Parameter-Efficient Fine-Tuning

Also known as PEFT

Overview

Parameter-efficient fine-tuning studies how to adapt pretrained models while training only a small subset of weights or adding lightweight modules, reducing resource costs and sometimes helping preserve existing capabilities. Common approaches include low-rank updates such as LoRA and quantized QLoRA, adapters and soft prompts, selective layer or rank tuning, and sparse or routed experts. Methods are evaluated on language benchmarks such as GLUE and mathematical reasoning, as well as vision, speech, multimodal, and continual-learning tasks, often with limited data or hardware. A central tension is that fewer trainable parameters do not necessarily mean lower memory use or faster training, and accuracy, robustness, and retention can vary with the model, task, and placement of updates.

Latest papers

646 papers42 papers in the last four weeks, down 51% on the four weeks before. 0.3% of all new papers.

Nov 19, 2026

Parameter-Efficient Adaptation of Self-supervised Vision Backbones Survives Distribution Shift in Radar Micro-Doppler HAR

Micro-Doppler radar enables contactless, privacy-preserving human activity recognition (HAR), but deployment robustness remains underexplored. We compared full fine-tuning, frozen probing, and parameter-efficient fine-tuning (PEFT) using subject-disjoint in-distribution cross-validation and few-shot cross-age, cross-orientation, and cross-dataset transfer. Full fine-tuning suffered negative cross-dataset transfer, while frozen features under-adapted on cross-orientation. Only PEFT avoided both failures. We paired DINOv3 with all-layer principal-singular-direction adaptation (PiSSA), to our knowledge the first application of a self-supervised image foundation backbone to radar micro-Doppler HAR. DINOv3-PiSSA ranked first among PEFT methods in every evaluated setting. Under our protocol, its Small variant matched the in-distribution accuracy of the prior Base-size PEFT state of the art using 4.5x fewer total parameters, with only 3% trainable. It surpassed a size-matched variant of that baseline in every evaluated setting, with transfer gains increasing with shift severity: 2.1, 6.6, and 8.9 percentage points, respectively. Ablations attributed improvements to both pretraining and adapter recipe, with the pretraining contributing most under compound shift. Code and protocol are available at https://github.com/KeryanChelouche/udoppler-ssl.
Nov 18, 2026

Continual Tool-Space Extension Learning for Multimodal Agents

Multimodal agents largely rely on generic tools such as cropping, search, and code execution, leaving specialist vision models outside their tool space. However, bringing such a model into an already trained agent is not straightforward. The challenge is to learn each new tool efficiently while retaining the ability to use existing ones. This work proposes COTEX, continual tool-space extension learning, that adds new tools at efficient adapter-level cost. COTEX locates two sets of parameter directions; one triggers a tool call and thereby opens a multi-turn reasoning trajectory, while the other determines which tool is chosen. The adapter of a new tool is restricted to the calling directions that leave earlier tool choices unaffected by orthogonal constraint. We further introduce a protocol that converts specialist models' existing task-specific annotations into verified tool-use trajectories. Experiments on two representative agents and three specialist tools show gains on ten vision-language benchmarks, with at most a 0.3-point drop on existing tools.
Nov 17, 2026

BQ-LoRA: Binary-Quantized Low-Rank Adapters as Implicit Regularizers for Parameter-Efficient Fine-Tuning

Parameter-efficient fine-tuning of quantized large language models is dominated by QLoRA, which combines 4-bit weight quantization with FP16 low-rank adapters. We ask whether the FP16 adapter precision is necessary. Replacing it with a single binary bit per adapter parameter, combined with a learned per-layer scale ss and a Scaled Straight-Through Estimator, reduces adapter storage by 13-16x. We introduce two variants -- NormScale, which binarizes under a norm-preserving scale with a clipped STE, and LearnThresh, which makes the binarization threshold learnable -- unified under a framework we call BQ-LoRA. Our main theoretical contribution is a tight characterization of the BQ-LoRA hypothesis class: it has bounded entrywise norm, a typical-set spectral norm bound that is Θ(r)Θ(r) times sharper than the worst case, a full-rank, near-isotropic update, and consequently a Rademacher complexity that yields an explicit O~(smax⁡αdout/(rn))\tilde{O}(s_{\max}α\sqrt{d_{out}/(rn)}) generalization bound. FP16 LoRA has no comparable a priori norm or isotropy control. On Qwen2.5 (1.5B-72B), BQ-LoRA exceeds QLoRA's benchmark averages at every scale by 0.8-3.5 points; on LLaMA-2 (7B/70B) it remains competitive at the same storage reduction but does not uniformly lead the strongest FP16-adapter baselines (LoftQ at 7B, QLoRA at 70B). On Qwen2.5-7B the binary adapter preserves the 4-bit base WikiText-2 PPL within 0.06 (9.86 vs. 9.80) while QLoRA worsens it by 0.59. The reduction applies to inference-time adapter storage and serving; training memory matches QLoRA and step time rises by under 4%. We translate the storage saving into a deployment study, BQ-SGMV, by extending Punica's SGMV kernel with a binary inner-product primitive: the same A100-80GB holds 15.6x more adapters in HBM, and our research prototype eliminates cache misses under a skewed synthetic workload.
Nov 16, 2026

