Low-Rank Adaptation

Also known as LoRA

Overview

Low-rank adaptation studies how to specialize pretrained models by representing weight changes as products of small matrices while keeping most backbone weights frozen. Methods choose where adapters are placed and how their rank is allocated, using approaches such as spectral or gradient-based initialization, adaptive rank selection, and quantized fine-tuning. For multiple tasks or users, adapter routing, expert mixtures, merging, and federated aggregation aim to share useful updates without interference. The central tension is how to gain enough capacity for new tasks while controlling memory and compute, preserving existing capabilities, and generalizing across data shifts; the best rank, placement, and composition strategy depend on the model and task.

Latest papers

579 papers65 papers in the last four weeks, down 19% on the four weeks before. 0.4% of all new papers.

Nov 12, 2026

Maximizing Capacity Utilization of LoRA Adapters for Mitigating the Learning-Forgetting Trade-off

Low-Rank Adaptation (LoRA) often underperforms full fine-tuning, especially in heterogeneous and continual-learning settings in which substantial forgetting occurs. In general, it suffers from an undesirable learning-forgetting trade-off across diverse fine-tuning settings. We identify a potential contributor: LoRA adapters, especially high-rank ones, frequently fail to realize the available capacity afforded by the configured rank. Specifically, we find that standard LoRA updates have effective rank far below their configured rank, limiting how broadly the update energy is distributed. Therefore, we propose training mechanisms that optimize the original objective while maximizing adapter effective rank. Our main method adds a simple, stable, differentiable, and SVD-free Rényi-2 effective-rank regularizer to the language-modeling loss. We also connect capacity maximization to norm-bounded updates. In mixed-domain rank sweeps on three models evaluated with multiple random seeds, a penalty weight of λ=0.01λ=0.01 lowers mean Wiki perplexity at every tested rank of at least 1616 relative to unpenalized LoRA. Mean training perplexity improves or increases by at most 0.020.02 in all but one model-rank comparison. Single-task and scheduling experiments further show that the balance between learning and retention depends on the configuration. We also observe that several LoRA variants in the literature also suffer from the identified capacity issue, and our approach provides a complementary axis of improvement to LoRA-based parameter-efficient fine-tuning. These results suggest that effective rank is not only a post-hoc diagnostic of adapter quality, but a trainable objective for improving the learning-forgetting frontier. Our code is available at https://github.com/tomohiro-sawada/lora-capacity-utilization
Nov 12, 2026

SALT: State-Aligned Low-Rank Transfer for Single-Temporal Cross-Dataset Structural Localization in Remote Sensing

Change detection in remote sensing conventionally relies on spatially aligned bi-temporal imagery, which may be unavailable or unreliable in practical scenarios. Recent single-temporal approaches address this limitation by learning structural evidence, such as building regions, from a labeled source dataset and transferring it to unseen target imagery using only the current image. However, adaptation to the source dataset can over-specialize pretrained features and reduce their cross-dataset generalization, a risk that this paper argues is higher when the feature adaptation is parameterized independently of the states used for prediction. To address this issue, the current study introduces SALT, a State-Aligned Low-Rank Transfer framework. SALT keeps a DINOv3 backbone frozen and learns negative and positive structural states that both score each patch and, through State-Aligned Low-Rank Adaptation (SALA), span the rank-2 subspace in which patch features are corrected. Directional State Alignment supervises these corrections to raise the evidence of positive patches and lower that of negative ones, while Geometry Preservation limits changes to the pairwise relations among pretrained patch features. Experiments on LEVIR-CD and the WHU aerial building dataset show that, at the pixel level and in both transfer directions, SALT outperforms four baselines reproduced under a common source-only protocol, improving over the strongest reproduced baselines by 9.45 and 5.79 points in pixel-level AP and by 27.69 and 14.25 points in IoU (LEVIR-CD to WHU and WHU to LEVIR-CD, respectively), with only 0.068M trainable parameters and 40.2 ms per image. In order to facilitate reproducibility and further research, the source code of SALT is available at https://github.com/gthpapadopoulos/SALT
Nov 12, 2026

Reasoning Supervision and Capability Retention in LoRA Domain Adaptation

Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model. In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels. Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy. Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation. Dual-adapter composite training remains on the same side of this comparison and yields no further gain. Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.
Nov 12, 2026

TokenStream: Multi-Subject Personalized Video Generation with Token-wise LoRA and Guided Attention

We introduce TokenStream, an optimization-based framework that composes independently learned subject identities in text-to-video generation without a reference-video training dataset or joint optimization of each new subject combination. Learning from a single reference image makes subject appearance difficult to separate from static-image bias, while merging independently trained adapters can mix identities. Token-wise LoRA (ToRA) restricts low-rank updates to the cross-attention value projections of subject-associated text tokens, providing a separate adaptation path for each subject. A T2V-guided stream copies queries and keys from the frozen base model during early denoising to reuse its layout and motion priors while retaining personalized values. We also introduce Multi-DB, a 30-subject, 122-prompt benchmark with appearance-free prompts and varied interactions. On Multi-DB at 1.3B, TokenStream improves identity preservation from 0.422 for DreamBooth to 0.467 and textual alignment from 0.652 to 0.688.
Nov 12, 2026

