NeurIPS 2026 论文列表

Low-Rank Adaptation (LoRA) has become a widely adopted technique in text-to-image diffusion models, enabling the personalisation of visual concepts such as characters, styles, and objects. However, existing approaches struggle to effectively compose multiple LoRA adapters, particularly in open-ended settings where the number and nature of required skills are not known in advance. In this work, we present LoRAtorio, a novel train-free framework for multi-LoRA composition that gates each adapter using only signals already inside the denoiser, specifically, patch-level cosine (dis)similarity between each LoRA's predicted noise and the base model's. We refer to this as ``intrinsic'' guidance because no external parameters are introduced. Our method is motivated by two key observations: (1) LoRA adapters trained on narrow domains produce unconditioned denoised outputs that diverge from the base model, and (2) when conditioned out of distribution, LoRA outputs show behaviour closer to the base model than when conditioned in distribution. These patch-level similarities are used to construct a spatially-aware weight matrix, which guides a weighted aggregation of LoRA outputs. To address domain drift, we further propose a modification to classifier-free guidance that incorporates the base model's unconditional score into the composition. We extend this formulation to a dynamic module selection setting, enabling inference-time selection of relevant LoRA adapters from a large pool. On the ComposLoRA benchmark, LoRAtorio achieves state-of-the-art performance, with a CLIPScore that does not deteriorate as the number of composed adapters grows, surpassing the strongest prior method by up to 1.3% in CLIPScore, and reaching a 76.92% GPT-4V win rate against the closest baseline. We further show that the framework is backbone-agnostic: the intrinsic-similarity principle transfers to rectified-flow architectures (FLUX.1-dev) without retraining. Code will be made available.

Low-rank adaptation (LoRA) attaches small low-rank matrices to a frozen backbone, and public repositories like Hugging Face host many such LoRAs per backbone, each for a different concept, character, style, or task. Users routinely want to combine several LoRAs in a single generation, but two existing strategies involve a trade-off with each other. First, combining the LoRA weights into a single update before inference (e.g., static merge) needs only one forward pass, but quality drops, because LoRAs trained for different concepts interfere when their weights are summed into a single update, distorting each LoRA's contribution. Second, running each LoRA in a separate forward pass and summing the outputs (e.g., composite) preserves output quality, but each LoRA adds a full backbone pass, so inference time scales linearly with the active LoRA count. To recover composite-level quality at merge cost, we study a property of LoRA pools overlooked by prior multi-LoRA work. When LoRAs in a pool are trained on related concepts, their low-rank factors overlap substantially. Using singular value decomposition, we find the minimal subspace that preserves 99% of the factor energy. On three diffusion pools, only 46% to 57% of the sum of individual LoRA ranks is enough to represent the pool, and the subspace grows sub-linearly with pool size. A 48-LoRA sub-pool of LoRA-Hub on FLAN-T5 NLP tasks shows comparable overlap (about 62%), while the diversity-curated LoRA-Retriever benchmark sits near the no-overlap baseline at 96%, indicating that across these five pools pool composition rather than modality predicts where overlap is large. We use this overlap to construct LoRASpace, a training-free pool-wide method where the subspace matrices are computed once per pool and reused across two inference modes: a static mode at merge inference cost, and a dynamic mode for input-dependent routing at substantially lower cost than composite. On the diffusion pools, static LoRASpace matches composite quality at merge cost, and dynamic LoRASpace achieves more than 6× speedup over composite at six active LoRAs on a 17-LoRA pool. In our ablation, we show that the quality gain comes from truncating to this low-rank shared subspace, which acts as a form of regularisation. On the near-orthogonal LoRA-Retriever benchmark, the framework predicts limited compression benefit, with gains coming from external retrieval coupling.

