Low-Rank Factorization

Recently, Muon and related spectral optimizers have demonstrated strong empirical performance as scalable stochastic methods, often outperforming Adam. Yet their behaviour remains poorly understood. We analyze stochastic spectral optimizers, including Muon, on a high-dimensional matrix-valued least squares problem. We derive explicit deterministic dynamics that provide a tractable framework for studying learning behaviour with a focus on (stochastic) SignSVD, which Muon approximates, and (stochastic) SignSGD, the latter serving as a proxy for Adam. Our analysis shows that for large batch size, SignSVD performs a square-root preconditioning with respect to the data covariance spectrum, while for small batch size smaller eigenmodes behave like SGD, slowing down convergence. We contrast with SignSGD which for generic covariance performs no preconditioning and has no transition, leading to different optimal learning rates and convergence characteristics. The two methods match up to a constant factor with isotropic data, but behave differently with anisotropic data. An analysis of a power law covariance model with data exponent $α$ and target exponent $β$ shows there are three phases in the $(α,β)$ plane: one where SignSGD is uniformly favored, one where SignSVD is uniformly favored, and a third where the two methods exhibit a trade-off in performance.

This paper relaxes the restrictive symmetry conditions adopted in [4], [5] and extends their universal feature selection framework to accommodate noisy observations as well as attribute structures that may exhibit directional preferences. We introduce the notion of weak spherical symmetry, quantified by second-moment distances, which allows controlled deviations from rotational invariance. Under this relaxed condition, we develop a universal feature selection framework based on the singular value decomposition of the canonical dependence matrix computed from noisy data. Our main result shows that the selected features achieve asymptotically optimal error exponents up to a residual term that depends on the symmetry deviation $δ$ and the noise levels $η_1, η_2$. When $δ, η_1, η_2$ are relatively small, our result recovers that of [5], thereby demonstrating that exact spherical symmetry is unnecessary. Overall, our findings highlight the robustness of the selection framework against second-moment deviations and observation noise, thereby broadening its applicability across diverse inference tasks and providing a theoretically grounded tool for universal feature selection in practical scenarios.

Muon and related norm-constrained matrix optimizers have become central to large-scale learning problems. They are formulated as a linear maximization oracle (LMO) over an ambient matrix-norm ball in unconstrained Euclidean space. However, these do not generalize cleanly to manifold-valued parameters such as low-rank factorizations, orthogonality constraints, or symmetric positive definite (SPD) matrices. Naively restricting the Muon LMO to the tangent space (i) breaks quotient symmetries and (ii) couples the tangent-space constraint with an ambient norm bound, thereby obstructing closed-form solutions on various manifolds of interest. We resolve both issues with a single observation: every Riemannian metric canonically lifts a unitarily invariant Euclidean norm to an intrinsic norm on each tangent space, and the resulting intrinsic norm constrained LMO is symmetry preserving. Building on this, we introduce intrinsic Muon (iMuon), a unified framework that yields closed-form updates on the fixed-rank, SPD, Stiefel, and Grassmann manifolds for any unitarily invariant norm, including the spectral, Frobenius, and nuclear norms. We establish convergence guarantees for both deterministic and stochastic iMuon with rate constants that depend only on the manifold dimension. Notably, on the fixed-rank manifold this constant depends only on the rank, making the rate independent of factor conditioning and removing the runtime factor-rescaling required by prior work. Experiments on LoRA finetuning of LLMs, image classification, and subspace learning illustrate the efficacy of the proposed approach.

Training large language models requires optimization algorithms that are not only statistically effective, but also computationally and memory efficient at extreme scale. Although Adam remains the dominant optimizer for large-scale language-model pretraining and fine-tuning, recent work has revisited nearly every component of the optimization stack: adaptive moment estimation, decoupled weight decay, memory footprint, curvature approximation, sign-based updates, large-batch stability, low-rank gradient structure, and matrix-wise orthogonalized updates. This survey reviews optimizer design for large language models through a systems-and-optimization lens. We organize the literature into classical first-order optimizers, adaptive optimizers, memory-efficient variants, second-order and curvature-aware methods, sign-based and discovered optimizers, low-rank and projection-based methods, and matrix-based optimizers such as Muon. We also discuss benchmarking methodology, including hyperparameter fairness, scale dependence, wall-clock efficiency, token efficiency, memory overhead, and downstream evaluation. We argue that optimizer research for LLMs is entering a new phase: moving from single-algorithm speedup claims toward rigorous, scale-aware comparisons that jointly evaluate convergence, stability, memory, and implementation complexity.

Neural operators have emerged as powerful data-driven solvers for PDEs, offering substantial acceleration over classical numerical methods. However, existing transformer-based operators still face critical challenges when modeling PDEs on complex geometries: directly processing over massive mesh points is computationally expensive, while operating in raw discretization coordinates may obscure the intrinsic geometry where physical interactions are more naturally expressed. To address these limitations, we introduce the Charted Axial Transformer Operator (CATO), a geometry-adaptive and derivative-aware neural operator for PDEs on general geometries. Instead of applying attention directly in the physical coordinate system, CATO learns a continuous latent chart that maps mesh coordinates into a learned chart space, where chart-conditioned axial attention efficiently captures long-range dependencies with reduced computational cost. In addition, CATO introduces a derivative-aware physics loss for steady-state PDEs that jointly supervises solution values, mesh-consistent gradients, and an auxiliary flux-like field, improving physical fidelity and reducing oversmoothing. We further provide a theoretical approximation result showing that, under a favorable chart, charted axial attention can represent low-rank axial solution operators with controlled error, and that small chart perturbations induce bounded approximation degradation. CATO achieves the best performance across all evaluated datasets, yielding an average improvement of approximately 26.76% over the strongest competing baselines while reducing the number of parameters by 81.98%. These results highlight the effectiveness of learning geometry-adaptive charts and derivative-aware physical supervision for accurate and efficient PDE operator learning.

Low-Rank Adaptation (LoRA) reparameterizes a weight update as a product of two low-rank factors, but the Jacobian $J_{G}$ of the generator mapping the factors to the weight matrix is rank-deficient, so the factor-space preconditioner $J_{G}^* {F}_t J_{G}$ induced by any ${W}$-space preconditioner ${F}_t$ is singular, and consequently the standard chain rule cannot be uniquely inverted to map a preconditioned ${W}$-space direction back to a factor-space update. We cast existing LoRA optimizers in a unified framework parameterized by two choices: (i) which invertible surrogate for $J_{G}^* {F}_t J_{G}$ to use, and (ii) which ${F}_t$ on ${W}$ to use. Existing methods occupy four families along these axes: factor-space adaptive updates, block-diagonal surrogates for $J_{G}^* J_{G}$, Frobenius-residual pseudoinverse methods, and Riemannian manifold constraint. Within this design space, a gradient-statistics-aware ${F}_t$ paired with a closed-form factor-space solve at ${O}((m+n)r)$ memory remains underexplored. We propose \textbf{AdaPreLoRA}, which fills this gap by adopting the Adafactor diagonal Kronecker preconditioner ${H}_t$ on ${W}$ and selecting from the resulting factor-space solution family the element minimizing an ${H}_t$-weighted imbalance between the two factor contributions; by construction, the resulting factor update is the closest LoRA approximation to the preconditioned ${W}$-space direction under the ${H}_t$-weighted norm. Across GPT-2 (E2E), Mistral-7B and Qwen2-7B (GLUE, ARC, GSM8K), and diffusion-model personalization, AdaPreLoRA is competitive with or improves over a representative set of LoRA optimizers while keeping peak GPU memory at the LoRA optimizer level.