Sustainable Federated Fine-Tuning with LoRA: Environmental Impact and Performance Footprint

Federated Learning (FL) with Parameter-Efficient Fine-Tuning (PEFT) enables efficient adaptation of Large Language Models (LLMs), yet its energy and carbon costs across hardware remain poorly understood. We evaluated federated Low-Rank Adaptation (LoRA) fine-tuning on NVIDIA A40, RTX 6000 Ada, and A100 GPUs using a 222^2 full factorial design with 135M/360M models and 10/100 local steps. CodeCarbon tracked energy and estimated CO2_2-equivalent emissions. Descriptive allocation of variation assigned 84.8-91.2% of energy and emissions variation to local steps within these ranges. On A40 and RTX 6000 Ada, configurations with fewer steps achieved comparable accuracy with roughly 6-8-fold lower energy consumption and higher accuracy per kWh. These findings support smaller local budgets for energy-efficient federated fine-tuning within the evaluated setting.
Nov 11, 2026

WaveletLoRA: Frequency-Aware Content-Style Decomposition for Personalized Image Generation

Content-style decomposition from a single image is a fundamental challenge in personalized image generation, aiming to separate subject identity from visual style for flexible recomposition. Existing approaches primarily operate in the spatial domain, without explicitly exploiting the distinct frequency characteristics naturally associated with content and style. In this paper, we propose WaveletLoRA, a novel frequency-aware framework for explicit content-style decomposition through wavelet-domain supervision. Specifically, we first apply a multi-level discrete wavelet transform (DWT) to decompose the reference image into low-frequency and high-frequency subbands. We then assign a dedicated LoRA branch to each subband, where the low-frequency branch models global appearance for style acquisition, while the high-frequency branches capture structural details for content preservation. To further improve disentanglement quality, we introduce a dense Mixture-of-Experts aggregation module together with a frequency-domain regularization objective, which reduces style leakage and enhance content fidelity. In addition, we establish WaveBench and develop a VLM-based evaluation protocol that explicitly measures both attribute similarity and attribute separation for disentanglement assessment. Extensive experiments demonstrate that WaveletLoRA consistently outperforms existing methods and produces results that better align with human preference.
Nov 10, 2026

A Margin Perspective on LoRA: Robustness to Catastrophic Forgetting and Adapter Merging (MaLoRA)

Low-rank adaptation (LoRA) is the de facto method for parameter-efficient fine-tuning of large neural networks, achieving strong task adaptation with minimal trainable parameters and low optimization cost. Beyond this efficiency, practitioners have observed two striking properties: robustness to catastrophic forgetting and the ability to merge independently trained adapters into a single model that performs competitively across multiple tasks. These properties are central to continual and multi-task learning, yet remain poorly understood. In this work, we provide a theoretical explanation through a unified margin-based perspective. We analyze LoRA in multiclass linear classification under a near-orthogonal task regime with ℓ2\ell_2-regularization, and characterize the optimal LoRA adapter across regularization regimes. Our analysis shows that, in an intermediate regime of regularization parameters, the optimal adapter aligns with the max-margin solution on the fine-tuning data. Building on this characterization, we derive two key consequences. First, we obtain closed-form expressions for the margins on pre-training and fine-tuning data, revealing a precise margin trade-off: the regularization parameter controls the balance between retention and adaptation. Second, we analyze adapter merging, proving that merged models achieve positive margin on each task and deriving optimal mixing coefficients that balance margins across tasks and maximize the margin over their union. These results lead to a simple, training-free merging rule, which we term margin-based LoRA merging (MaLoRA). Experiments on modern architectures and real datasets validate our theoretical predictions, showing that MaLoRA matches or outperforms several adapter-merging baselines across a range of vision and language classification tasks.
Nov 10, 2026

TaoMate-R1: Temporal Role Specialization for Streaming Multi-Reference Video Generation

Streaming multi-reference video generation establishes a scene from specified people and objects, then continues it under changing instructions. We introduce TaoMate-R1, a streaming multi-reference generator that organizes adaptation parameters through temporal role specialization. It assigns separate low-rank adapters to scene establishment and continuation on a frozen backbone. Each adapter performs both denoising and context writing for its role; Self Gradient Forcing jointly trains the roles through the history they construct. This allocation contrasts with operation-based specialization, which separates denoising from context writing across all chunks. A fixed-replay gradient analysis characterizes the update directions admitted by the two sharing schemes. On 32 multi-reference cases, we compare the schemes at the same two-adapter budget. Across ten approximately two-minute streams with changing instructions and shared teacher-audio conditioning, TaoMate-R1 reduces the mean boundary luminance jump from 1.96 to 0.613, with higher boundary feature similarity. It also achieves higher mean CLIP and DINO temporal similarity across short interactions, long streams, and public reference compositions. These results support grouping adaptation parameters by scene establishment and continuation for streaming multi-reference generation. Our code and model are released at https://github.com/TaoLiveAIGC/TaoMate-H3.
Nov 9, 2026