On-policy distillation (OPD) has emerged as an efficient post-training paradigm for large language models. However, existing studies largely attribute this advantage to denser and more stable supervision, while the parameter-level mechanisms underlying OPD's efficiency remain insufficiently understood. In this work, we argue that OPD's efficiency stems from a form of ``foresight'': it establishes a direct and stable path from the initial model to the final model early in training. This foresight manifests in two aspects. First, at the \textbf{Module-Allocation Level}, OPD identifies regions with low marginal utility and concentrates updates on modules that are more critical to reasoning. Second, at the \textbf{Update-Direction Level}, OPD exhibits stronger low-rank concentration, with its dominant subspace aligning closely with the final update subspace early in training. Motivated by this theory, we propose \textbf{EffOPD}, a plug-and-play acceleration method that speeds up OPD by searching for an effective step size and extrapolating along the current update direction. EffOPD requires no additional trainable modules or complex hyperparameter tuning, and achieves up to $3.5\times$ training acceleration while maintaining comparable final performance. Overall, our findings provide a parameter-dynamics perspective for understanding the efficiency of OPD and offer practical insights for designing more efficient post-training methods for large language models. Our code is available at: https://anonymous.4open.science/r/EffOPD-7C58.

We study the sequential decision-making problem where at each time the learner selects an assortment of bounded size and receives a ranking of the selected items. Ranking feedback arises naturally when human evaluators or automated LLM judges compare multiple options simultaneously. It is often easier and more reliable than assigning absolute scores, which are typically poorly calibrated across evaluators, while yielding substantially more information than a single pairwise comparison. We focus on rankings that are constructed according to the random utility model with linear utilities, which includes the popular Plackett-Luce model as a special case. We show that improved regret minimisation is possible under full rank breaking with a refined analysis based on 1-factorisations and Baranyai's theorem that fully exploits independence in the pairwise comparisons. This gives a tighter confidence interval around the true utility parameter which is crucial in our analysis for regret that improves with maximum assortment size. Furthermore, it matches the regret lower bound for our model in case the learner is permitted to play multi-sets of the maximum bounded size. Our work gives a principled approach to learning from ranking feedback that shows consistent improvement with the maximum assortment size.

Semantic visual classification commonly aligns visual embeddings with semantic spaces induced by pretrained text encoders. However, such spaces are typically high-dimensional, weakly aligned with visual feature geometry, and insufficiently discriminative for downstream classification tasks. We propose TaSS, a framework for learning a task-adaptive low-dimensional semantic space that preserves semantic structure while explicitly optimizing class separability. TaSS is constructed through a parameterized geometry transformation of pretrained text embeddings that enlarges inter-class margins via distance reshaping and improves intra-class compactness through prototype-preserving semantic constraints. The resulting semantic space maintains meaningful semantic relationships while producing a discriminative geometry tailored for classification. Frozen visual encoder features are subsequently aligned to TaSS using a lightweight MLP, enabling improved discriminative learning without fine-tuning pretrained backbones. We further provide theoretical analysis showing that TaSS preserves semantic consistency while enforcing enhanced inter-class separation and controlled intra-class variance. Extensive experiments on skeleton-based human action recognition, video classification, and image classification benchmarks demonstrate consistent improvements over existing semantic alignment approaches, with gains exceeding 3 percentage points on average and up to 8.5 points on challenging datasets.

Learning in neural systems arises from synaptic changes that reshape the representations underlying behavior. While low-rank recurrent neural networks (RNNs) have emerged as a powerful framework for linking connectivity to function, a theoretical understanding of their learning process remains elusive. Here, we extend the low-rank framework from activity to learning by deriving gradient-descent dynamics directly in a reduced overlap space. We formulate a closed-form, low-dimensional system of ODEs that governs learning in this space, exact for linear RNNs and asymptotically exact for nonlinear RNNs in the large-$N$ Gaussian limit. Central to our analysis is a distinction between two classes of overlaps: _loss-visible_ overlaps, which fully determine network activity, output, and loss, and _loss-invisible_ overlaps, which do not affect function but are required to describe learning. We illustrate the consequences of this decomposition through two phenomena. First, we show that learning can serve as a perturbation that exposes differences in connectivity between functionally equivalent networks. Second, we show that loss-invisible overlaps can act as memory variables that encode training history, and characterize the conditions under which this occurs. Finally, we present several testable predictions for biological learning experiments derived from our theory.