Large language models (LLMs) have rapidly grown in scale, creating substantial memory and computational costs that hinder efficient deployment. Singular value decomposition (SVD) has emerged as an effective post-training compression technique, but existing SVD-based methods rely on static rank truncation, applying a fixed prefix of singular components to all inputs regardless of their diversity. We identify two limitations of this static design: the optimal rank varies across individual prompts, and the selected rank is sensitive to the choice of calibration set, leading to suboptimal performance across diverse inputs. To address these challenges, we propose $\textbf{PARSE}$, a post-training framework for $\textbf{P}$rompt-$\textbf{A}$ware $\textbf{R}$ank $\textbf{S}$election as $\textbf{E}$xperts in SVD-compressed LLMs. PARSE trains a linear router offline to perform prompt-aware rank selection, decoupling it from calibration information by supervising the router against dense-model outputs on a large-scale corpus. We further observe that rank-selection patterns are shared across semantically similar prompts and remain stable across decoding steps, allowing appropriate rank subsets to be served directly from a pattern cache at inference. Complemented by expert memory aggregation and kernel fusion for system-level efficiency, PARSE is orthogonal to existing SVD-based pipelines and consistently improves both model quality and inference efficiency. Integrated with four representative SVD-based methods, PARSE improves average task accuracy by up to 10% at a compression ratio of 0.6 on LLaMA-7B, and achieves up to 2.5 $\times$ prefill and 2.4 $\times$ decode speedup over native SVD execution.

We present a data-adaptive method for parameter-efficient fine-tuning of large neural networks. Standard low-rank adaptation methods improve efficiency by restricting each layer update to a fixed low-rank form, but this static parameterization can be too rigid when the appropriate correction depends on the input and on the evolving depth-wise computation of the network. Our approach replaces a purely layer-local adapter with a shared queryable memory of low-rank update atoms. For each block of layers, the model forms a query from the current low-rank state and a running summary of previous blocks, uses this query to retrieve a content-dependent combination of shared update components via attention, and applies the resulting routed operator within the low-rank bottleneck. In this way, the method retains the efficiency and scalability of low-rank adaptation while allowing the effective update to vary across inputs and to share reusable structure across layers. The resulting architecture provides a principled middle ground between static LoRA-style updates and fully generated parameter updates: it remains compact and parameter-efficient while supporting dynamic, context-sensitive adaptation. Further, we incorporate instruction-regularization by augmenting routing logits with a language-induced prior over update atoms, thereby biasing the selection of low-rank transformations toward semantically relevant directions without generating unconstrained parameter updates. Experiments on noisy non-linear regression tasks and LLM fine-tuning suggest that this queryable update-memory formulation can improve final test performance and training stability compared to standard low-rank adaptation, while using a comparable number of trainable parameters.

SVD-based Low-rank compression reduces transformer parameters and nominal FLOPs, but these savings often translate poorly into real LLM serving speedups. We show that this gap is largely a runtime problem: factorized checkpoints fragment execution paths, and the resulting overhead differs substantially between prefill and autoregressive decode. We present FlashSVD v1.5, a unified inference runtime for serving SVD-compressed transformers. FlashSVD v1.5 maps diverse public SVD compression families to a common factorized representation and combines phase-specific kernels with dense-KV decode, packed MLP execution, and per-layer CUDA-graph replay to reorganize the low-rank serving path into a thin runtime. Across representative decoder-serving settings, FlashSVD v1.5 achieves up to 2.55x decode and 2.39x end-to-end speedup, and it attains 1.48x average decode and 1.44x average end-to-end speedup across multiple popular SVD compression families. These results suggest that practical low-rank acceleration requires runtime co-design, not compression algorithms alone. Our code is available at: https://github.com/Zishan-Shao/FlashSVD.

Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of large models by decomposing weight updates into low-rank matrices, significantly reducing storage and computational overhead. While effective, standard LoRA lacks mechanisms for uncertainty quantification, leading to overconfident and poorly calibrated models. Bayesian variants of LoRA address this limitation, but at the cost of a significantly increased number of trainable parameters, partially offsetting the original efficiency gains. Additionally, these models are harder to train and may suffer from unstable convergence. In this work, we propose a novel framework for parameter-efficient Bayesian fine-tuning, demonstrating that effective uncertainty quantification can be achieved in very low-dimensional parameter spaces. The proposed method achieves strong performance with improved calibration and generalization while maintaining computational efficiency. Our empirical findings show that, with the appropriate projection of the weight space uncertainty can be effectively modeled in a low-dimensional space, and weight covariances exhibit low ranks.

Adaptive rank allocation for LoRA, allocating more parameters to important layers and fewer to unimportant ones, consistently improves efficiency under supervised fine-tuning (SFT). We investigate whether this success transfers to reinforcement learning, specifically Group Relative Policy Optimization (GRPO). Using gradient-magnitude profiling on Qwen 2.5 1.5B with GSM8K, we find that it does not: proportional rank allocation degrades accuracy by 4.5 points compared to uniform allocation (70.0% vs. 74.5%), despite using identical parameter budgets. We identify two mechanisms behind this failure. First, the gradient landscape under GRPO is fundamentally flatter than under SFT, the max-to-min layer importance ratio is only 2.17x, compared to >10x reported in SFT literature. All layers carry meaningful gradient signal; none are truly idle. Second, we discover a gradient amplification effect: non-uniform allocation widens the importance spread from 2.17x to 3.00x, creating a positive feedback loop where high-rank layers absorb more gradient while low-rank layers are progressively silenced. Our results suggest that gradient importance does not predict capacity requirements under RL, and that naive transfer of SFT-era rank allocation to alignment training should be avoided.

Random-feature neural networks (RFNNs), including architectures with fixed hidden layers and analytically determined output weights, offer fast training but often suffer from issues due to dense representations of the hidden layer activation. Their reliance on dense feature mappings and least squares solvers can limit scalability and numerical stability, particularly for high-dimensional or stiff systems. Specifically, the activation matrix is observed to be low-rank and extremely ill-conditioned. In this work, we propose a sparse framework for RFNNs that integrates structured sparsity into the hidden layer activations that increases the rank and employs Sparse Singular Value Decomposition (sSVD) for solving the resulting linear least squares problem scalably and efficiently while catering to the bad condition number. We explore the theory behind Lanczos-Golub-Kahan Bidiagonalization technique for sparse SVD and conduct some experiments to identify some limitations and justify the requirement for orthogonalization step in our application. Then, we demonstrate that the proposed method maintains or improves solution accuracy for solving the benchmark one-dimensional steady convection-diffusion equations case having stronger advection, while achieving substantial gains in training efficiency and robustness compared to standard dense implementations.

How few parameters do we really need to forecast a periodic time series? An hourly electricity series, reshaped as a 24-row matrix with one column per day, is approximately rank-1: a daily shape modulated by a daily level (median centered rank-1 energy 0.82 on GIFT-Eval). Should we learn the shape? Smoothing, shrinkage, and low-rank fits all seem like obvious upgrades over the simple average of the last K=2 cycles. On all 97 GIFT-Eval configurations, we tested 8 such alternatives (e.g., Fourier, EWMA, James-Stein, rank-r SVD): none significantly beats the frozen baseline under Holm correction; two are significantly worse. The resulting method, FLAIR, is (a) Effective: matches PatchTST on aggregate GIFT-Eval (relMASE 0.838 vs 0.849); (b) Compact: 28 scalars for hourly, 57 for weekly; (c) Fast: 22 minutes on one CPU core of a MacBook Pro; (d) Closed-form & Hands-Off: one SVD per period candidate, GCV-averaged Ridge, no GPU, no pre-training, no per-task tuning. In the high-rank-1, many-cycle regime, extra flexibility is estimation noise.

Low-rank gradient compression reduces communication in distributed training by representing updates with rank-$r$ factors. Dion is a recent method that approximates Muon, a spectral optimizer that orthogonalizes momentum, using one step of power iteration followed by column normalization (rescaling each column of the right factor to unit length). This makes it compatible with fully sharded data parallel training, but it converges more slowly than full-rank spectral methods. We show that this gap is geometric: column normalization does not yield the rank-$r$ polar factor that Muon implicitly targets, so the resulting direction violates the dual-norm constraint of the low-rank spectral geometry, and the rate picks up an extra factor of $\sqrt{r}$ even though the low-rank approximation of the gradient itself is accurate. The same mismatch enters the smoothness term and the error-feedback recursion in the analysis, which has a knock-on effect on empirical performance. We propose Orth-Dion, which replaces column normalization with QR orthogonalization of the right factor. Under non-Euclidean smoothness, with $L_r$ the curvature constant along rank-$r$ directions, Orth-Dion attains rate $O(\sqrt{L_r/T})$, matching exact spectral methods at the same per-step communication cost as Dion. The proof removes the bounded-drift assumption common in prior error-feedback analyses via a self-consistent fixed-point argument, and uses a time-averaged contraction that only requires the error sequence to contract on average rather than at every step. Experiments on large-scale language model pre-training validate the predicted $\sqrt{r}$ scaling and show that Orth-Dion closes the convergence gap to Muon at Dion's communication cost.

