ICLR 2026 论文列表

The quadratic complexity of softmax attention presents a major obstacle for scaling Transformers to high-resolution vision tasks. Existing linear attention variants often replace the softmax with Gaussian kernels to reduce complexity, but such approximations lack theoretical grounding and tend to oversuppress mid-range token interactions. We propose LaplacianFormer, a Transformer variant that employs a Laplacian kernel as a principled alternative to softmax, motivated by empirical observations and theoretical analysis. To address expressiveness degradation under low-rank approximations, we introduce a provably injective feature map that retains fine-grained token information. For efficient computation, we adopt a Nyström approximation of the kernel matrix and solve the resulting system using Newton--Schulz iteration, avoiding costly matrix inversion and SVD. We further develop custom CUDA implementations for both the kernel and solver, enabling high-throughput forward and backward passes suitable for edge deployment. Experiments on ImageNet show that LaplacianFormer achieves strong performance-efficiency trade-offs while improving attention expressiveness. Our anonymous repository is at \href{https://anonymous.4open.science/r/sdfasfsdgsfgdrf}{\textcolor{black}{https://anonymous.4open.science/r/sdfasfsdgsfgdrf}}.

Transformers are widely used across data modalities, and yet the principles distilled from text models often transfer imperfectly. In this paper, we analyze Transformers through the lens of rank structure. Our focus is on the time series setting, where the structural properties of the data remarkably differ from those of text or vision. Time-series embeddings, unlike text or vision, exhibit sharply decaying singular spectra: small patch sizes and smooth continuous mappings concentrate the data into low-rank subspaces. From this, we prove that the associated $Q/K/V$ projections admit accurate low-rank approximations, and that attention layers become compressible in proportion to the decay of the embedding spectrum. We introduce the concept of *flow-of-ranks*, a mechanism by which nonlinear mixing across depth inflates the rank, explaining why early layers are most amenable to compression and why rank schedules should grow with depth. Guided by these results, we compress Chronos, a large time series foundation model, achieving a reduction of $65\\%$ in inference time and $81\\%$ in memory without loss of accuracy. These findings provide principled guidance for allocating width, depth, and heads in time series foundation models, and for exploiting their inherent compressibility.

We introduce GateSkip, a simple residual-stream gating mechanism that enables token-wise layer skipping in decoder-only LMs. Each Attention/MLP branch is equipped with a sigmoid-linear gate that compresses the branch’s output before it re-enters the residual stream. During inference we rank tokens by the gate and skip low-importance ones using a per-layer budget. While early-exit or router-based Mixture-of-Depths models are known to be unstable and need extensive retraining, our smooth, differentiable gates fine-tune stably on top of pretrained models. On long-form reasoning, we save up to 15% compute while retaining >90% of baseline accuracy. On instruction-tuned models we see accuracy gains at full compute and match baseline quality near 50% savings. The learned gates give insight into transformer information flow (e.g., BOS tokens act as anchors), and the method combines easily with quantization, pruning, and self-speculative decoding.

Fine-tuning pretrained models is a standard and effective workflow in modern machine learning. However, robust fine-tuning (RFT), which aims to simultaneously achieve adaptation to a downstream task and robustness to adversarial examples, remains challenging. Despite the abundance of non-robust pretrained models in open-source repositories, their potential for RFT is less understood. We address this knowledge gap by systematically examining RFT from such non-robust models. Our experiments reveal that fine-tuning non-robust models with a robust objective, even under small perturbations, can lead to poor performance, a phenomenon that we dub _suboptimal transfer_. In challenging scenarios (eg, difficult tasks, high perturbation), the resulting performance can be so low that it may be considered a transfer failure. We find that fine-tuning using a robust objective impedes task adaptation at the beginning of training and eventually prevents optimal transfer. However, we propose a novel heuristic, _Epsilon-Scheduling_, a schedule over perturbation strength used during training that promotes optimal transfer. Additionally, we introduce _expected robustness_, a metric that captures performance across a range of perturbations, providing a more comprehensive evaluation of the accuracy-robustness trade-off of diverse models at test-time. Extensive experiments on wide range of configurations (six pretrained models and five datasets) show that _Epsilon-Scheduling_ successfully prevents _suboptimal transfer_ and consistently improves expected robustness.