Low-Rank Adaptation (LoRA) is the prevailing approach for efficient large language model (LLM) fine-tuning. Building on this paradigm, recent studies have proposed alternative initialization strategies and architectural modifications, reporting substantial improvements over vanilla LoRA. However, these gains are often demonstrated under fixed or narrowly tuned hyperparameter settings, despite the known sensitivity of neural networks to training configurations. In this work, we systematically re-evaluate nine representative LoRA variants alongside vanilla LoRA through extensive hyperparameter searches. Across tasks spanning mathematical reasoning, commonsense reasoning, code generation, and instruction following at diverse model scales, we find that different LoRA methods favor distinct learning rate ranges. Crucially, once learning rates are properly tuned, all methods achieve similar peak performance (within 1-2\%), with only subtle rank-dependent behaviors. These results suggest that vanilla LoRA remains a competitive baseline and that improvements reported under a single training configuration may not reflect consistent methodological advantages. Finally, a second-order analysis attributes the differing optimal learning rate ranges to variations in the largest Hessian eigenvalue, aligning with classical learning theories.

In many modern applications of reinforcement learning (RL), the natural reward for a task of interest is inherently sparse: a reward of 0 is given everywhere except when the task is completed, when a reward of 1 is given. Training a policy to maximize such a sparse reward requires solving a challenging credit assignment problem, leading to slow or ineffective RL improvement. We propose a simple approach to transform a sparse outcome reward into a dense process reward. Our approach relies on training a discriminator to distinguish between previous successful and unsuccessful trajectories, and using this discriminator to incentivize the RL-learned policy to match the state-action visitations of successful trajectories, while avoiding those of unsuccessful trajectories. By incentivizing the policy to match the visitations over all states, not just those that correspond to task success, this reward provides dense feedback on whether progress is being made towards task completion, and, we show, provably achieves this without changing the optimal policy. Focusing on finetuning of robotic control policies, we demonstrate that our approach leads to significantly faster RL finetuning performance on both simulated and real-world manipulation tasks, as compared to simply maximizing the sparse outcome reward.

Softmax attention in Transformers suffers from quadratic complexity in sequence length, making it impractical for long-context applications. Linear attention alleviates this issue by replacing the exponential kernel with alternative functions that enable linear-time computation. Among existing linear attention approaches, recent studies have shown that polynomial kernels are particularly effective, as they exhibit sparse and spiky behavior similar to softmax attention, which emphasizes large dot products while suppressing irrelevant interactions. However, the exact computation of high-degree polynomial kernels is infeasible for high-dimensional representations. As a result, prior work relies on approximate polynomial kernels. This introduces a non-negligible approximation error. In this paper, we show that, from the perspective of polynomial kernel approximation, existing linear attention methods are still suboptimal. We propose Learnable Low-Rank Polynomial Sketch (LLoPS), a principled and flexible framework for approximating polynomial kernels with linear attention. Our method learns a low-rank polynomial sketch that provably achieves a strictly smaller approximation error than existing approaches. Experiment results show that LLoPS achieves the highest performance across extensive benchmarks, comparing with various linear attention baselines.

Domain adaptation faces a fundamental paradox in the cold-start regime. When target data is scarce, statistical methods fail to distinguish relevant source domains from irrelevant ones, which often leads to negative transfer. In this paper, we address this challenge by leveraging expert textual descriptions of the target domain, a resource that is often available but overlooked. We propose a probabilistic framework that translates these semantic descriptions into a choice model, namely a Language-Induced Prior (LIP), that learns the preferences from a pre-trained Large Language Model (LLM). The LIP is then integrated into an Expectation-Maximization (EM) algorithm to identify source relevance. Methodologically, this framework is compatible with any parametric model where a likelihood is available. It allows the LIP to guide the selection of sources when target signals are weak, while gradually refining these choices as samples accumulate. Theoretically, we prove that the estimator roughly matches an oracle cold-start MSE under a correct prior, while remaining asymptotically consistent regardless of the quality of the LIP. Empirically, we validated the framework on a descriptive (Gaussian estimation), a predictive (C-MAPSS dataset), and a prescriptive task (MuJoCo Hopper).

Vision-language models (VLMs) deliver strong multimodal reasoning capabilities, but their large computational cost and high parameter counts make deployment challenging on resource-constrained devices. Low-rank decomposition has emerged as a promising compression technique, yet existing methods often optimize local matrix reconstruction error, rely on uniform or heuristic rank allocation, and focus mainly on attention projections while leaving feed-forward networks underexplored. In this paper, we propose~\textit{LSVD}, a loss-aware low-rank approximation framework for efficient low-precision VLMs. LSVD derives a curvature-weighted SVD objective from a second-order approximation of the model loss and uses Kronecker-factored Fisher information to guide decomposition toward downstream performance rather than reconstruction alone. We further introduce a loss-aware cross-layer rank allocation strategy based on calibration gradients, enabling more effective parameter budgeting across layers. Finally, we extend low-rank compression to FFN layers through a hybrid scheme that combines SVD with quantization. The evaluation results show that LSVD achieves over $2.3\times$ decoding speedup over previous work while preserving strong accuracy under low-precision inference.