Building on the abstract Generalized Singular Value Decomposition (GSVD) theory of Brown et al. [2025], we prove that most modern neural architectures admit a generalized SVD representation in which they are left-invertible before a final linear layer, with no change in input-output behavior. Furthermore, the left-invertible nonlinear portion of the input-output behavior can be made to be \emph{norm preserving}, meaning that perturbations in the left-invertible ``embedding'' (the activations prior to the final linear layer in this representation) correspond proportionally to changes in the input, i.e., distance in feature space can be calibrated directly to distance in input space. We provide a data-driven algorithm for estimating this representation from trained models and propose a model architecture that naturally facilitates the decomposition. We then provide a proof-of-concept that the learned representation can be used to identify adversarial perturbations to model inputs, and develop the theory necessary for future applications to areas such as model bias and invertibility.

Most first-order optimizers treat matrix-valued parameters as vectors, ignoring the intrinsic geometry of hidden-layer weights in neural networks. Muon addresses this mismatch by updating along the polar factor of a momentum matrix, but its theoretical understanding has lagged behind practice. In particular, practical implementations incorporate Nesterov momentum, compute the polar factor only approximately, and operate with stochastic gradients that may be heavy-tailed. We close this gap by developing a convergence theory for Muon with Nesterov momentum and inexact polar decomposition in non-convex matrix optimization under heavy-tailed noise. Our analysis builds on a unified framework for inexact polar decomposition that captures practical iterative approximations such as Newton-Schulz and quantifies how their errors propagate through the optimization dynamics. Under this framework, we establish an optimal iteration and sample complexity of $O \left(\varepsilon^{\frac{-(3α-2)}{(α-1)}} \right)$ for finding an $\varepsilon$-stationary point, where $α\in(1,2]$ denotes the heavy-tail index. For the inexact-polar setting with $σ_1=0$, we also provide guarantees that do not require prior knowledge of $α$. We analyze a randomized low-rank polar decomposition that is substantially more efficient than full-space methods while remaining compatible with our theory. Numerical experiments further demonstrate the effectiveness of the proposed inexact and randomized variants.

Recent extensive research has demonstrated that the enhanced reasoning capabilities acquired by models through Reinforcement Learning with Verifiable Rewards (RLVR) are primarily concentrated within the rank-1 components. Predicated on this observation, we employed Periodic Rank-1 Substitution and identified a counterintuitive phenomenon: RLVR may exhibit implicit reward overfitting to the training dataset. Specifically, the model can achieve satisfactory performance on the test set even when its rewards remain relatively low during the training process. Furthermore, we characterize three distinct properties of RL training: (1) The effective rank-1 component in RLVR don't maintain other model knowledge except mathematical reasoning capability. (2) RLVR fundamentally functions by optimizing a specific singular spectrum. The distribution of singular values of almost all linear layers in RLVR-trained model behaves like heavy-tailed distribution. (3) the left singular vectors associated with rank-1 components demonstrate a stronger alignment tendency during training, which echoes the discovery that RLVR is optimizing sampling efficiency in essence. Taken together, our findings and analysis further reveal how RLVR shapes model parameters and offer potential insights for improving existing RL paradigms or other training paradigms to implement continual learning.

Federated LoRA enables parameter-efficient adaptation of large language models under decentralized data and limited client resources.However, directly averaging LoRA factors is representation-dependent: the same intrinsic update admits infinitely many gauge-equivalent factorizations, so factor-level aggregation can change under arbitrary coordinate choices while the underlying update remains unchanged. This reveals a semantic mismatch in existing federated LoRA aggregation rules. We propose \textbf{GLoRA}, a gauge-aware server representation for federated LoRA.Instead of aggregating raw factors, GLoRA estimates a consensus update subspace from client projectors and aggregates client updates in shared reference coordinates, thereby representing semantic update aggregation entirely in low-rank form. To support heterogeneous client capacities, GLoRA further provides a rank-compatible readout that instantiates adapters of different ranks from the same server state without dense update reconstruction. Experiments on GLUE and SuperNI show that GLoRA consistently outperforms federated LoRA baselines under data, resource, and task heterogeneity, including heterogeneous client ranks, sparse participation, larger backbones, and unseen-task evaluation. GLoRA also achieves a favorable efficiency--performance trade-off, suggesting that effective federated LoRA requires not merely averaging low-rank factors, but defining a semantically meaningful server-side representation for aggregation.

The massive computational costs of scaling modern deep learning architectures have driven the widespread use of parameter-efficient low-rank structures, such as LoRA and low-rank factorization. However, theoretical guarantees for their expressive power are less explored, often relying on restrictive priors like a pretrained base matrix, ReLU activations or non-verifiable singularity conditions. We first investigate the limits of neural networks constrained strictly to low-rank manifolds without pretrained dense priors. We demonstrate a theoretical paradox: while purely rank-1 layers can exactly interpolate arbitrary scalar datasets, they collapse for function approximations. To overcome this bottleneck without surrendering parameter efficiency, we introduce a unified \textit{Structural Correspondence} framework. We prove that augmenting low-rank layers with only a minimal sparse diagonal component, say a Diagonal plus Low-Rank (DLoR) structure, is sufficient to reach Universal Approximation. We show that any full-rank transformation can be exactly reconstructed using these DLoR components by trading off network width (additive decomposition) or depth (multiplicative decomposition). By tracking asymptotic Taylor remainders, we prove that DLoR neural networks fully restore the Universal Approximation Theorem for general activation functions. Finally, we establish that multiplicative depth provides superior parameter-to-expressivity scaling compared to additive width. Our results show that dense matrices and specific activation functions are not topological prerequisites for universal expressivity.

Nonconvex methods have emerged as a dominant approach for low-rank matrix estimation, a problem that arises widely in machine learning and AI for learning and representing high-dimensional data. Existing analyses for these methods often require additional regularization to mitigate nonconvexity, even though such regularization is often unnecessary in practice. Moreover, most analyses rely on problem-specific arguments that are difficult to generalize to more complex settings. In this paper, we develop a theoretical framework for studying nonconvex procedures across a broad class of low-rank matrix estimation problems. Rather than focusing on a specific model, we reveal a fundamental mechanism that explains why nonconvex procedures can behave well in low-rank estimation. Our key device is a {\it benign regularizer} that does not alter the original update rule, but yields an equivalent locally strongly convex formulation of the algorithm. This perspective uncovers a disguised convexity inherent in the nonconvex procedure and provides a new route to theoretical guarantees for nonconvex low-rank matrix estimation.

Perceptual image compression focuses on preserving high visual quality under low-bitrate constraints. Most existing approaches to perceptual compression leverage the strong generative capabilities of generative adversarial networks or diffusion models, at the cost of substantial model complexity. To this end, we present an efficient perceptual image compression method that exploits the long-range modeling capability and linear computational complexity of state space models, with a particular focus on Mamba. Unlike existing methods that rely on an inherently fixed scanning order and consequently impair semantic continuity and spatial correlation, we develop a semantic-aware Mamba block (SAMB) to enable scanning guided by dynamically clustered semantic features, thereby alleviating the strict causality constraints and long-range information decay inherent to Mamba. Inspired by singular value decomposition, we design an SVD-inspired redundancy reduction module (SVD-RRM) that performs a low-rank approximation on the latent features by introducing a learnable soft threshold, leading to channel-wise redundancy information reduction. The proposed SAMB is integrated into both the encoder and decoder of the compression framework, whereas the SVD-RRM is incorporated only in the encoder. Extensive experiments demonstrate that our method performs favorably against state-of-the-art approaches in terms of rate-distortion-perception tradeoff and model complexity. The source code and pretrained models will be available at https://github.com/Jasmine-aiq/SAMIC.