State-Message Interaction Adapter for Cross-Domain Few-Shot Segmentation

Cross-domain few-shot segmentation adapts natural-image pretrained encoders to medical or industrial domains from a few pixel-annotated support masks. Within frozen DINOv3, intermediate per-head attention messages already carry target-background distinctions that baseline final-layer similarities and predictions do not consistently reflect. We propose the State-Message Interaction Adapter (SMIA) to learn how per-head messages should update each receiving token representation while keeping pretrained attention and FFN parameters frozen. Inserted right after per-head aggregation, SMIA conditions the transformation of each head's aggregated output (message) on each patch's pre-attention feature (state), allowing similar contextual messages to produce distinct or opposing updates depending on the receiving position. Because scarce support masks cover only limited state-message combinations, a shared low-rank bilinear map confines the learned interactions to low-dimensional state, message, and output subspaces. Added before each adapted block's frozen FFN, these updates yield adapted features that are refined by frozen FFNs and deeper blocks for linear readout. With only 0.51M trainable parameters and single-pass, support-free inference, SMIA achieves six-dataset mean mIoUs of 77.40% at 5 shots and 80.99% at 10 shots across 20 medical and industrial tasks, surpassing the strongest prior methods by +3.61 and +3.93 percentage points. In controlled 10-shot comparisons across three benchmarks on the same DINOv3 backbone, SMIA also outperforms parameter-matched AdaptFormer, LoRA, and ParaX with a better precision-recall balance.
Nov 16, 2026

Tight PAC-Bayes Generalisation Guarantees for Large Language Model Judges

Large language model (LLM) judges are increasingly used to monitor the outputs of LLM-based AI agents, yet formal guarantees on the reliability of these judges remain largely absent, limiting confidence in their use for safety-critical oversight. We consider certifying such a judge and define its risk over safety classifications of prompt--response pairs drawn from a deployed generator. We apply a PAC-Bayes framework to this setting, refining existing compression-based methodology with the use of tighter bounds and more efficient coding. We show that standard parameter-efficient fine-tuning can yield non-vacuous and even tight certificates for safety judges, for both classification and Brier risk, without holding out data for validation. We introduce LoRA-GT, which shares a single quantisation scale across layers and achieves the tightest certificates with minimal degradation in predictive performance. Across five guardrail models and four benchmarks, we find that certificate tightness is governed more by adaptation complexity than by base model scale. Finally, we show how far quantisation can be pushed before the resulting increase in risk outweighs the bound tightening gained from shorter descriptions. Together, these results provide a practical route toward certifiable scalable oversight and offer insight into how the complexity of safety adaptations shapes generalisation.
Nov 14, 2026

MarineSparkLM-4B: A Lightweight Language Model for Marine and Offshore Engineering Built by Parameter-Efficient Adaptation