Given the inherent unpredictability of packet loss in vehicular wireless communications, V2X collaborative perception can yield practical benefits only if agents can achieve reliable collaboration under lossy and low-bandwidth communication conditions. Existing dense BEV feature fusion methods depend on redundant BEV feature exchange, which is infeasible in low-bandwidth scenarios, while compact-communication methods aggressively compress messages but can hardly recover the missing feature content after packet loss. In this paper, we present LR-V2X, a loss-resilient, latent-space reconstruction framework that converts corrupted received latents (even under severe 90\% packet loss) into a spatial prior and then reconstructs the missing BEV information from this informative prior and using ego context as condition. Notably, the model can be trained under complete communication conditions and can be directly applied to lossy conditions at test time, eliminating the need for training under numerous lossy conditions. Experiments on DAIR-V2X and V2XREAL show that LR-V2X delivers the strongest robustness under severe packet loss and preserves reliable collaboration as communication quality degrades. And it reduces communication overhead by $64\times$ compared to dense BEV feature fusion baselines. Code will be publicly released.

Orthogonal parameter-efficient fine-tuning adapts pretrained weights through structure-preserving multiplicative transformations, but existing methods often conflate two distinct design choices: the subspace in which adaptation occurs and the transformation applied within that subspace. This paper introduces LOFT, a low-rank orthogonal fine-tuning framework that explicitly separates these two components. By viewing orthogonal adaptation as a multiplicative subspace rotation, LOFT provides a unified formulation that recovers representative orthogonal PEFT methods, including coordinate-, butterfly-, Householder-, and principal-subspace-based variants. More importantly, this perspective exposes support selection as a central design axis rather than a byproduct of a particular parameterization. We develop a first-order analysis showing that useful adaptation supports should be informed by the downstream training signal, motivating practical gradient-informed support selection strategies. Across language understanding, visual transfer, mathematical reasoning, and multilingual out-of-distribution adaptation, LOFT recovers principal-subspace orthogonal adaptation while gradient-informed supports improve the efficiency–performance trade-off under matched parameter, memory, and transform budgets. These results suggest that principled support selection is an important direction for improving orthogonal PEFT.

Distilled diffusion models accelerate image generation by reducing the number of denoising steps, but often suffer from degraded image quality. To mitigate this trade-off, test-time optimization methods improve quality, yet their iterative nature incurs substantial computational overhead and leads to slow inference, limiting practical usability. Recent hypernetwork-based approaches amortize this process during training, but still require costly noise modulation in high-dimensional latent spaces. In this work, we propose LENS (Low-frequency Eigen Noise Shaping), an efficient noise modulation framework that operates in a low-dimensional subspace. Our approach is motivated by the observation that low-frequency components of the noise largely determine the global structure and visual fidelity of generated images. Based on this observation, we provide a theoretical justification for restricting modulation to the low-frequency subspace and derive a principled training objective. Building on this, LENS employs a lightweight, standalone network to selectively modulate these components, enabling efficient and targeted noise modulation. Extensive experiments demonstrate that LENS achieves competitive image quality while reducing FLOPs by 400–700$\times$, model parameters by 25–75$\times$, and inference-time overhead by 10–20$\times$ compared to prior methods.

We consider the problem of learning to adapt a foundation model in a federated setting, particularly the most realistic and general setting: 1) When the local data sets are sampled from different distributions but we want to learn a globally adapted model, 2) Where local agents enter and leave the federation asynchronously at each time tick which is beyond the control of the learning algorithm, and 3) Where the goal is continuous adaptation so that after each time tick, the learned adapter generalizes accurately for all participants that have been seen during training. We propose a simple idea called federated *library-based adaptation* (*LEAN*) for this setting. In library-based adaptation, the system maintains a pool, or "library" of rank-1 "basis pairs". Groups of these pairs are distributed to local sites to be trained in LoRA fashion, and then separated and *shuffled* at communication. Library-based adaptation is designed to avoid problems with more conventional methods, such as those based on averaging. In particular, we demonstrate *LEAN* outperforms traditional averaging baselines in both communication and computation cost efficiency across a broad range of important settings, including heavy data skewness and high asynchrony.