Generative Recommender (GR) inference places embedding hot caches (EMB) and KV caches in direct competition for limited GPU HBM: allocating more memory to one improves its efficiency but degrades the other. Existing systems optimize them in isolation, overlooking that the optimal EMB-KV allocation ratio can shift by up to 0.35 across workload regimes, leaving 20-30% latency improvement unrealized. While online reallocation is required to close this gap, naive approaches introduce H2D refill traffic on the critical path, causing P99 SLO violations. To address this, we present HELM, which jointly manages HBM allocation and request routing at runtime through two key components: (1) Adaptive Memory Allocation, a three-layer PPO-based controller (frozen base policy, online residual adapter, and burst-aware recovery controller) that achieves $32\,\mathrm{μs}$ decision latency while staying within 0.024-0.029 of the offline-optimal ratio; and (2) EMB-KV-Aware Scheduling, which routes requests by jointly considering KV residency, embedding locality, and node load to avoid routing inefficiencies under heterogeneous allocations. Evaluations on three production-scale datasets over a 32-node A100 cluster show that HELM reduces P99 latency by 24-38% over the best static policy and achieves 93.5-99.6% SLO satisfaction across Steady, Trend, and Burst workloads, significantly outperforming state-of-the-art baselines without sacrificing throughput.

Large Language Models (LLMs) have achieved remarkable capabilities, but their immense computational demands during training remain a critical bottleneck for widespread adoption. Low-rank training has received attention in recent years due to its ability to significantly reduce training memory usage. Meanwhile, applying 2:4 structured sparsity to weights and activations to leverage NVIDIA GPU support for 2:4 structured sparse format has become a promising direction. However, existing low-rank methods often leave activation matrices in full-rank, which dominates memory consumption and limits throughput during large-batch training. Furthermore, directly applying sparsity to weights often leads to non-negligible performance degradation. To achieve efficient pre-training of LLMs, this paper proposes ELAS: Efficient pre-training of Low-rank LLMs via 2:4 Activation Sparsity, a novel framework for low-rank models via 2:4 activation sparsity. ELAS applies squared ReLU activation functions to the feed-forward networks in low-rank models and implements 2:4 structured sparsity on the activations after the squared ReLU operation. We evaluated ELAS through pre-training experiments on LLaMA models ranging from 60M to 1B parameters. The results demonstrate that ELAS maintains performance with minimal degradation after applying 2:4 activation sparsity, while achieving training and inference acceleration. Moreover, ELAS reduces activation memory overhead, particularly with large batch sizes. Code is available at ELAS Repo.

KV-cache quantizers usually optimize storage-space reconstruction, even though attention reads keys through logits and values through attention-weighted readout. We argue that persistent cache error should be measured in model-visible coordinates. For keys, the visible object is score error modulo constant shifts; this yields HeadQ, a key-side method that stores a low-rank residual side code in a calibration-learned query basis and applies it as an additive logit correction. For values, fixed-attention readout gives an $A^2$-weighted token-distortion surrogate. Across six models, Fisher/score-space error predicts attention KL far better than raw key MSE; same-budget counterexamples, null-space interventions, query-PCA controls, and wrong-sign HeadQ falsify storage-MSE alternatives. Matched Pythia checkpoints localize the main anomaly to a small-model low-entropy route-flip boundary. In K-only WikiText-103 decode experiments with dense values, HeadQ removes roughly $84$--$94\%$ of the excess perplexity on the strongest 2-bit rows; in an auxiliary full-KV 2-bit composition, HeadQ plus an $A^2$ value policy improves all six models.

We introduce the Yat kernel $k_{b,\varepsilon}(\mathbf{w},\mathbf{x})=\frac{(\mathbf{w}^\top\mathbf{x}+b)^2}{\|\mathbf{x}-\mathbf{w}\|^2+\varepsilon},\qquad b\ge 0,\ \varepsilon>0,$ a rational hidden-unit primitive whose units are Mercer sections over a shared input/weight space. For $b\ge 0$ the kernel is PSD; for $b>0$ it dominates a scaled inverse-multiquadric (IMQ) in the Loewner order, yielding fixed-kernel universality, characteristicness, and strict positive definiteness on every compact domain. The polynomial numerator opens nonradial alignment channels absent from finite IMQ expansions, witnessed by the directional far-field trace $T_\infty g_\varepsilon(\cdot;\mathbf{w},b)(\mathbf{u})=(\mathbf{u}^\top\mathbf{w})^2$. Algebraically, a second finite difference in the bias recovers any IMQ atom from three positive-bias Yat atoms exactly, sharp at three atoms in every dimension at exact pointwise equality. A trained shared-$(b,\varepsilon)$ Yat layer is therefore a finite learned-center expansion in a fixed universal characteristic RKHS, with closed-form norm $\boldsymbolα^\top\mathbf{K}\boldsymbolα$ and explicit diagonal $(\|\mathbf{x}\|^2+b)^2/\varepsilon$ driving a Rademacher generalization bound.

A method is presented for accelerating inference in transformer language models by exploiting the low effective rank of the token activation manifold at each layer. The method decomposes each activation vector into a subspace component and a residual, computes the linear-layer output on the subspace component via a cached low-rank weight image at reduced memory bandwidth, and applies a per-token gate that determines whether the residual correction is computed or skipped. The gate ensures that the output distribution is preserved to within a controllable tolerance. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates effective speedups of 3.0x to 10.5x on linear-layer weight reads with perplexity ratios below 1.00 and top-1 token agreement above 98%. The method requires no retraining, no architectural modification, and no approximation of the attention mechanism. At the operating point (k = 256, ε = 0.05) on GPT-J 14 6B, the accelerated model produces character-for-character identical output to the baseline.

Adapting large pretrained models to diverse tasks is now routine, yet the two dominant strategies of parameter-efficient fine-tuning (PEFT) and low-rank compression are typically composed in sequence. This decoupled practice first compresses and then fine-tunes adapters, potentially misaligning the compressed subspace with downstream objectives and squandering a global parameter budget. To overcome this limitation, we introduce JACTUS (Joint Adaptation and Compression with a Task-aware Union of Subspaces), a single framework that unifies compression and adaptation. From a small calibration set, JACTUS estimates input and pre-activation gradient covariances, forms their orthogonal union with the pretrained weight subspace, performs a projected low-rank approximation inside this union, allocates rank globally by marginal gain per parameter, and trains only a compact core matrix. This explicitly mitigates the potential misalignment between the compressed subspace and downstream objectives by coupling the directions preserved for compression with those required for adaptation, yielding a deployable low-rank model that avoids retaining full frozen weights while enabling fast and robust tuning. On vision, JACTUS attains an average 89.2% accuracy on ViT-Base across eight datasets at 80% retained parameters, surpassing strong 100% PEFT baselines (e.g., DoRA 87.9%). On language, JACTUS achieves an 80.9% average on Llama2-7B commonsense QA at the same 80% retained-parameter budget, outperforming 100% PEFT (e.g., DoRA 79.7%) and exceeding prior compress-then-finetune pipelines under the same ratained-parameter budget. We will release code.

We introduce the Z-Domain Neural Operator (ZNO), a causal neural operator whose layers are stable low-rank multiple-input multiple-output (MIMO) rational filters parameterized directly in the $z$-plane. ZNO addresses a limitation of existing operator learning methods, many of which are primarily tailored for continuous-time problems, while a large class of system-identification problems is intrinsically discrete-time. The $z$-domain form expresses stability as a unit-disk pole constraint and makes learned discrete-time poles directly readable. The model combines low-rank channel mixing, smooth stable pole reparameterization, causal recurrence, and an optional short finite impulse response (FIR) branch in a single $z$-domain rational recurrent layer. Across controlled discrete system-identification experiments, ZNO's advantage is most evident when the target dynamics are stable rational systems with lightly damped poles near the unit circle. Under matched parameter budgets, ZNO is not uniformly dominant; however, with validation-selected configurations, the same architecture can achieve the lowest mean error across the controlled tasks. A five-bin difficulty sweep over near-unit-circle / long-memory dynamics shows that ZNO has the lowest mean error across memory regimes, from short (approximately 10 steps) to long (approximately 100-200 steps). On five public nonlinear system-identification benchmarks, ZNO is competitive with neural operator and state-space baselines, achieving the lowest mean error on benchmarks whose dynamics align with stable rational discrete-time filters, while classical or state-space baselines remain preferable on some systems. These results position ZNO as a strong model for stable rational discrete-time dynamics, especially in near-unit-circle and long-memory regimes, but not as a universal replacement for specialized system-identification methods.