Marine and offshore engineering text is spread across preprints, open-access journals, encyclopedias and institutional documents, and the amount of it is small by pretraining standards. We ask how far a lightweight language model for marine and offshore engineering can be taken with parameter-efficient adaptation on a single workstation-class device. Starting from the 4-billion-parameter OceanGPT-basic-4B-Instruct, we train two independent LoRA adapters, each updating 33.0 million of 4.06 billion parameters (0.81%). Training took 4 hours 40 minutes for the first adapter and 10 minutes for the second on one NVIDIA DGX Spark. The first adapter is trained by next-token prediction on a 61.6-million-token corpus covering ten engineering themes. The corpus was built from 13,879 source documents, of which 7,069 were retained after deduplication and a two-tier keyword filter that removes cross-domain lexical collisions. Its validation loss falls from 2.082 to 1.976 over 1,500 steps. Direct testing showed that this adapter could still drift off topic on a retrieval-grounded ballast-water prompt, so we trained a second adapter on 2,295 question-answer pairs (114 hand-written and 2,181 generated by the base model from corpus passages) with the loss restricted to answer tokens. On 20 held-out prompts scored by hand for accuracy, topicality and specificity (maximum 120), the plain base model, the continuation adapter, the instruction adapter and the instruction adapter with retrieval scored 42, 56, 84 and 88. Absolute accuracy stays low (14 of 40 without retrieval and 20 of 40 with it). We report the full pipeline, two negative results (a reverted keyword fix and an instruction run that overfit within 100 steps), and the limits of the evaluation, which uses one small prompt set, no inter-rater agreement measure and model-generated reference answers.
Nov 13, 2026

PREFACE: Separating Shared Corrections from Prefix Conditioning in PEFT

Does input-conditioned parameter-efficient fine-tuning (PEFT) improve because it uses the current input, or because it adds a useful learned correction? We investigate this distinction with PREFACE, a bounded channel controller for PEFT-adapted projections that requires no test-time gradients. The controller reads observed context; in multiple-choice likelihood scoring, its state is independent of the candidate continuation. Our central study freezes IA3, tuned LoRA, and DoRA adapters and matches controller architecture, initialization, batches, and updates across current-prefix and learned-global controls. Across ARC-Challenge and SciQ, current-prefix control improves all six mean accuracies over frozen adapters, but none of its paired contrasts with learned-global control excludes zero. Replacing its corrections with their training-prefix mean changes accuracy by at most 0.34 percentage points; input-varying corrections contribute only 1.8-2.8% of correction RMS. On ARC test, frozen-adapter current/global controls score 25.68%/25.77%, distinct from the earlier jointly trained 23.1% to 26.4% gain over IA3. We derive and evaluate static compilation of fixed corrections, separating deployment simplification from statistical equivalence. Broader GPT-2, Pythia, and Qwen experiments show family-, task-, and scoring-dependent gains, with modest recent-model effects. The study provides an intervention-based evaluation of conditional PEFT: improvements over a static adapter do not by themselves demonstrate a benefit from input matching.
Nov 13, 2026

Adaptable Watermarking: Transforming Image Watermarking Models via Lightweight Adapters

We explore the adaptability of neural image watermarking models; the extent to which one watermarking method can be transformed into another through lightweight adaptation. We introduce a framework that enables encoder-decoder watermarking architectures to become modular and upgradable, supporting multiple watermarking styles without re-training full models. Adaptation is performed by attaching a compact parameter-efficient adapter and fine-tuning under a multi-domain residual alignment loss (spatial, gradient, and frequency) while preserving bit accuracy via a secret-loss term. We evaluate this approach across a suite of recent watermarking models, including TrustMark, InvisMark, RoSteALS, HiDDeN, and WAM, demonstrating that adapted encoders closely reproduce target residuals and that adapted decoders recover messages with near-optimal accuracy under standard augmentations and distortions. We show that cross-model adaptation, with applications in interoperable and upgradable watermarking, can be achieved with as few as 1-2M trainable parameters per component (e.g., the 2M decoder adapter used in our multi-decoder inference pipeline).
Nov 13, 2026

Direction or Magnitude? Learning Scalars over Task-Calibrated Directions

Parameter-efficient fine-tuning (PEFT) aims to preserve adaptation quality while reducing training costs. We propose SCALAR, which achieves performance comparable to LoRA with substantially fewer trainable parameters, lower peak training memory, and higher throughput. SCALAR constructs a fixed input basis from task activations and assigns a random output dictionary to each input direction. Fine-tuning learns only layer-specific scalar coefficients, reducing gradient and optimizer-state storage. Grouped computation aggregates output atoms before applying the adapter to token activations, while dictionary sharing across compatible layers reduces fixed-basis storage. Experiments on commonsense reasoning, mathematical reasoning, code generation, and knowledge evaluation after instruction tuning demonstrate effective adaptation at small trainable budgets. On commonsense reasoning, SCALAR achieves accuracy comparable to LoRA with about 4% of its trainable parameters. System measurements on Qwen2.5-7B show 20% lower peak training memory and 6% higher throughput than LoRA. The combination of adaptation quality, a compact trainable state, and training efficiency makes SCALAR a promising approach to resource-constrained fine-tuning.
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 12, 2026