Train High, Ship Low: Efficient LoRA Synchronization for Distributed On-Policy Post-Training

Transmitting LoRA adapters to rollout workers can bottleneck distributed on-policy post-training over bandwidth-constrained networks. Lowering the training rank may limit the capacity a workload needs, while adaptive optimizer updates are difficult to compress. Rollout workers can instead reconstruct updates from gradients by maintaining replicated optimizer states. We introduce HiLo, a compressor-agnostic framework that replaces parameter transmission with compressed gradient dissemination and local optimizer replay. By applying the same decoded gradients, the learner and workers maintain consistent policy states at the same version under reproducible execution. We further develop Spar, a budget-aware gradient compressor that combines spectral truncation, adaptive rank allocation and error feedback to decouple training rank from communication cost. Experiments across on-policy distillation and reward-guided post-training show that HiLo with Spar maintains learning quality close to full synchronization while transmitting as little as 1.6% of the adapter bytes, and accelerates end-to-end updates by up to 2.2×2.2\times under bandwidth constraints.
Nov 8, 2026

PROLA: Principal-Orthogonal Low-rank Adaptation for Predictive Spatiotemporal Weather Downscaling

High-resolution weather prediction is important for resolving local atmospheric patterns missed by coarse global forecasts. Large pretrained weather backbones have improved global forecasting, but adapting them to produce future high-resolution fields from recent coarse states remains challenging and costly. We formulate this setting as Predictive Spatiotemporal Weather Downscaling (PSWD), a task that jointly refines spatial and temporal resolution while predicting future high-resolution trajectories. To this end, we propose a framework that decomposes prediction into spatial and temporal stages using a shared pretrained backbone, fixed numerical scaffolds, and lightweight residual heads. For backbone adaptation, we introduce PROLA (PRincipal-Orthogonal Low-rank Adaptation), which allocates a fixed low-rank budget between pretrained weights' principal directions and their orthogonal complement. PROLA further rescales rank-1 optimizer updates using gradient signals, improving adaptation without increasing trainable rank. Experiments in the PSWD setting show that PROLA outperforms representative low-rank adaptation baselines. Further experiments across diverse datasets and tasks support the broader applicability of PROLA, showing that its benefits extend beyond the primary weather downscaling setting.
Nov 7, 2026

Zeroth-Order Riemannian Optimization on Fixed-Rank Manifolds for LLM Fine-Tuning

Zeroth-order (ZO) optimization estimates descent directions from function values alone, which removes the activation storage required by backpropagation and makes it attractive for memory-efficient fine-tuning of large language models (LLMs). Existing low-rank ZO methods either parameterize the additive update through Euclidean factors with a non-unique representation, or restrict each finite-difference direction to be low-rank. In this work, we propose to constrain each additive update block to a fixed-rank matrix manifold while keeping the pretrained weights frozen, so the learned update remains rank-constrained throughout fine-tuning. On this geometry, we build a Riemannian ZO gradient estimator that projects a random matrix onto the tangent space, normalizes it to correct the scale mismatch induced by projection, and further propose Zeroth-order Riemannian Gradient Descent (ZO-RGD) and its smoothed variant, ZO-RGD-S. We show that the ambient finite difference matches its retracted counterpart, the normalized estimator recovers the Riemannian gradient in expectation, and ZO-RGD converges under trajectory-regularity conditions. Experiments on LLM fine-tuning tasks show that the proposed algorithms are competitive with strong forward-only baselines across model families and tasks, with gains in several settings.
Nov 7, 2026

RBLA+: Efficient Rank-Specific LoRA Adapter Extraction for Federated Fine-Tuning with Heterogeneous Clients

Federated fine-tuning with low-rank adaptation (LoRA) reduces communication and computational overhead by training and exchanging compact trainable LoRA factors. Clients may use different LoRA ranks to meet adapter communication budgets and accommodate different local adaptation needs. The server must process these differences during aggregation and construct adapters at each client's rank for local training. We distinguish pre-aggregation alignment, which handles factor dimensions and scaling conventions before aggregation and rank-specific LoRA adapter extraction, which constructs client-sized pairs from the aggregated global model. Two common approaches to rank-specific extraction are direct rank truncation and Singular Value Decomposition (SVD) of the dense update. Direct prefix truncation keeps the leading columns of one aggregated factor and the corresponding rows of the other. It is cheap but can increase reconstruction error. The dense SVD approach first forms the product of aggregated LoRA factors. It then decomposes this matrix and retains the leading singular components to construct adapters at the target rank. This gives the optimal rank-constrained approximation in Frobenius norm but incurs significant computational overhead. To address this trade-off, we propose RBLA+ as an extension of RBLA. It uses established compact QR/core-SVD decomposition to minimize Frobenius error for the extracted LoRA update without forming the dense update. It also accounts for client and server scaling conventions during upload and broadcast. Across vision and language tasks, RBLA+ improves accuracy over RBLA in multiple heterogeneous settings. In CPU benchmarks with synthetic adapters, RBLA+ has lower extraction time than the dense partial-SVD baseline. RBLA+ preserves each client's prescribed adapter size and per-round communication payload.
Nov 6, 2026