Low-rank decomposition is a compelling approach for compressing large language models, but its effectiveness hinges on selecting which singular-vector bases to retain for a target task. Existing methods such as Basel adapt singular-value coefficients on downstream data and prune bases with small re-learned magnitudes, a heuristic that can be misaligned with task performance because it ignores the local geometry of the loss landscape. We present Basis Selection with Importance (BSI), a principled low-rank compression framework that ranks and prunes bases by directly estimating the expected loss increase incurred when each basis is removed. BSI derives a derivative-based importance score from a second-order Taylor expansion of the task loss with respect to singular values, combining first-order sensitivity and second-order curvature to quantify pruning impact. To make this criterion practical for LLMs, we develop an efficient Hessian-diagonal estimator by adapting the Hutchinson randomized-probing method to loss curvature with symmetric parameter perturbations. We provide a comprehensive theoretical analysis, including loss-increase bounds under basis pruning, explicit propagation of Hessian-diagonal estimation error into these bounds, variance characterization tied to the Hessian spectrum, high-probability sample-complexity guarantees for achieving a target estimation accuracy, and guidance on perturbation intensity. Extensive experiments on mathematical reasoning benchmarks demonstrate that BSI consistently outperforms state-of-the-art low-rank decomposition baselines, with especially strong improvements under deep compression.

Modern deep learning usually treats models as separate artifacts: trained independently, specialized for particular purposes, and replaced when improved versions appear. This thesis studies model merging as an alternative paradigm: combining independently trained neural networks directly in weight space, with little or no optimization and without requiring access to the original training data. The thesis considers two main regimes. In the single-task setting, where models share an objective but differ in initialization, we introduce C$^2$M$^3$, a cycle-consistent merging algorithm based on Frank-Wolfe optimization. C$^2$M$^3$ aligns multiple networks into a shared, reference-free parameter space, making weight averaging meaningful without privileging any individual model. In the multi-task setting, where models are fine-tuned for different downstream tasks from a common pretrained initialization, we first develop a theoretical account of task vectors as approximate gradients. This explains both the effectiveness and the limitations of task arithmetic. Building on this view, we show that task vectors inherit the low-rank structure of gradients and introduce Task Singular Vectors (TSV), a decomposition that enables compression and interference reduction through TSV-Merge. We then present MASS, an input-adaptive routing method that uses TSV geometry to select task-relevant subspaces at inference time. Finally, we introduce MERGE$^3$, an evolutionary merging framework that uses Item Response Theory to reduce evaluation costs by up to 50$\times$ while preserving solution quality. Together, these contributions provide theoretical and algorithmic foundations for model merging, supporting a paradigm in which learned capabilities can be composed, reused, and extended across models.

Exponential growth in the scale of modern foundation models has led to the widespread adoption of Low-Rank Adaptation (LoRA) as a parameter-efficient fine-tuning technique. However, standard LoRA implementations disregard the varying intrinsic dimensionality of model layers and enforce a uniform rank, leading to parameter redundancy. We propose Post-Optimization Adaptive Rank Allocation (PARA), a data-free compression method for LoRA that integrates seamlessly into existing fine-tuning pipelines. PARA leverages Singular Value Decomposition to prune LoRA ranks using a global threshold over singular values across all layers. This results in non-uniform rank allocation based on layer-wise spectral importance. As a post-hoc method, PARA circumvents the training modifications and resulting instabilities that dynamic architectures typically incur. We empirically demonstrate that PARA reduces parameter count by 75-90% while preserving the predictive performance of the original, uncompressed LoRA across multiple vision and language benchmarks. Code will be published upon acceptance.

Low-Rank Adaptation (LoRA), which leverages the insight that model updates typically reside in a low-dimensional space, has significantly improved the training efficiency of Large Language Models (LLMs) by updating neural network layers using low-rank matrices. Since the generation of adversarial examples is an optimization process analogous to model training, this naturally raises the question: Do adversarial perturbations exhibit a similar low-rank structure? In this paper, we provide both theoretical analysis and extensive empirical investigation across various attack methods, model architectures, and datasets to show that adversarial perturbations indeed possess an inherently low-rank structure. This insight opens up new opportunities for improving both adversarial attacks and defenses. We mainly focus on leveraging this low-rank property to improve the efficiency and effectiveness of black-box adversarial attacks, which often suffer from excessive query requirements. Our method follows a two-step approach. First, we use a reference model and auxiliary data to guide the projection of gradients into a low-dimensional subspace. Next, we confine the perturbation search in black-box attacks to this low-rank subspace, significantly improving the efficiency and effectiveness of the adversarial attacks. We evaluated our approach across a range of attack methods, benchmark models, datasets, and threat models. The results demonstrate substantial and consistent improvements in the performance of our low-rank adversarial attacks compared to conventional methods.

EigenDecomposition (ED) is at the heart of many computer vision algorithms and applications. One crucial bottleneck limiting its usage is the expensive computation cost, particularly for a mini-batch of matrices in deep neural networks. Our previous work proposed a dedicated QR-based ED algorithm for batched small matrices (dim${<}32$). This short paper targets the limitation and proposes a batch-efficient Divide-and-Conquer based ED algorithm for larger matrices. The numerical test shows that for a mini-batch of matrices whose dimensions are smaller than $64$, our method can be much faster than the Pytorch SVD function.

We show that replacing the rolling SVD of AdamW updates with a rolling SVD of loss gradients changes the diagnostic by 1-2 orders of magnitude. Performing SVD on the loss gradient instead of the AdamW update increases the measured perturbative coupling between SED directions and Linear Centroid Hypothesis (LCH) features from $\bar{R}_k \approx 3$--$9\times$ to $100$--$330\times$ across four single-task modular arithmetic operations, eliminating the apparent operation dependence in the original measurement. On a multitask transformer with a shared encoder, update-based SED gives $\bar{R}_k \leq 1$ -- an apparent failure of the diagnostic -- while per-operation gradient-based SED recovers $\bar{R}_k = 20$--$45\times$ across all four operations. Gradient aggregation across competing tasks is the main obstruction; performing SVD on per-task gradients resolves it. A causal intervention shows that constraining attention updates to any rank-3 subspace (whether SED-derived or random) accelerates grokking by approximately $2.3\times$ across random seeds and operations, while removing the rank-3 component has negligible effect under proper gradient-projection methodology. The SED-LCH coupling is therefore a strong diagnostic of where feature formation concentrates in parameter space, but it is not a unique causal pathway: the natural full-rank AdamW attention update is highly rank-redundant under our hyperparameters.

Low-Rank Adaptation (LoRA) has become the standard for fine-tuning large pre-trained models at reduced computational cost. However, its low-rank point-estimate updates limit expressiveness, leave a persistent gap relative to full fine-tuning accuracy, and provide no built-in uncertainty quantification, limiting its applicability in settings where reliability matters as much as accuracy. We introduce BaLoRA, a Bayesian extension of LoRA with a novel input-adaptive Bayesian parameterization of LoRA matrices that adds minimal parameters and compute. Surprisingly, not only does the Bayesian extension yield well-calibrated uncertainty estimates, but the adaptive noise injection underlying our approach also significantly improves prediction accuracy, narrowing the gap with full fine-tuning across both natural language reasoning and vision tasks. When applied to band gap prediction in metal-organic frameworks, BaLoRA produces zero-shot test-time uncertainty estimates that correlate more strongly with model error than a trained ensemble of LoRA models, and improve monotonically with compute without sacrificing accuracy.

Federated learning (FL) offers a promising distributed learning paradigm for internet of vehicles (IoV) applications. However, it faces challenges from communication overhead and dynamic environments. Model compression techniques reduce computing and communication burden yet create trade-offs between compression ratios and vehicle participation strategies. In this paper, we propose a lightweight FL algorithm named federated learning with dynamic low-rank adaptation (Fed-DLoRA), which is combined with low-rank adaptation (LoRA) to effectively reduce parameters and communication costs while enhancing training efficiency. The convergence analysis of Fed-DLoRA is conducted through stochastic gradient descent optimization coupled with singular value decomposition. This analysis establishes the theoretical relationships among LoRA rank, vehicular scheduling strategies and the model's convergence characteristics. Building on these insights, we formulate a joint optimization problem aimed at maximizing system performance. To address this problem, we propose an adaptive rank, bandwidth and vehicle selection (ARBVS) algorithm that integrates enumeration with greedy optimization strategies. The algorithm provides efficient rank selection and resource scheduling strategies for each FL communication round, thereby achieving effective performance improvements for the FL system. Experimental results demonstrate that Fed-DLoRA achieves superior performance compared to conventional federated learning approaches, exhibiting enhanced accuracy, faster convergence, and improved communication efficiency.