Electroencephalogram (EEG) classification plays a key role in medical diagnosis and brain–computer interfaces, but remains challenging due to low signal-to-noise ratios and high inter-subject variability. As a result, many existing approaches rely on subject-specific models, which fail to exploit shared structure in neural signals and do not generalize to unseen subjects. To address these limitations, we propose LAtte, a framework that combines Lorentz attention with a hyperbolic InceptionTime-based encoder to improve cross-subject generalization in EEG classification. The model explicitly decomposes EEG signals into a learned baseline component and task-relevant deviations, enabling more structured representation learning. To further improve robustness and adaptability, we incorporate subject-specific low-rank adaptation (LoRA) modules at both encoder and decoder levels, augmented with a Lorentz boost–based LoRA mechanism and hyperbolic projection layers to reduce overfitting in geometric representations. We evaluate LAtte with and without finetuning in three settings: subject-specific, subject-conditional, and leave-one-subject-out (LOSO) on five established EEG datasets, achieving a consistent improvement in performance over current state-of-the-art methods for smaller datasets and maintaining performance for larger datasets.

Fast linear algebra in deep learning usually comes with a choice: fixed geometry and exact computation, as in the Fourier transform, or adaptive geometry paid for by dense parameters, random features, or low-rank surrogates. To move beyond this trade-off, we introduce LAPLEX, a class of exact, trainable (phased) Laplace-kernel operators. A LAPLEX layer is a typically full-rank dense matrix, implicitly defined by learnable coordinate anchors, with FFT-like scaling. Consequently, it supports trainable matrix--vector operations at vector dimensions up to $10^9$ on modern GPUs. As a neural layer, it yields compact projections and classification heads interpretable as soft, trainable routing models. The same primitive also serves as an efficient Gram operator, enabling high-dimensional covariance models on flattened images of dimension $3 \cdot 10^6$ that preserve visible spatial structure without imposing convolutional bias. These applications reflect a single principle: dense geometry can be learned without storing a dense matrix, which enables data-adaptive global interactions in regimes where ordinary dense layers are out of reach. In this sense, LAPLEX separates expressivity from storage cost: it behaves like a dense trainable matrix, but is represented and applied through a small structured set of parameters.

Mixed-precision computations are a hallmark of the current stage of AI, driving the progress in large language models towards efficient, locally deployable solutions. This article addresses the floating-point computation of compositionally-rich functions, concentrating on transformer inference. Based on the rounding error analysis of a composition $f(g(x))$, we provide an adaptive strategy that selects a small subset of components of $g(x)$ to be computed more accurately while all other computations can be carried out with lower accuracy. We then explain how this strategy can be applied to different compositions within a transformer and illustrate its overall effect on transformer inference. We study the effectiveness of this algorithm numerically on GPT-2 models and demonstrate that already very low recomputation rates allow for improvements of up to two orders of magnitude in accuracy.

At deployment, a multivariate forecaster may expose only a recent input window and a frozen point prediction, with no access to ground truth, calibration residuals, repeated inference, or model internals. We study deployment-constrained forecast risk localization: ranking future timesteps by likely error before they are observed. We introduce SubDx, a zero-parameter diagnostic that audits the forecast against the local cross-channel subspace of the input history. Each predicted channel vector receives an off-subspace residual score, computed with no labels, no learned parameters, and less than 1% runtime overhead on Traffic. Under a linear factor model, we derive closed-form score distributions and a two-population AUROC expression showing how heterogeneous off-subspace error visibility makes the ranking identifiable. On Traffic (N=862), SubDx reaches within-horizon AUROC 0.90 and reduces MSE by 62% at 50% coverage; on Solar (N=137), it reaches AUROC 0.85 with 67% MSE reduction. The same post-hoc score applies to frozen pretrained forecasters, with Chronos reaching AUROC 0.838 on Traffic without fine-tuning. Across trained backbones, pretrained models, and synthetic controls, the results show that local cross-channel consistency can provide a practical abstention signal for redundant multivariate systems.