Quantum-inspired orthogonal adapters for certifiably robust medical image classification

Medical image classifiers require not only high accuracy but also reliable predictions under small input perturbations. Certified robustness provides mathematical guarantees of prediction stability, but preserving these guarantees when adapting pretrained models to new tasks remains challenging. We propose a fine-tuning approach based on quantum-inspired orthogonal butterfly layers that preserves the Lipschitz bounds of pretrained certified networks. Inspired by quantum circuits of reconfigurable beam splitters, these layers preserve orthogonality by construction using only O(nlog⁡n)\mathcal{O}(n\log n) trainable parameters. Their structured design enables efficient classical implementation and offers a path toward future quantum implementations with logarithmic circuit depth under suitable connectivity assumptions. Controlled experiments on MNIST and MedMNIST demonstrate improved certified robustness compared with conventional alternatives and, in the DermaMNIST transfer, greater resistance to the tested attacks than low-rank adaptation (LoRA). We then apply it to trichoscopic image classification for alopecia diagnosis and alopecia areata activity assessment. By adapting a certified ImageNet-pretrained network, butterfly fine-tuning achieves approximately twice the certified accuracy of conventional fine-tuning and LoRA, with comparable classification accuracy on the same backbone and data split and similar empirical resistance to the tested adversarial attacks. Larger robustness guarantees, including certificates covering certain image-processing transformations, can be obtained at the cost of reduced accuracy. These results highlight the potential of quantum-inspired butterfly layers for parameter-efficient adaptation of medical image classifiers while preserving the mathematical guarantees underlying certified robustness.
Nov 11, 2026

Paired Geometric Supervision for Generalizable Deepfake Detection

We present OrthoPairGTL, a geometric training framework for cross-domain deepfake detection, built on a partially frozen DINOv3 backbone (∼\sim300M parameters), and make three contributions: (1) We re-evaluate six recent transfer-learning detectors (Forensics Adapter, Effort, three GenD variants and FakeSTormer) using the authors' released checkpoints under a single preprocessing pipeline across 17 cross-domain datasets (2018-2025), so that the reported gaps reflect method rather than per-method pipeline tuning. (2) We model the transformation between paired real and fake samples, training Parameter-Efficient Fine-Tuning (PEFT) adapters with an orthogonal pair decomposition, hyperspherical regularisation, and (in the GSD-LoRA variant) a set-level coherence objective. (3) Under LoRA, this regularisation gives a more stable optimisation profile: GSD-LoRA's post-peak validation degradation is 4.2×4.2\times shallower than plain LoRA at the same parameter footprint, and the time-averaged validation AUC across training is 1.82 pp higher. On the 17 datasets, GSD-LoRA matches or exceeds every re-evaluated baseline (92.82% vs. 91.58% for the strongest; paired Wilcoxon signed-rank p=0.015p=0.015, 13/17 per-dataset wins), with the largest gains on modern reenactment-heavy benchmarks (DSv1.1, DSv2). Under strict parameter budgets, LayerNorm-only tuning (≈\approx0.03% of the backbone) is the most stable configuration and matches the strongest baseline with a mean of 91.97%.
Nov 9, 2026

3R-Adapter: Retrieval, Rewiring, and Refinement for Efficient Adaptation of 3D Reconstruction Model

Feed-forward 3D reconstruction models have demonstrated strong generalization under large-scale pre-training, yet they struggle in long-tail scenarios such as low-light reconstruction, joint human-scene reconstruction, and test-time adaptation. We observe relational attention structure collapse, manifested as the failure of cross-view correspondences, as a primary cause of these limitations. While adaptation offers a practical way to bridge this gap, existing parameter-efficient fine-tuning (PEFT) approaches mainly perform channel-wise feature adjustment and lack mechanisms to explicitly adapt relational structures, leaving correspondence failures unresolved. In this work, we formulate adaptation of 3D reconstruction models as a problem of correspondence recovery under domain shift and propose 3R-Adapter, a PEFT paradigm for long-tail 3D reconstruction scenarios. Our method consists of three components: Deformable Retrieval for task-adaptive relation candidate retrieval, Relational Rewiring for reconstructing cross-view relational structures, and Consistency Refinement for injecting the rewired relations into stable predictions. Extensive experiments on low-light reconstruction, joint human-scene reconstruction, and test-time adaptation demonstrate that our approach significantly improves robustness over existing PEFT methods while maintaining efficient lightweight adaptation. Our code is released at https://github.com/PKU-OV3-LAB/3R-Adapter.
Nov 9, 2026