FedLoVA: Value-Only Aggregation for Federated LoRA Fine-Tuning of Large Language Models

Low-Rank Adaptation (LoRA) is a popular parameter-efficient fine-tuning (PEFT) method for fine-tuning of large language models (LLMs), especially in federated learning, due to its strong performance and communication efficiency. In practice, LoRA is applied to the query and value projections of transformer layers without considering the distinct roles of these components. In this work, we analyze the impact of these projections in federated fine-tuning of LLMs using LoRA and observe that the value projection converges significantly faster than the query projection. Motivated by this finding, we propose FedLoVA, a federated fine-tuning framework that trains LoRA adapters on both query and value projections locally while sharing only the value projection with the server. This approach reduces communication overhead while improving convergence efficiency. Experiments on natural language understanding, multilingual sentiment analysis, and generation tasks show that FedLoVA reduces the communication cost by up to 7.8×7.8\times compared to existing federated LLM fine-tuning baselines while achieving comparable or better performance. The code is available at: https://github.com/EnsiehKhazaei/FedLoVA.
Nov 6, 2026

Breaking the Privacy Promise: Gradient Inversion Attacks on Federated Low-Rank Adapted Vision Transformers

Parameter-efficient fine-tuning (PEFT) has been increasingly adopted to fine-tune pretrained models with low computational cost and is therefore also preferred in the federated learning paradigm to reduce computational and communication overhead. In this work, we investigate the sparsely explored problem of the risk of private data leakage in LoRA-adapted Vision Transformers (ViTs), where low-rank updates weaken the gradient signal available to a server with access to individual client updates. We show that quality of reconstruction is limited by two coupled ambiguities: reduced information in LoRA gradients and missing spatial cues from frozen positional embeddings. To address this, we propose a spatially anchored generative inversion framework that incorporates server-prescribed trainable positional embeddings into ViT blocks during fine-tuning at clients. The resulting positional gradients act as layer-wise spatial anchors, providing additional positional cues, while the generative prior keeps reconstructions on a realistic image manifold, reducing duplicate-data and patch-permutation artifacts, {which produce similar gradients as those produced when the target client images are used}. Adding our proposed spatial anchoring to GIAS improves PSNR by 10.14 dB, SSIM by 0.45, and LPIPS by 0.098 on CIFAR-10, as well as PSNR gains of 7.83 and 11.74 dB on high-resolution CelebA and CARS datasets, respectively, in the FedSGD setting. Such gains in reconstructed image quality metrics are also maintained at higher batch sizes and in FedAvg settings. These results indicate that LoRA-based PEFT does not, by itself, prevent data leakage in FL, and carefully incorporated positional embeddings can help in exposing private information.
Nov 6, 2026

Steerable Multi-Objective Group-Relative Preference Optimization with Jointly Trained Low-Rank Adapters

Language model deployment often requires different trade-offs between competing objectives as user preferences change, yet standard post-training methods typically produce a single policy for a fixed objective weighting. We introduce LambdaMix, a new parameter efficient method for learning a preference conditioned family of policies with group-relative online reinforcement learning. LambdaMix jointly trains one low-rank adapter per objective over a frozen base model. One preference vector selects the adapter mixture that generates responses and weights the separately normalized rewards used to update the adapters. A single training run supports changing preferences at inference without further optimization. We experiment with four models on helpfulness and harmlessness and on solving math problems while following instructions. We also evaluate the two largest models on rewriting comments to preserve wording while reducing clues about personal attributes. On the first two tasks, LambdaMix achieves higher hypervolume than all evaluated conditioned and policy merging baselines across all four models. LambdaMix uses fewer training completions and stores fewer adapter parameters than a set of GRPO specialist policies. With architecture and initialization fixed, varying adapter mixtures and objective weights together gives wider score ranges and higher coverage than varying either alone.
Nov 6, 2026