Convolutional neural networks (CNNs) have become increasingly difficult to deploy in resource-constrained environments due to their large memory and computational requirements. Although low-rank compression methods can reduce this burden, most existing approaches compress spatial and channel redundancy independently and therefore do not fully exploit the localised structure within convolutional feature maps. This paper proposes a hierarchical spatio-channel low-rank compression framework for CNNs that exploits redundancy across spatial regions and channel activations. Unlike conventional methods, which apply a uniform decomposition across an entire layer, the proposed approach first partitions feature maps into spatial regions, then groups channels according to their co-activation patterns within each region, and finally applies rank-adaptive SVD to each resulting spatio-channel cluster. The method is evaluated on an AlexNet-based brain tumour MRI classification model and compared with Global SVD and Tucker decomposition under $3\times$ and $6\times$ compression budgets. Our method outperforms both baselines, reducing FLOPs from $8.21\,\mathrm{G}$ to $1.55\,\mathrm{G}$ ($81.1\%$ reduction), achieving a $1.38\times$ inference speed-up, and increasing classification accuracy from $87.76\%$ to $89.80\%$. The method also improves the macro $F_1$-score and performance on challenging classes such as meningioma. A hyper-parameter trade-off analysis demonstrates that the framework provides Pareto-optimal configurations, enabling control over the balance between compression and predictive performance. Moderate clustering with adaptive rank selection yields strong results. Bootstrap standard errors are reported for all classification metrics.

One of the major limits of kernel ridge regression (KRR) is that storing and manipulating the kernel matrix K_n for n samples requires O(n^2) space, which rapidly becomes unfeasible for large n. Nystrom approximations reduce the space complexity to O(nm) by sampling m columns from K_n. Uniform sampling preserves KRR accuracy (up to epsilon) only when m is proportional to the maximum degree of freedom of K_n, which may require O(n) columns for datasets with high coherence. Sampling columns according to their ridge leverage scores (RLS) gives accurate Nystrom approximations with m proportional to the effective dimension, but computing exact RLS also requires O(n^2) space. (Calandriello et al. 2016) propose INK-Estimate, an algorithm that processes the dataset incrementally and updates RLS, effective dimension, and Nystrom approximations on-the-fly. Its space complexity scales with the effective dimension but introduces a dependency on the largest eigenvalue of K_n, which in the worst case is O(n). In this paper we introduce SQUEAK, a new algorithm that builds on INK-Estimate but uses unnormalized RLS. As a consequence, the algorithm is simpler, does not need to estimate the effective dimension for normalization, and achieves a space complexity that is only a constant factor worse than exact RLS sampling.

Mid-infrared astronomy from the ground faces critical challenges in accurately detecting and quantifying sources due to the dominant spatially and time-variable background noise. Moreover, chopping and nodding, the traditional methods for dealing with these background issues, will not be technically feasible on the next generation of extremely large telescopes. This limitation requires the development of novel computational methods for a robust background reduction. We present and evaluate a novel method named LOw-RAnk Background ELimination (LORABEL) to improve the sensitivity of mid-infrared astronomical observations, without the need for classical telescope nodding, source masking, or other overheads in observing time. We applied a low-rank background-reduction strategy to (1) data taken on the ground with the VISIR with synthetically injected sources, and (2) airborne data from SOFIA. We compared the performance of our new method to classical chopping and nodding techniques, and analysed the effect on source photometry and detection precision for different observational scenarios. In regimes with a low signal-to-noise ratio (S/N $<5$) in the ground-based VISIR data, LORABEL reduces variation in the photometric error with respect to chopping differences alone and even the classical chop-nod sequence, at the cost of introducing a bias. Secondly, we demonstrate that LORABEL increases detection precision in comparison to traditional background-reduction methods. For the SOFIA dataset, we achieve a $20-100$ fold decrease in mean background flux with respect to the traditional chop-nod method while preserving most of the source flux. Our findings suggest that LORABEL is applicable to a wider range of instrumental observation, that is, both ground-based and airborne, and it is a suitable tool in the context of faint-source detection.

Large-scale kernel ridge regression (KRR) is limited by the need to store a large kernel matrix K_t. To avoid storing the entire matrix K_t, Nystrom methods subsample a subset of columns of the kernel matrix, and efficiently find an approximate KRR solution on the reconstructed matrix. The chosen subsampling distribution in turn affects the statistical and computational tradeoffs. For KRR problems, recent works show that a sampling distribution proportional to the ridge leverage scores (RLSs) provides strong reconstruction guarantees for the approximation. While exact RLSs are as difficult to compute as a KRR solution, we may be able to approximate them well enough. In this paper, we study KRR problems in a sequential setting and introduce the INK-ESTIMATE algorithm, that incrementally computes the RLSs estimates. INK-ESTIMATE maintains a small sketch of K_t, that at each step is used to compute an intermediate estimate of the RLSs. First, our sketch update does not require access to previously seen columns, and therefore a single pass over the kernel matrix is sufficient. Second, the algorithm requires a fixed, small space budget to run dependent only on the effective dimension of the kernel matrix. Finally, our sketch provides strong approximation guarantees on the distance between the true kernel matrix and its approximation, and on the statistical risk of the approximate KRR solution at any time, because all our guarantees hold at any intermediate step.

Estimating the number of components is a fundamental challenge in unsupervised learning, particularly when dealing with high-dimensional data with many components or severely imbalanced component sizes. This paper addresses this challenge for classical Gaussian mixture models. The proposed estimator is simple: center the data, compute the singular values of the centered matrix, and count those above a threshold. No iterative fitting, no likelihood calculation, and no prior knowledge of the number of components are required. We prove that, under a mild separation condition on the component centers, the estimator consistently recovers the true number of components. The result holds in high-dimensional settings where the dimension can be much larger than the sample size. It also holds when the number of components grows to the smaller of the dimension and the sample size, even under severe imbalance among component sizes. Computationally, the method is extremely fast: for example, it processes ten million samples in one hundred dimensions within one minute. Extensive experimental studies confirm its accuracy in challenging settings such as high dimensionality, many components, and severe class imbalance.

Recovering latent structure from count data has received considerable attention in network inference, particularly when one seeks both cross-group interactions and within-group similarity patterns in bipartite networks, which is widely used in ecology research. Such networks are often sparse and inherently imperfect in their detection. Existing models mainly focus on interaction recovery, while the induced similarity graphs are much less studied. Moreover, sparsity is often not controlled, and scale is unbalanced, leading to oversparse or poorly rescaled estimates with degrading structural recovery. To address these issues, we propose a framework for structured sparse nonnegative low-rank factorization with detection probability estimation. We impose nonconvex $\ell_{1/2}$ regularization on the latent similarity and connectivity structures to promote sparsity within-group similarity and cross-group connectivity with better relative scale. The resulting optimization problem is nonconvex and nonsmooth. To solve it, we develop an ADMM-based algorithm with adaptive penalization and scale-aware initialization and establish its asymptotic feasibility and KKT stationarity of cluster points under mild regularity conditions. Experiments on synthetic and real-world ecological datasets demonstrate improved recovery of latent factors and similarity/connectivity structure relative to existing baselines.

Post-training quantization (PTQ) is essential for deploying large diffusion transformers on resource-constrained hardware, but aggressive 4-bit quantization significantly degrades generative performance. Low-rank approximation methods have emerged as a promising solution by appending auxiliary linear branches to restore performance. However, current state-of-the-art approaches assume these branches must retain high precision (W16A16) and rely on heavy, data-dependent calibration for initialization. We challenge both limitations with LoRaQ (Low-Rank Approximated Quantization), a simple, data-free calibration approach that optimizes quantization error compensation. By overcoming the need for high-precision branches, LoRaQ enables the first fully sub-16 bit pipeline, allowing the low-rank branch itself to be quantized. We demonstrate that, at equal memory overhead, LoRaQ outperforms the state-of-the-art methods in their native implementations on Pixart-$Σ$ and SANA. We also analyze mixed-precision configurations, showing that setups such as W8A8, W6A6, and W4A8 for the low-rank branch, alongside a W4 main layer, yield superior results while maintaining a fully quantized architecture compatible with modern mixed-precision hardware.

