Experiments are conducted on 18 different tasks across three different categories namely commonsense reasoning, math reasoning and summary-dialogue. Latency measurements made on NPU (Samsung Galaxy S25+) as well as GPU (NVIDIA H100) platforms show that the proposed zFLoRA adapters introduce zero to negligible latency overhead.

A long-standing challenge in developing accurate recommendation models is simulating user behavior, mainly due to the complex and stochastic nature of user interactions. Towards this, one promising line of work has been the use of Large Language Models (LLMs) for simulating user behavior. However, aligning these general-purpose large pre-trained models with user preferences necessitates: (i) effectively and continously parsing large-scale tabular user-item interaction data, (ii) overcoming pre-training-induced inductive biases to accurately learn user specific knowledge, and (iii) achieving the former two at scale for millions of users. While most previous works have focused on complex methods to prompt an LLM or fine-tune it on tabular interaction datasets, our approach shifts the focus to extracting robust textual user representations using a frozen LLM and simulating cost-effective, resource-efficient user agents powered by fine-tuned Small Language Models (SLMs). Further, we showcase a method for training multiple low-rank adapters for groups of users or \textit{persona}, striking an optimal balance between scalability and performance of user behavior agents. Our experiments provide compelling empirical evidence of the efficacy of our methods, demonstrating that user agents developed using our approach have the potential to bridge the gap between offline metrics and real-world performance of recommender systems.

Modern neural network based speech recognition models are required to continually absorb new data without re-training the whole system, especially in downstream applications using foundation models, having no access to the original training data. Continually training the models in a rehearsal-free, multilingual, and language agnostic condition, likely leads to catastrophic forgetting, when a seemingly insignificant disruption to the weights can destructively harm the quality of the models. Inspired by the ability of human brains to learn and consolidate knowledge through the waking-sleeping cycle, we propose a continual learning approach with two distinct phases: factorization and centralization, learning and merging knowledge accordingly. Our experiments on a sequence of varied code-switching datasets showed that the centralization stage can effectively prevent catastrophic forgetting by accumulating the knowledge in multiple scattering low-rank adapters.

--Endoscopic surgery is the gold standard for robotic-assisted minimally invasive surgery, offering significant advantages in early disease detection and precise interventions. However, the complexity of surgical scenes, characterized by high variability in different surgical activity scenarios and confused image features between targets and the background, presents challenges for surgical environment understanding. T o address this limitation, we explore multi-task learning, which utilizes the interrelated features between tasks to enhance overall task performance. In this paper, we propose EndoARSS, a novel multi-task learning framework specifically designed for endoscopy surgery activity recognition and semantic segmentation. Built upon the DINOv2 foundation model, our approach integrates Low-Rank Adaptation to facilitate efficient fine-tuning while incorporating T ask Efficient Shared Low-Rank Adapters (TESLA) to mitigate gradient conflicts across diverse tasks. Additionally, we introduce the Spatially-A ware Multi-Scale Attention that enhances feature representation discrimination by enabling cross-spatial learning of global information within complex surgical environments.In order to evaluate the effectiveness of our framework, we present three novel datasets, MTLESD, MTLEndovis and MTLEndovis-Gen, tailored for endoscopic surgery scenarios with detailed annotations for both activity recognition and semantic segmentation tasks. Extensive experiments demonstrate that EndoARSS achieves remarkable performance across multiple benchmarks, significantly improving both accuracy and robustness in comparison to existing models. These results underscore the potential of EndoARSS to advance AI-driven endoscopic surgical systems, offering valuable insights for enhancing surgical safety and efficiency. Endoscopy has become the gold standard for therapeutic interventions, offering significant improvements in early disease detection and advancing robotic-assisted minimally invasive surgery (RMIS) [1]. During endoscopic procedures, the endoscope serves as a critical tool for visualizing surgical instruments and tissues. Despite these advancements, endoscopic surgery remains technically demanding with the high risk of bleeding and perforation complications, requiring both highly skilled surgeons and instruments that offer exceptional flexibility and precision [3], [4]. Augmenting the display with intelligently selected key targets in the endoscopic view can further enhance the surgeon's capabilities.

Despite advances using low-rank adapters and quantization, pretraining of large models on consumer hardware has not been possible without model sharding, offloading during training, or per-layer gradient updates. To address these limitations, we propose Low-Rank Adapters for Quantized Training (LoQT), a method for efficiently training quantized models. LoQT uses gradient-based tensor factorization to initialize low-rank trainable weight matrices that are periodically merged into quantized full-rank weight matrices. Our approach is suitable for both pretraining and fine-tuning models. We demonstrate this for language modeling and downstream task adaptation, finding that LoQT enables efficient training of models up to 7B parameters on a 24GB GPU.

In this paper, we introduce Symmetric Low-Rank Adapters, an optimized variant of LoRA with even fewer weights. This method utilizes Low-Rank Symmetric Weight Matrices to learn downstream tasks more efficiently. Traditional LoRA accumulates fine-tuning weights with the original pre-trained weights via a Singular Value Decomposition (SVD) like approach, i.e., model weights are fine-tuned via updates of the form $BA$ (where $B \in \mathbb{R}^{n\times r}$, $A \in \mathbb{R}^{r\times n}$, and $r$ is the rank of the merged weight matrix). In contrast, our approach, named SymLoRA, represents fine-tuning weights as a Spectral Decomposition, i.e., $Q \, diag(Λ)\, Q^T$, where $Q \in \mathbb{R}^{n\times r}$ and $Λ\in \mathbb{R}^r$. SymLoRA requires approximately half of the finetuning weights. Here, we show that this approach has negligible losses in downstream efficacy.

The adoption of Foundation Models in resource-constrained environments remains challenging due to their large size and inference costs. A promising way to overcome these limitations is post-training compression, which aims to balance reduced model size against performance degradation. This work presents Any Compression via Iterative Pruning (ACIP), a novel algorithmic approach to determine a compression-performance trade-off from a single stochastic gradient descent run. To ensure parameter efficiency, we use an SVD-reparametrization of linear layers and iteratively prune their singular values with a sparsity-inducing penalty. The resulting pruning order gives rise to a global parameter ranking that allows us to materialize models of any target size. Importantly, the compressed models exhibit strong predictive downstream performance without the need for costly fine-tuning. We evaluate ACIP on a large selection of open-weight LLMs and tasks, and demonstrate state-of-the-art results compared to existing factorisation-based compression methods. We also show that ACIP seamlessly complements common quantization-based compression techniques.