From LoRA to Language: Recovering Skill Documents from Adapters

Agent skills are short documents that teach a language model how to handle a class of tasks. To save context, these documents are increasingly compiled into low-rank adapters (LoRA), which keep the behaviour but lose the text: an adapter is hard to read, edit, audit or move to another model. We ask whether the skill document can be written back from the adapter and its frozen base model, and we score each recovered document by the share of the original's key instructions it states and by the tasks it solves. Reading the weights alone recovered little with the methods we tried; a learned map from weights to text passed on about one bit of the document. Running the adapter worked much better. The Executing Writer, a small writer trained on top of the adapted model, recovers 39-44% of the key instructions of held-out skills, against 1.3-1.5% with another skill's adapter, and on adapters fitted to 41 external skills its documents restore 85-91% of the original's task gain. These strongest results are on adapters whose training stored the document; adapters trained on behaviour alone are much harder for trained writers (1-2% of instructions). For these adapters we test an agent that runs the base and adapted models and writes a skill from where they differ. On ALFWorld its skills recover 58-71% of a LoRA's gain and carry over to models that cannot load the LoRA, lifting Llama-3.1-8B from 5% to 55%. Whether such a skill comes from the adapter depends on what the adapter learned: when its behaviour cannot be guessed, the adapter is the source; when it can, a reward loop without the adapter does as well.
Nov 5, 2026

Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement Learning

Low-rank adaptation (LoRA) is commonly used in both supervised fine-tuning (SFT) and reinforcement learning (RL), often under similar optimizer settings. Yet, how the different dynamics of the two stages manifest in the low-rank increment that LoRA optimizes remains unclear. Tracking the singular spectrum and subspaces of LoRA updates on mathematical reasoning tasks, we find that RL concentrates the spectrum at every learning rate, while SFT expands it at moderate learning rates; moreover, RL rotates the singular subspaces much faster than SFT at small learning rates. Interpolating between SFT and RL one ingredient at a time, we pinpoint group-mean baseline subtraction as the source of both contrasts. A quadratic multi-index model updated via LoRA unveils the underlying mechanism: group-mean baseline subtraction tilts the balance between two competing forces, coherent fitting and gradient noise, toward spectral concentration and faster subspace rotation, while forcing the risk-favorable learning rate to be low. Guided by these dynamics, we tailor LoRA updates to each stage: in SFT, we flatten the spectrum of the full update to spread adaptation across singular directions, whereas in RL, we flatten the part of the update outside the current LoRA increment span to promote subspace rotation. Experiments on mathematical reasoning tasks across multiple models show that the respective configurations for SFT and RL yield the best performance among the compared methods in most evaluated settings.
Nov 5, 2026

Axis-LoRA: Probing and Adapting Embedding Models for Personal-Memory Retrieval

As LLMs are increasingly asked questions about a user's personal experiences, retrieval must distinguish memories by time, place, person, and event. Current personal-memory pipelines rely on frozen dense embedding models to embed memories for retrieval, but we find that these models do not capture the four axes well. Swapping a memory's date or place barely shifts its embedding, and a trained projection on top of the frozen embedding does not recover the missing signal. The problem lives inside the model, where each axis is encoded by a distinct attention-head pattern. Guided by these findings, we propose Axis-LoRA, a per-axis adapter that places four LoRA branches inside the embedding model and combines them via a per-query gate. Across three standard embedding models evaluated on LongMemEval, ATM-Bench, and LoCoMo, Axis-LoRA improves 4-benchmark mean R@10 by +1.28+1.28 to +1.56+1.56~pp while training only a small fraction of the base parameters on a small axis-counterfactual training set. Code is available at https://github.com/dohoonkim-ai/axis-lora.
Nov 5, 2026

Adapters as Data: The Latent Space of LoRAs

LoRA adapters are usually treated as finished artifacts rather than as a data modality. Prior work has shown that LoRA adapters populate a low-dimensional manifold that a linear PCA basis already organizes well enough to support different operations such as sampling, editing, and inversion. We ask what a learned, nonlinear latent adds beyond this. We train a transformer autoencoder on the same population of LoRA identity adapters, mapping each into a latent space and compare every operation against the identical operation applied directly to the raw adapter weights. The latent preserves the capabilities already available in weight space, sampling diverse identities and editing individual attributes. Beyond preservation, we show that the latent is far more robust to displacement than weight space, degrading into distorted but still valid faces rather than texture. Fixed-dimensional weight-space representations cannot natively host adapters of more than one rank, so we extend our sequential autoencoder to hold mixed-rank adapters within one shared latent. This gives a path toward treating adapter weights of varying architecture as a single, general modality rather than one tied to a fixed rank, and points to a learned, nonlinear latent as the mechanism that could carry this generalization further, toward heterogeneous populations of adapters.
Nov 5, 2026

Align Before Aggregation: Basis-Consistent Federated LoRA under Heterogeneous Ranks