Matrix-level low-rank compression is a promising way to reduce the cost of large language models, but running compression and evaluating the resulting models on language tasks can be prohibitively expensive. Can compression-induced degradation be predicted before committing to this compute? We systematically analyze the Qwen3 and Gemma3 model families across four representative low-rank compression methods: vanilla SVD, two ASVD variants, and SVD-LLM. We find that stable rank and information density, measured in bits per parameter, dominate performance degradation. The interaction term $γ\cdot \barρ_s$, defined as compression ratio times stable rank, is a robust predictor of accuracy degradation, achieving leave-one-out cross-validation Pearson correlations of $0.890$ for attention layers and $0.839$ for MLP layers. We provide theoretical intuition for why this predictor succeeds by connecting it to standard SVD truncation bounds and error composition mechanisms in transformer layers. These findings enable a predict-then-compress workflow: compute $γ\cdot \barρ_s$ from weights, estimate degradation, and invest compute only in desirable configurations.

Adapting foundation models under resource budgets relies heavily on Parameter-Efficient Fine-Tuning (PEFT), with LoRA being a standard modular solution. However, LoRA suffers from spectral interference. Low-rank updates often concentrate energy on the leading singular directions of pretrained weights, perturbing general capabilities and causing catastrophic forgetting and fragile multi-adapter merging. To resolve this, we propose HiP-LoRA, a spectrum-aware adaptation framework. Utilizing the cached singular value decomposition (SVD) of pretrained layers, HiP-LoRA decomposes updates into two channels: a principal channel within the dominant singular subspace, and a residual low-rank channel in the orthogonal complement. A singular-value-weighted stability budget on the principal channel continuously balances pretrained behavior preservation with task-specific plasticity. Experiments on Llama-3.1-8B demonstrate that under matched budgets, HiP-LoRA drastically reduces pretraining degradation and multi-adapter MergeFail, robustly outperforming baselines in interference-sensitive tasks like continual tuning and knowledge editing.

Recursive architectures such as Tiny Recursive Models (TRMs) perform implicit reasoning through iterative latent computation, yet the geometric structure of these reasoning trajectories remains poorly understood. We investigate the activation manifold of TRMs during recursive unrolling and find that activations occupy an effectively linear, low-dimensional subspace whose principal directions can be tracked dynamically with cheap power iterations. This suggests that weight-sharing concentrates iterative computation along a small number of dominant eigendirections, and we find that this concentration varies sharply across computational sites. We exploit this structure through LASER (Low-Rank Activation SVD for Efficient Recursion), a dynamic compression framework that maintains an evolving low-rank basis via matrix-free subspace tracking with a fidelity-triggered reset mechanism, achieving ${\sim}60\%$ activation memory savings with no statistically significant accuracy degradation. Our analysis raises questions about how recursive architectures allocate representational capacity during implicit reasoning, and whether this concentration can be exploited to improve the efficiency and stability of latent computation.

This work introduces Structured 3D-SVD as a practical framework for the reconstruction, compression, and analysis of biological volumetric data. Inspired by the logic of matrix singular value decomposition (SVD), the proposed approach represents third-order volumetric data in the spatial domain and supports progressive reconstruction through ordered quasi-singular coeffients. The experimental evaluation was carried out on two biological volumetric datasets: one full-volume scan of a fish and another of a brain. The results show that Structured 3D-SVD achieves reconstruction quality close to that of Tucker decomposition while requiring shorter computation times and outperforms canonical polyadic decomposition (CPD) in both accuracy and runtime. In addition, a progressive reconstruction analysis shows that relatively low truncation levels are sufficient to preserve the main volumetric structures, while higher truncation levels lead to more detailed reconstructions.

Supervised Fine-Tuning (SFT) accelerates taskspecific large language models (LLMs) development, but the resulting proliferation of finetuned models incurs substantial memory overhead. Delta compression addresses this by retaining a single pre-trained LLM with multiple compressed delta weights. However, existing methods fail on models fine-tuned with largescale datasets. We find that larger SFT data scale amplifies delta parameter magnitude, singular values, and entropy, exacerbating compression errors. To tackle this, we propose DQRELO (Delta Compression via Quantization and Residual Low-Rank), a novel training- and data-free delta compression method. It combines coarse-grained one-bit quantization to capture the dominant structure of the delta, followed by compensated residual low-rank approximation to recover fine-grained details from the smaller residual error. Experiments on various LLMs spanning dense and MoE architectures across multiple domains under this challenging setting demonstrate that DQRELO outperforms existing methods. Moreover, we establish key design principles for delta compression through extensive empirical analysis, demonstrating how task difficulty, architecture, and layer positioning create predictable patterns that can guide optimal compression strategies in production systems.

Large pre-trained language models are increasingly adapted to downstream tasks using parameter-efficient fine-tuning (PEFT), but existing PEFT methods are typically deterministic and unimodal, making them poorly suited for low-resource multimodal settings where predictive uncertainty and cross-modal reliability both matter. We introduce CALIBER (Context-Aware Low-rank Inference with Bayesian Embedding Regularization), a multimodal uncertainty-aware PEFT framework for audio-text learning. CALIBER extends Bayesian low-rank adaptation by conditioning the variational posterior in the adapter space on per-layer, token-level text-audio cross-attention. Specifically, text-derived low-rank features attend to frame-level audio embeddings to produce localized acoustic context, which then modulates the mean and variance of a compact stochastic latent matrix within the rank-$r$ adapter space. This design treats audio not only as an additional feature source, but as a contextual reliability signal that shapes both adaptation and confidence. By confining stochasticity to a low-dimensional latent component, CALIBER retains the computational efficiency and scalability of PEFT while enabling heteroscedastic multimodal uncertainty estimation. Experimental results across diverse text and audio backbones show that CALIBER consistently matches or improves upon text-only Bayesian PEFT and conventional multimodal transfer-learning baselines, with token-level cross-attention yielding the most consistent gains. Our findings demonstrate that localized cross-modal conditioning is an effective and lightweight mechanism for uncertainty-aware multimodal adaptation.

Vision-Language Models (VLMs) have revolutionized multi-modal learning by jointly processing visual and textual information. Yet, they face significant challenges due to the high computational and memory demands of processing long sequences of vision tokens. Many existing methods rely on local heuristics, such as attention scores or token norms. However, these criteria suffer from positional bias and information dispersion, limiting their ability to preserve essential content at high pruning ratios and leading to performance degradation on visually detailed images. To address these issues, we propose SVD-Prune, a training-free, plug-and-play token pruning method based on Singular Value Decomposition. It decomposes the vision token feature matrix and selects the top-k tokens using statistical leverage scores, ensuring only tokens contributing most to the dominant global variance are preserved. Experiments show that SVD-Prune consistently outperforms prior pruning methods under extreme vision token budgets, maintaining strong performance even with 32 and 16 vision tokens.

Large language model (LLM) evaluation platforms increasingly rely on pairwise human judgments. These data are noisy, sparse, and non-uniform, yet leaderboards are reported with limited uncertainty quantification. We study this as semiparametric inference for a low-rank latent score tensor observed through pairwise comparisons under Bradley-Terry-Luce-type models. This places LLM evaluation in a new tensor completion setting with structured observations, non-uniform sampling, and pairwise contrasts. Our target is a smooth functional $ψ(T^\star)$, including linear estimands such as ability gaps and nonlinear ones such as win probabilities. We derive the information operator on the low-rank tangent space, the efficient influence function, and the semiparametric efficiency bound, then construct a one-step debiased estimator with asymptotic normality. A central challenge is that the information operator is anisotropic and does not commute with the tangent-space projection, creating a bottleneck absent from isotropic models. We introduce a score-whitening method that equalizes local Fisher information and restores stable inference at the optimal sample-complexity scale. Our results provide a principled framework for uncertainty quantification in LLM evaluation and more broadly for inference on low-rank structures from pairwise data.