VeilTune: Secure Fine-Tuning of Private LoRA Adapters on Public Pretrained Language Models

Fine-tuning public pretrained language models with lightweight adapters such as LoRA has become the mainstream way to adapt them to downstream tasks. A service provider often trains its adapter on a data owner's samples, so two assets need protection: the data owner's samples and the provider's adapter. Existing secure fine-tuning protocols protect the samples but reveal to the data owner intermediate results of every training step, such as the hidden representations or logits of its inputs. We show that a single step suffices to copy the adapter: from the hidden representations of one mini-batch of 32 samples, the data owner fits an adapter that agrees with the provider's on 99.2% of test inputs, and logits alone yield about 94%. Protecting both assets therefore requires that no value that depends on the adapter be revealed throughout training. We present VeilTune, to the best of our knowledge the first protocol that fine-tunes a pretrained language model with backpropagation entirely in secure multi-party computation (MPC). VeilTune computes only what the loss depends on, replaces generic MPC operators with training-aware nonlinear operators that stay accurate on the values that training produces, and obliviously skips the steps corrupted by truncation failures. On four few-shot tasks with RoBERTa-large, VeilTune nearly matches plaintext training and is 5.3 points more accurate on average than SecP-Tuning, the state-of-the-art secure fine-tuning protocol. On SST-2, it achieves a 3.5x end-to-end speedup and a 3.6x reduction in communication over SecP-Tuning, and 26x and 17x over full fine-tuning in MPC.
Nov 8, 2026

TID System for Speech-COSER Challenge 2026: ASR Track

We describe the system submitted by Telefónica Innovación Digital to the Speech-COSER Challenge 2026, which features spontaneous speech from elderly speakers recorded in rural Spain. For ASR, we perform CTC-based pseudo-forced alignment scores audio-text match quality to guide a two-stage LoRA curriculum, followed by post-hoc magnitude pruning of the LoRA matrices to enhance generalization. An ensemble of seven systems combined via ROVER achieves 36.25% unnormalized and 22.17% normalized WER.
Nov 8, 2026

Full Fine-Tuning is not the Problem: Why Adam Fails and SGD Succeeds in Few-Shot CLIP Adaptation

Few-shot CLIP adaptation is often treated as a parameterization problem. Full fine-tuning is assumed to overfit, so parameter-efficient fine-tuning (PEFT) methods are preferred. In this paper, we show that this conclusion conflates parameterization with optimization. Across 11 datasets, 3 CLIP backbones, and 5 shot levels, dense full fine-tuning with plain SGD outperforms representative PEFT methods under the reported evaluation protocol, whereas Adam/AdamW collapses. We trace this collapse to unreliable support-set preconditioning. The few-shot second moment systematically underestimates the corresponding full-data statistic, already at initialization. Adam's inverse-square-root normalization converts this error into an oversized adaptive update. Matching Adam's first-step update budget to SGD resolves most of the collapse, whereas numerator-side alignment does not explain the failure. Shampoo and SOAP also fail, so richer preconditioning alone does not resolve the problem. Theory formalizes how few-shot estimation error is amplified into unsafe updates. SGD succeeds because it avoids unreliable preconditioners and stays close to pretrained CLIP. Interpolation probes show low-loss connections along tested directions, and curvature probes show favorable directional curvature. These results identify optimizer-induced preconditioning, not dense full fine-tuning itself, as the main source of failure in few-shot CLIP adaptation.
Nov 8, 2026