Federated fine-tuning of large language models (LLMs) under heterogeneous client capacities is increasingly important, and low-rank adaptation (LoRA) with heterogeneous ranks makes it practical. Existing server-side aggregators compose local matrix updates ΔWiΔW_i from client-specific LoRA factors (Ai,Bi)(A_i,B_i) before aggregating them. Although these matrix updates lie in the same weight space Rm×n\mathbb{R}^{m\times n} and can be averaged directly, the LoRA factors that generate them may expressed client-specific local basis. In the typical compose--aggregate--factorize--dispatch loop, this basis inconsistency can make the rank-rir_i reference prefix dispatched to lower-rank clients unstable after server-side refactorization. In this work, we propose FedAbA, a basis-consistent federated LoRA aggregation framework for heterogeneous ranks. FedAbA first extracts each client's local basis, aligns it to the corresponding prefix of the global reference basis, and constructs a basis-consistent reconstruction of each local matrix update before aggregation. It then aggregates these reconstructions with weights that combine client data size and consistency score, followed by exact coefficient-space SVD refactorization and gauge fixing before rank-prefix dispatch. Our theoretical analysis clarifies the role of basis consistency and supports the proposed consistency-aware aggregation mechanism. Extensive experiments show that FedAbA outperforms representative federated LoRA baselines, with the largest gains under the strongest rank heterogeneity.
Nov 5, 2026

From What Is Used to What Is Missing: Rethinking Dynamic Rank in LoRA

Parameter-efficient fine-tuning with LoRA is governed by rank, which controls parameter budget and locally accessible task updates. Existing dynamic-rank methods mainly focus on capacity allocation, leaving unresolved whether the current adaptation space is sufficient for the task. We characterize the first-order information boundary of standard factor gradients, showing that their unobservable subspace coincides with the normal space of the low-rank manifold. Thus, factor gradients alone cannot distinguish under-optimization within the current space from task-relevant directions beyond its first-order reach. This motivates a geometric view of rank evolution, where tangent information optimizes existing degrees of freedom and the normal spectrum quantifies the value of new ones. Under a local quadratic model, we show that leading normal singular directions yield the locally optimal rank expansion, with spectral energy determining its gain. Building on this characterization, we propose GEAR, which turns the tangent--normal decomposition into rank evolution through tangent-space optimization, matrix-free normal-spectrum estimation for rank expansion, and task-induced geometric initialization. Across language and vision benchmarks, GEAR achieves the best average performance in all evaluated settings with substantially fewer trainable parameters. On LLaMA3-8B, it raises the average commonsense score from 86.61 to 88.45 while reducing trainable parameters from 42.0M to 14.4M.
Nov 4, 2026

The Behavioral Rate of Fine-Tuning: Compression as a Measure of Useful Information

Low-Rank Adaptation (LoRA) adapts a frozen language model through a compact learned update, making it a natural setting for asking how much information fine-tuning needs to write. Yet adapter rank, parameter count, or numerical reconstruction error do not directly tell us how many bits are required to preserve the acquired behavior. We therefore define an operational behavioral rate R⋆R^\star as the smallest serialized size of a compressed adapter that, once reloaded into the frozen base model, retains a fixed fraction of its held-out fine-tuning gain. Across LoRA fine-tuning experiments, this rate increases with distinct training content at fixed optimization budgets, but is not determined by intuitive metrics like source description length, fine-tuning gain, or weight-space reconstruction error. Most importantly, the same datasets exhibit substantially different, and even reordered, rates across frozen base models, showing that fine-tuning compressibility is a property of the dataset--receiver pair rather than of the dataset alone. However, we show that correction spectra measured on the data are able to rank these rates within a receiver. Finally, we study a striking noise-reducing effect of compression: adapters trained on deliberately corrupted datasets fall far below the frozen model, yet after compression score above it (23.7% to 82.9% on GSM8K, against 77.6% for frozen Qwen2.5-7B), consistently across tested families and sizes. The damage is unevenly concentrated in the adapter's leading singular directions, which compression appears to filter. Away from its training task, the compressed corrupted adapter also generalizes better than a clean fine-tune: on harder math it retains 97% of base accuracy, versus 44% for the clean adapter.
Nov 4, 2026

Model Merging Does Not Require Uniform LoRA Ranks

Model merging combines separately fine-tuned experts into a single model without retraining. Standard practice trains every expert at the same LoRA rank, which couples the rank choice across experts even though each domain has its own best rank. We show this coupling is a property of the merge method, not of merging. We train 36 experts on three domains (mathematics, instruction following, multilingual) at four ranks and three model scales, then merge with six methods at uniform ranks and at each domain's own best rank. Giving experts different ranks costs averaging and utility-based selection nothing, and gains them up to 1.8 pp of multi-domain accuracy, while the conflict-resolution methods are inconsistent and can lose up to 4 pp. The split has a simple cause. Merging keeps only a fixed share of the pooled spectral components, and methods that choose them by size discard many that would have reduced loss. They also tend to discard the same expert: the instruction adapter has the smallest singular values, so it is dropped almost entirely and IFEval falls by 16 to 25 pp. Choosing components by their effect on loss instead keeps all three experts, and leaves accuracy nearly flat across an 8×\times rank range (0.7 pp, against 6.7 pp for TIES at 4B scale). Rank can therefore be chosen per expert whenever the merge operator does not select by magnitude.
Nov 3, 2026