Prompt learning has become a dominant paradigm for adapting vision-language models (VLMs) such as CLIP to downstream tasks without modifying pretrained weights. While extending prompts to both vision and text encoders across multiple transformer layers significantly boosts performance, it dramatically increases the number of trainable parameters, with state-of-the-art methods requiring millions of parameters and abandoning the parameter efficiency that makes prompt tuning attractive. In this work, we propose MMLoP (Multi-Modal Low-Rank Prompting), a framework that achieves deep multi-modal prompting with only 11.5K trainable parameters, comparable to early text-only methods like CoOp. MMLoP parameterizes vision and text prompts at each transformer layer through a low-rank factorization that constrains prompts to a compact subspace, providing parameter efficiency while motivating the need for our complementary regularization components. To further close the accuracy gap with state-of-the-art methods, we introduce three complementary components: a self-regulating consistency loss that anchors prompted representations to frozen zero-shot CLIP features at both the feature and logit levels, a uniform drift correction that removes the global embedding shift induced by prompt tuning to preserve class-discriminative structure, and a shared up-projection that couples vision and text prompts through a common low-rank factor to enforce cross-modal alignment. Extensive experiments across three benchmarks and 11 diverse datasets demonstrate that MMLoP achieves a highly favorable accuracy-efficiency tradeoff, outperforming the majority of existing methods including those with orders of magnitude more parameters, while achieving a harmonic mean of 79.70% on base-to-novel generalization. Code is available at https://github.com/sajjad-ucsb/MMLoP.

Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model's performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model's geometry, as captured by existing metrics, primarily reflects training choices rather than performance.

Foundation models have achieved remarkable success, yet their growing parameter counts pose significant computational and memory challenges. Low-rank factorization offers a promising route to reduce training and inference costs, but the community lacks a stable recipe for training models from scratch using exclusively low-rank weights while matching the performance of the dense model. We demonstrate that Large Language Models (LLMs) can be trained from scratch using exclusively low-rank factorized weights for all non-embedding matrices without auxiliary "full-rank" guidance required by prior methods. While native low-rank training often suffers from instability and loss spikes, we identify uncontrolled growth in the spectral norm (largest singular value) of the weight matrix update as the dominant factor. To address this, we introduce Spectron: Spectral renormalization with orthogonalization, which dynamically bounds the resultant weight updates based on the current spectral norms of the factors. Our method enables stable, end-to-end factorized training with negligible overhead. Finally, we establish compute-optimal scaling laws for natively low-rank transformers, demonstrating predictable power-law behavior and improved inference efficiency relative to dense models.

As unsupervised pretraining becomes increasingly ubiquitous in reinforcement learning, a more thorough theoretical understanding of these methods becomes of equal importance to their empirical success. We focus on the setting of unsupervised learning via interaction, where the forward-backward (FB) representation learning serves as a prototypical and popular example. In this paper, we shed light on FB by formally contextualizing the method within a broader class of recent methods that use regression to obtain a low-rank approximation of a successor measure ratio. Our analysis clarifies when FB representations can exist and how the low-rank approximation converges in practice. Building upon the theory, we propose a variant of FB that is both more amenable to theoretical understanding and simpler to optimize in practice. Experiments in didactic settings, as well as in $10$ state-based and image-based continuous control domains, demonstrate that our method converges to desired representations with $10^5 \times$ smaller errors than FB, achieving $+24\%$ improved zero-shot performance on average. We also demonstrate that zero-shot policies inferred by our algorithm provide an efficient initialization if the user prefers further fine-tuning on downstream tasks. Our project website is available at https://chongyi-zheng.github.io/onestep-fb.

Parameter-efficient continual learning aims to adapt pre-trained models to sequential tasks without forgetting previously acquired knowledge. Most existing approaches treat continual learning as avoiding interference with past updates, rather than considering what properties make the current task-specific update naturally preserve previously acquired knowledge. From a knowledge-decomposition perspective, we observe that low-rank adaptations exhibit highly imbalanced singular value spectra: a few dominant components absorb most of the adaptation energy, thereby (i) more likely to disrupt previously acquired knowledge and (ii) making the update more vulnerable to interference from subsequent tasks. To enable explicit balance among components, we decouple the magnitude of the task update from its directional structure and formulate it as a constrained optimization problem on a restricted Stiefel manifold. We address this problem using a projected first-order method compatible with standard deep-learning optimizers used in vision-language models. Our method mitigates both backward and forward forgetting, consistently outperforming continual learning baselines. The implementation code is available at https://github.com/haodotgu/EBLoRA.

Mixture-of-Experts (MoE) models scale neural networks by conditionally activating a small subset of experts, where the router plays a central role in determining expert specialization and overall model performance. However, many modern MoE systems still adopt linear routers in raw high-dimensional representation spaces, where representation mismatch, angular concentration, and scale-sensitive scoring can jointly undermine routing discriminability and stable expert specialization. In this work, we propose Low-rank & Lipschitz-controlled Routing (L2R), a unified routing framework that reshapes both the routing space and scoring geometry. L2R performs expert assignment in a shared low-rank latent routing space and introduces Saturated Inner-Product Scoring (SIPS) to explicitly control the Lipschitz behavior of routing functions, yielding smoother and more stable routing geometry. In addition, L2R incorporates a parameter-efficient multi-anchor routing mechanism to enhance expert expressiveness. Experiments on an OLMoE-based language MoE model and a ViT-based ImageNet setting show improved overall performance in both domains; OLMoE diagnostics further show improved routing geometry and expert discrimination.

Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning methods, such as LoRA, are widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches typically assign identical LoRA ranks to all expert modules, ignoring the heterogeneous specialization of pretrained experts. This uniform allocation leads to a resource mismatch: task-relevant experts are under-provisioned, while less relevant ones receive redundant parameters. To address this, we propose DR-LoRA, a Dynamic Rank LoRA framework for fine-tuning pretrained MoE models. Specifically, DR-LoRA initializes all expert LoRA modules with a small active rank and uses an expert saliency score, which combines routing frequency and gradient-based rank importance, to identify which experts would benefit most from additional capacity. It then periodically expands the active ranks of the task-critical expert LoRA, progressively constructing a heterogeneous rank distribution tailored to the target task. Experiments on three MoE models across six tasks show that DR-LoRA consistently outperforms LoRA and other strong baselines, demonstrating that task-adaptive heterogeneous rank allocation is an effective strategy to improve active capacity utilization in MoE fine-tuning.

Model merging has emerged as a promising paradigm for enabling multi-task capabilities without additional training. However, traditional basic merging methods often experience performance degradation due to parameter conflicts, even when applied to similar tasks. While recent personalized merging frameworks successfully preserve task-specific information to maintain performance, they typically incur storage overhead. In this paper, we propose Decomposition, Thresholding, and Scaling (DTS), an approximation-based personalized merging framework that pushes task-specific storage efficiency. DTS first applies singular value decomposition to the task-specific information and retains only a small subset of singular values and vectors. It then introduces a novel thresholding strategy that partitions singular vector elements into groups and assigns a scaling factor to each group. To enable generalization to unseen tasks, we further extend DTS with a variant that fuses task-specific information in a data-free manner based on the semantic similarity of task characteristics. Extensive experiments demonstrate that DTS consistently outperforms state-of-the-art baselines while requiring only 1% extra storage per task. Furthermore, experiments on unseen tasks show that the DTS variant achieves significantly better generalization performance. Our code is available at https://github.com/krumpguo/DTS.

As deep learning (DL) techniques become integral to various applications, ensuring model fairness while maintaining high performance has become increasingly critical, particularly in sensitive fields such as medical diagnosis. Although a variety of bias-mitigation methods have been proposed, many rely on computationally expensive debiasing strategies or suffer substantial drops in model accuracy, which limits their practicality in real-world, resource-constrained settings. To address this issue, we propose a fairness-oriented low rank factorization (LRF) framework that leverages singular value decomposition (SVD) to improve DL model fairness. Unlike traditional SVD, which is mainly used for model compression by decomposing and reducing weight matrices, our work shows that SVD can also serve as an effective tool for fairness enhancement. Specifically, we observed that elements in the unitary matrices obtained from SVD contribute unequally to model bias across groups defined by sensitive attributes. Motivated by this observation, we propose a method, named FairLRF, that selectively removes bias-inducing elements from unitary matrices to reduce group disparities, thus enhancing model fairness. Extensive experiments show that our method outperforms conventional LRF methods as well as state-of-the-art fairness-enhancing techniques in terms of improving fairness while maintaining accuracy. Additionally, an ablation study examines how major hyper-parameters may influence the performance of processed models. To the best of our knowledge, this is the first work utilizing SVD not primarily for compression but for fairness enhancement of DL classifications.