Not All Layers Are Equal in Image-to-Video Transfer

Parameter-efficient image-to-video transfer adapts frozen image foundation models to video via lightweight trainable adapters. Existing methods insert the same adapter at every transformer layer, implicitly assuming all layers require equal temporal modeling. A layer-wise sensitivity analysis across four video-native model families and two pretraining paradigms reveals that this assumption is wrong: temporal sensitivity increases with depth, and the pretraining paradigm determines its growth profile. Vision-only models exhibit a gradual rise, while vision-language models show early suppression followed by late recovery. This finding motivates STRAP, a training-free adapter placement rule that derives a spatial-to-temporal transition boundary from a paradigm-matched video model. Spatial adapters are assigned before this boundary and temporal adapters after it, without modifying the adapter architecture itself. On four fine-grained action recognition benchmarks, STRAP improves accuracy by up to 11.8% over uniform placement, reduces trainable parameters by 29 to 48%, and lowers inference time by up to 10%. On standard benchmarks, it matches accuracy with fewer parameters. Code is available at https://github.com/th-nesh/STRAP.
Nov 7, 2026

The Surprising Effectiveness of Deleting Weights in LLM Post-Training

How little of a pretrained large language model (LLM) has to change for it to acquire new capabilities? We find that zeroing fewer than 0.05% of its weights, with no other modification, matches full fine-tuning (FFT) at scale across three LLM post-training regimes: reinforcement learning with verifiable rewards, supervised fine-tuning, and on-policy distillation. Our method, Deletion Tuning, learns a binary mask over the gate projection of transformer blocks. Two practical advantages follow: the trained mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint, and Deletion Tuning forgets far less, which addresses a central challenge in continual learning. Under reinforcement learning, its prior-task accuracy stays near the base level while FFT, LoRA, and random sparse fine-tuning lose up to 5.4 points. To understand why masking alone suffices, we analyze the geometry of the learned weight changes under reinforcement learning using stable rank, spectrum drift, and principal-weight overlap. Compared with LoRA and random sparse fine-tuning, Deletion Tuning is closest to FFT across the three measures.
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 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

Backbone-Informed Adaptation after EEG Pretraining: A Capacity-Controlled Study

Electroencephalography (EEG) foundation models learn representations that combine information across channels and time. This structure motivates adapters that model the same relationships during downstream learning. However, it is not clear whether such a design outperforms other candidate adapters. Moreover, in order to arrive at a fair comparison between candidate adapters, we also need to consider their trainable parameters. Here, we ask whether a backbone-informed (native) adapter makes better use of a given capacity by comparing it against a token-wise multilayer perceptron control that transforms each feature vector independently. We approximately match their trainable parameter counts and keep the pretrained encoder fixed. Across two EEG foundation models and tasks covering emotion recognition, sleep staging, and EEG event classification, the native adapter shows no consistent advantage. For one backbone, differences between the designs are small relative to variation across runs. For the other, the token-wise control performs better on emotion recognition and EEG event classification. On sleep staging, the native adapter's advantage over the token-wise control is reduced when we also train the component that processes sequences of EEG segments. The relative performance of the two adapters depends on the backbone, task, and downstream training protocol. These findings motivate evaluating backbone-informed adapters against capacity-matched alternatives.
Nov 4, 2026

MATA: Motion-aligned Temporal Adapter

Few-shot fine-grained video action recognition requires recognizing novel action classes from only a few labeled examples. The difficulty is that visually similar actions often share actors, scenes, poses, and objects, differing only in the order, timing, and short duration of motion changes. We study this problem under three practical constraints: few tunable parameters, low compute cost, and use of the video modality alone. We propose Motion-aligned Temporal Adapter (MATA), a parameter-efficient adapter on top of a frozen CLIP backbone. MATA specializes CLIP features along the temporal dimension using two components: temporal phase alignment through learnable temporal positional embeddings, and a local motion transition operator through depth-wise convolution. With only ~3.6M tunable parameters, MATA achieves state-of-the-art performance across multiple fine-grained benchmarks while remaining robust to both visual corruptions and temporal disruptions induced by frame dropping. Code is available at https://github.com/prakashchhipa/MATA-Motion-aligned-Temporal-Adapter
Nov 3, 2026

Training Specialist LLMs on Your Phone: Forward-Only Optimization via Block-Spectral Coordinates