SaMA: Morpho Adaptation via Asymmetric Expansion of Kronecker Product

Low-rank adaptation (LoRA) has attracted significant attention in the parameter-efficient fine-tuning (PEFT) landscape. However, its design suffers from two fundamental limitations. First, its invariance manifold induces flat directions in the loss landscape, creating an optimization bottleneck for adaptive optimizers. Second, its effective rank scales only linearly with the parameter budget. In this paper, we propose Scalable Morpho Adaptation (SaMA), a PEFT method grounded in a principled generalization of the Kronecker product. By relaxing the shared-block constraint in the perfect-shuffle factorization, SaMA spans from low to full rank and achieves a practical rank that scales quadratically with the parameter budget, up to full-rank. Furthermore, we show that its invariance manifold collapses to a diagonal group, making it structurally more optimizer-friendly. Empirically, SaMA consistently outperforms strong PEFT baselines on commonsense and arithmetic reasoning benchmarks while being more parameter-efficient.
Nov 3, 2026

LIRA: A More Interpretable Dynamic Loss-Informed Rank Allocation for Low-Rank Adaptation

Low-rank adaptation (LoRA) has become a widely used parameter-efficient fine-tuning method for large pretrained models. However, standard LoRA typically assigns a fixed rank profile despite layer- and stage-dependent adaptation needs, while recent adaptive-rank variants often rely on heuristic or first-order signals that provide only an indirect link between rank reallocation and local loss behavior. This leaves the rationale behind each rank reallocation insufficiently interpretable. Motivated by this gap, we propose LIRA (Loss-Informed Rank Allocation), a dynamic-rank LoRA method designed to make online rank reallocation more interpretable. LIRA scores each layer by the estimated magnitude of the local loss perturbation associated with its recent LoRA-factor movements under a second-order approximation, making the basis for each rank change interpretable term by term. By periodically applying DELTA (Displacement-Evaluated Loss-Trajectory Allocation), a displacement-based algorithm for LoRA rank reallocation, LIRA produces discrete effective ranks under matched budgets while remaining compatible with standard (A,B) adapters. On RoBERTa-base GLUE under the La-LoRA/LoRA protocol, LIRA outperforms the uniform-LoRA and adaptive-rank baselines reported under this protocol; Trajectory EMA attains performance competitive with Fixed-point Fisher at lower estimation cost.
Nov 2, 2026

EpiLoRA: Fusing Many LoRA Adapters into Low-Precision GEMMs for Serving and RL

Low-rank adapters (LoRA) let one base model serve many fine-tuned variants: multi-tenant engines keep a single copy of the base and batch the requests of many tenants, each with its own adapter. As base models move to 8-bit (FP8) and 4-bit formats, their general matrix multiplications (GEMMs) become cheaper, but the adapters do not: engines run them as separate small GPU kernels beside the GEMM, whose latency takes a growing share of every decoding step. An adapter projects its input down to a low rank and back up, and our key observation is that the GEMM needs the down-projection only in its epilogue, the final stage in which it writes its output. EpiLoRA, a plugin for the vLLM engine, therefore launches the down-projection first and the low-precision GEMM as a dependent kernel that streams its weights meanwhile and adds every tenant's up-projection in its epilogue. On NVIDIA Hopper GPUs, EpiLoRA decodes for a single adapter with up to 24% lower time per token than vLLM's fastest configuration, and faster than SGLang and TensorRT-LLM; with many tenants, its adapter kernels deliver more throughput than vLLM's on most GPU-model pairs we tested. The same low-precision base also trains: a LoRA reinforcement-learning trainer that computes on the engine's FP8 weights needs no 16-bit copy, so a 27B model trains on one 48 GB GPU, and an engine that then keeps activations in 16 bits removes most of the remaining mismatch between the model generating training data and the one trained.
Nov 2, 2026

BISECT: Bilevel Spectral Decomposition for Low-Rank Continual Learning

Exemplar-free continual learning on pre-trained Vision Transformers increasingly relies on Low-Rank Adaptation (LoRA), yet prevailing methods govern the low-rank update space through strict subspace isolation. By projecting each incoming task into the approximate null space of historical token activations, isolation-based learners equate any cross-task feature overlap with catastrophic interference. We show that this orthogonality premise induces a structural bottleneck we term spectral starvation: when down-projection bases are locked before task optimization, the gradient norm reaching the trainable up-projection is upper-bounded by the energy of current-task features captured by the chosen basis. Because natural visual tasks share low- and mid-level covariance structure, historical null tails contain barely 5.5%5.5\% of incoming task variance, starving new-class plasticity while blocking positive consolidation along shared directions. We propose BISECT (Bilevel Spectral Decomposition), a continual LoRA framework that partitions a fixed rank-rr budget into two orthogonal functional blocks: an invariant carrier spanned by the dominant modes of the joint historical-and-current Gram matrix, and a novelty complement extracted via generalized Rayleigh whitening followed by orthogonal complement projection. After single-task training, BISECT folds both blocks back into the frozen backbone using a closed-form precision gate that scales each rank-1 invariant atom by its current-to-joint Rayleigh ratio while preserving novelty atoms at unit gain. Across eight incremental settings on ImageNet-R, ImageNet-A, CIFAR-100, CUB-200, and DomainNet, BISECT outperforms prompt, adapter, and LoRA baselines with zero extra inference parameters or FLOPs, and remains complementary to sparse analytic expansion.
Nov 2, 2026

ASCoR: Asymmetric Semantic-Spatial Learning for Remote Sensing Domain Generalization

Source-only domain generalization (DG) for remote sensing semantic segmentation seeks reliable predictions in unseen domains without target-domain data during training. Although vision foundation models offer transferable knowledge, category recognition remains sensitive to source-specific visual cues, while task-specific adaptation may weaken pretrained transferability. We propose ASCoR, an asymmetric semantic-spatial framework built on SAM3. Its text-conditioned semantic branch uses class-name prompts to provide explicit category guidance. Probe-Once Dynamic Adaptive LoRA distributes a fixed adaptation budget across encoder blocks to limit source-specific specialization. A complementary task-specific dense branch learns multi-scale spatial representations using a separate frozen SAM3 image encoder. A lightweight reliability router integrates the two branches through bounded spatial correction based on local prediction reliability. All components are trained using source-domain supervision only. Experiments on the ISPRS, LoveDA, and OpenEarthMap benchmarks cover nine cross-domain transfers. ASCoR achieves a state-of-the-art mean IoU of 68.61% across six ISPRS transfers, outperforming the previous best method by 2.69 percentage points. It also achieves the best results on all three LoveDA/OpenEarthMap transfers, with 59.07%, 47.95%, and 49.38% mIoU on Rural-to-Urban, Urban-to-Rural, and Urban-to-OpenEarthMap, respectively. Ablation studies, feature visualizations, and spatial reliability analyses support the effectiveness and complementarity of the proposed components. Code: https://github.com/WangZitong42/ASCoR.
Nov 2, 2026

Unlocking the Potential of Vision Foundation Models for Multi-Task Dense Prediction via Mixture of Task Adapters

Multi-task dense prediction seeks a single, coherent representation that supports simultaneously discriminative semantic recognition and regularized geometric/structural inference. For vision foundation models, pretrained semantic invariances conflict with the frequency and smoothness biases of dense objectives; joint fine-tuning and single-route adapters therefore compress competing gradients into one update channel, causing negative transfer and unstable cross-task agreement. We frame this as two-level adaptation: task-generic alignment that preserves priors, and task-aware specialization that resolves late-stage conflicts. Based on this principle, we propose D2MoE , a Mixture of Task Adapters that explicitly factorizes these roles within a frozen DINOv3 backbone. D2MoE introduces Task-Generic Low-Rank Adaptation (TG-LoRA) in attention projections throughout the network to provide a stable, shared adaptation channel, and a Task-Adaptive Mixture of Experts (TA-MoE) injected only at the final attention block of each stage to enact sparse, task-conditioned perturbations of token mixing in a single forward pass. A Semantic Prior Injector further couples multi-level task features with frozen foundation priors to reinforce semantic coherence while preserving boundary and structural fidelity. Experiments across diverse benchmarks substantiate that this structured factorization yields robust multi-task improvements with a small trainable footprint, establishing a principled and scalable route to prior-preserving, foundation-centric multi-task dense perception.
Nov 2, 2026

LoRA-Patch: Defending Against Backdoored LoRA Adapters via Directional Steering

Third-party LoRA adapters introduce serious backdoor risks: malicious adapters can maintain normal task performance while embedding hidden, trigger-activated behaviors, making them difficult to detect through standard utility evaluations. Existing removal-based defenses typically sanitize backdoored models by updating, pruning, distilling, or mixing model parameters. However, these approaches are misaligned with adapter-based deployment, where the trusted base model is expected to remain frozen. In this work, we propose LoRA-Patch, a lightweight backdoor defense tailored to LoRA adapters. Our key insight is that backdoor injection distorts the directional geometry of a poisoned adapter relative to its trusted base model. Building on this insight, LoRA-Patch suppresses trigger-activated behaviors by realigning adapter directions through layer-wise steering, without modifying the base model weights. Extensive experiments across five backdoor attacks and three models show that LoRA-Patch effectively mitigates trigger-activated backdoors, outperforming existing baselines while preserving benign task performance. Moreover, LoRA-Patch is computationally efficient, constructing patches within 20 seconds across all evaluated models and adapters. Our code is publicly available at https://github.com/FOXVEDA/LoRA-Patch.