Large language models (LLMs) on personal devices keep user data local, but limited memory restricts their general capability. Since most user-specific tasks do not require the breadth of frontier models, a practical solution is to adapt a shared on-device model into lightweight specialists. However, backpropagation is prohibitive in memory, and users rarely provide labeled data. We present MobileSpec, a system that trains specialists with forward passes alone and learns from automatically checked task outcomes, as in reinforcement learning with verifiable rewards (RLVR). Because forward-only updates are built from random perturbations, training is efficient only when these perturbations resemble the weight changes of RLVR, which largely preserve the singular values of pretrained weights and rotate their singular vectors. The core algorithm of MobileSpec, FuRA-S, produces such perturbations while training only the block-spectral coordinates of pretrained weights, namely the singular values of their column blocks. As we show theoretically and empirically, rescaling these blocks independently shifts their relative contributions to each weight matrix and thereby rotates its singular vectors. Since this update can be rewritten as a transformation of the layer input, FuRA-S also adapts INT4 models without modifying their weights. With specialists of only 1-5 MB that update 0.015-0.05% of the parameters, MobileSpec raises the success rate of a phone-use agent from 6-23% up to 89-100% and trains a personalized 8B model on an iPhone 16 Pro overnight. On RLVR benchmarks, FuRA-S is competitive with full-parameter evolution strategies (ES) and outperforms ES over LoRA adapters and quantized ES on INT4 models.
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

Mamba3R: Fixed-Capacity Selective State Memory for Long-Sequence 3D Reconstruction

Streaming 3D reconstruction must maintain geometric quality over long inputs within a fixed GPU budget. A pretrained recurrent model provides bounded state, but freezing its backbone also fixes its learned state-observation interactions. We introduce Mamba3R, a jointly designed state-memory adapter that provides a trainable adaptation path without updating the backbone or task heads. A selective state-space mixer jointly transforms native state features and a fixed-capacity auxiliary memory, explicit read and write paths carry information across frames, and a decomposed feedback path regulates what the added memory feeds back. Only 1.28 million adapter parameters are trained. Under a shared periodic-reset protocol, it lowers trajectory error and improves accumulated surface and depth accuracy over recurrent baselines on public indoor and outdoor datasets and robot-recorded traversals, gaining up to 40 percent on one metric. Matched comparisons support both the choice of the mixing operator and the utility of carrying the auxiliary memory across frames. Allocated memory remains unchanged and throughput moves by at most 1.4 percent from one to two thousand frames. Replaying its recorded traversals offline, the platform's RTX 4060 averages 11.7 frames per second within 5.1 GiB with the same checkpoint, a mean throughput covering the 10 Hz sensor rate.
Nov 1, 2026

EEG-HealthLLM: Hierarchical Adaptation of Large Language Models for Unified Heterogeneous EEG Health Assessment

Electroencephalography (EEG) provides a non-invasive means for neurological and mental health assessment, yet existing EEG models are typically developed for individual tasks and are difficult to unify across heterogeneous signal configurations and assessment objectives. Large language models (LLMs) offer a common interface for multiple tasks, but adapting heterogeneous EEG signals to a shared LLM raises a key challenge: how to preserve task-specific signal characteristics while enabling effective language-side parameter sharing. To address this problem, we propose EEG-HealthLLM, a unified and parameter-efficient framework for heterogeneous EEG health assessment. We introduce an Asymmetric Hierarchical Adaptation Strategy: task-specific EEG adapters preserve heterogeneous spatio-temporal dynamics, while a unified shared LoRA is injected into the LLM backbone (e.g., Phi-3) to capture shared high-level assessment semantics across tasks. We further establish EEG-HealthBench by reorganizing five existing EEG resources into paradigms covering epileptic seizure detection, major depressive disorder assessment, anxiety risk assessment, depressive symptom risk assessment, and emotional state assessment under a unified subject-independent protocol. Extensive experiments show that EEG-HealthLLM achieves an average balanced accuracy (BAC) of 69.58%, ranking first among the evaluated methods in both average BAC and macro-F1. Compared with independent single-task LoRA adaptation, it improves average BAC from 67.70% to 69.58% while reducing the aggregate language-side adapter footprint by 80.0%. Further analyses reveal task-dependent cross-task interactions and trade-offs, supporting the proposed hierarchical adaptation strategy. The implementation is publicly available at https://github.com/Marui-88/EEG-HealthLLM.
Oct 31, 2026

Scute: Exact Batched Serving for PEFT Adapters Beyond LoRA

Existing serving systems for large language models batch requests for many different LoRA adapters on one GPU. However, these systems reject adapters from other parameter-efficient fine-tuning (PEFT) methods such as DoRA, (IA), VeRA and OFT. One solution to this is to convert these adapters to LoRA, but this induces a significant memory and accuracy cost. Instead, we show that, under some modifications, a batched LoRA kernel can serve all of these methods exactly and efficiently. We achieve this by writing each method in one normal form, a LoRA computation with six slots for an adapter's own vectors and matrices. Some slots can merge into an adapter's weights at load time; the others remain live and induce a cost. We give ways to efficiently pay this cost and predict it from a zero-parameter byte model. We package this system, Scute, and build it on top of vLLM. Scute serves all five methods, coming as checkpoints from the Hugging Face PEFT library, in one batch, and it matches that library with a mean error of at most 1.5% of each adapter's effect. Under a burst of requests on an RTX A6000, a mix of these methods takes 5-17% more decode time than the same number of LoRA adapters. Furthermore, using Scute, the same GPU can serve 64 to 1024 (IA) and VeRA adapters at 1.5x to at least 2.5x the request rate of their LoRA conversions.
Oct 30, 2026

ProLand: A New Infeasible Manifold Optimization Method for LoRA Fine-Tuning

Low-rank adaptation (LoRA) enables parameter-efficient fine-tuning but can suffer from redundant basis directions. Constraining the basis factor to the Stiefel manifold addresses this redundancy through orthonormality. Landing methods allow temporarily infeasible iterates to explore a neighborhood of the manifold while progressively restoring orthogonality without per-step retractions. To improve optimization efficiency while retaining this flexibility, we introduce ProLand (Product-space Landing), a retraction-free optimizer that combines coordinated factor updates with Muon-style spectral directions. By designing the updates in the space of the represented weight increment, ProLand separates objective optimization from orthogonality restoration, supporting efficient feasibility recovery. This coordinated construction also enables Muon-style directions to guide objective optimization while preserving that separation. Experiments cover four GLUE validation tasks on DeBERTa-v3-base and seven language benchmarks on Qwen2.5-3B, Llama-3.2-3B, and Moonlight-16B-A3B across multiple adapter ranks. Across these 11 tasks, ProLand achieves the highest average score among the compared methods in every tested encoder and decoder model-rank configuration.
Oct 30, 2026

CAPER: Circuit-Aware Parameter-Efficient Reasoning in LLMs for Analog Verification

Large language models (LLMs) are increasingly used to generate analog circuits, yet every netlist must also be verified. We propose LLMs as verification assistants for analog circuits that flag unmet specifications and reliability issues that appear under process, voltage, and temperature (PVT) variation. These issues emerge from how circuit stages combine, a structure that generic parameter-efficient fine-tuning (PEFT) methods such as LoRA do not capture. To this end, we introduce CAPER, a circuit-aware PEFT method that adds a small branch alongside LoRA to read each stage's properties from the LLM and combine them from the relevant stages. In deciding whether a feedback loop is negative, LoRA and a high-rank adapter stay at chance, whereas CAPER answers every test case correctly. For comparator latches, where 11 of 90 LLM-written designs that work under nominal conditions fail under PVT variation, CAPER likewise learns to flag such latent failures, while generic adapters stay near chance. These results show that adding circuit knowledge to PEFT makes such analog verification checks learnable through fine-tuning.
Oct 30, 2026

Breaking Information Islands in Sparse Tuning via Small-World Connectivity

Sparse tuning is widely used to adapt large language models due to its parameter efficiency. However, its effectiveness depends not only on the number of trainable parameters, but also on the feature-interaction topology of the sparse update's structural support. Limited or uneven connectivity in this support may isolate subsets of feature dimensions, forming information islands that restrict direct task-specific interactions through the trainable update. To address this issue, we propose Small-World Induced Fine-Tuning (SWIFT), a sparse tuning framework inspired by small-world connectivity. SWIFT combines block-diagonal sparse updates for local feature mixing with a fixed deterministic scattering permutation that routes the mixed outputs across feature groups without additional trainable parameters. We further show theoretically that certain random sparse-update models and block-diagonal updates can exhibit information islands, whereas a sufficiently scattering permutation guarantees positive block-level expansion when the trainable block-diagonal factor consists of equal-sized, structurally dense square blocks. Extensive experiments on commonsense reasoning, natural language generation, and image classification demonstrate that SWIFT outperforms the evaluated parameter-efficient fine-tuning (PEFT) baselines on average and can match or surpass full fine-tuning. Implementation is available at https://github.com/xianchaoguan/SWIFT