Long-sequence processing is a critical capability for modern large language models. However, the self-attention mechanism in the standard Transformer architecture faces severe computational and memory bottlenecks when processing long sequences. While trainable sparse attention methods offer a promising solution, existing approaches such as NSA introduce excessive extra parameters and disrupt the conventional pretrain-on-short, finetune-on-long workflow, resulting in slow convergence and difficulty in acceleration. To overcome these limitations, we introduce Dense-Sparse Switchable Attention framework (DSSA), a trainable sparse attention that seamlessly adapts models from short to long sequences. Specifically, DSSA reuses dense attention parameters through parameter-free architecture modification, maintaining consistency between short and long sequence processing. Additionally, DSSA ensures computational efficiency across all sequence lengths, by using dense attention for short inputs and smoothly transitioning to sparse attention for long sequences. To achieve practical acceleration, we further introduce an efficient implementation of DSSA that significantly reduces the computational overhead. Our experiments on long-context understanding and chain-of-thought reasoning demonstrate that DSSA is $4\times$ faster than dense attention while retaining 98.1% and 99.7% of the performance, respectively. We will release all associated implementations to facilitate future research on efficient attention.

Multivariate time series (MTS) anomaly diagnosis, which encompasses both anomaly detection and localization, is critical for the safety and reliability of complex, large-scale real-world systems. The vast majority of existing anomaly diagnosis methods offer limited theoretical insights, especially for anomaly localization, which is a vital but largely unexplored area. The aim of this contribution is to study the learning process of a Transformer when applied to MTS by revealing connections to statistical time series methods. Based on these theoretical insights, we propose the Attention Low-Rank Transformer (ALoRa-T) model, which applies low-rank regularization to self-attention, and we introduce the Attention Low-Rank score, effectively capturing the temporal characteristics of anomalies. Finally, to enable anomaly localization, we propose the ALoRa-Loc method, a novel approach that associates anomalies to specific variables by quantifying interrelationships among time series. Extensive experiments and real data analysis, show that the proposed methodology significantly outperforms state-of-the-art methods in both detection and localization tasks.

Continual Learning (CL) requires a model to learn multiple tasks in sequence while maintaining both stability—preserving knowledge from previously learned tasks, and plasticity—effectively learning new tasks. Orthogonal projection has emerged as an effective and popular paradigm in CL, where it partitions the gradient space of previously learned tasks into two orthogonal subspaces: a primary subspace and a minor subspace. New tasks are learned effectively within the minor subspace, thereby reducing interference with previously acquired knowledge. However, existing orthogonal projection methods struggle to achieve an optimal balance between plasticity and stability, as it is hard to appropriately partition the gradient space. In this work, we consider a continual learning paradigm based on Low-Rank Adaptation (LoRA), which has gained considerable attention due to its efficiency and wide applicability, and propose a novel approach for continual learning, called SplitLoRA. We first provide a theoretical analysis of how subspace partitioning affects model stability and plasticity. Informed by this analysis, we then introduce an effective method that derives the optimal partition of the gradient space for previously learned tasks. This approach effectively balances stability and plasticity in continual learning. Experimental results on multiple datasets demonstrate that the proposed method achieves state-of-the-art performance. The code is available at https://anonymous.4open.science/r/SplitLoRA-FB45.

Pre-trained vision language models have shown remarkable performance on visual recognition tasks, but they typically assume the availability of complete multimodal inputs during both training and inference. In real-world scenarios, however, modalities may be missing due to privacy constraints, collection difficulties, or resource limitations. While previous approaches have addressed this challenge using prompt learning techniques, they fail to capture the cross-modal relationships necessary for effective multimodal visual recognition and suffer from inevitable computational overhead. In this paper, we introduce MoRA, a parameter-efficient fine-tuning method that explicitly models cross-modal interactions while maintaining modality-specific adaptations. MoRA introduces modality-common parameters between text and vision encoders, enabling bidirectional knowledge transfer. Additionally, combined with the modality-specific parameters, MoRA allows the backbone model to maintain inter-modality interaction and enable intra-modality flexibility. Extensive experiments on standard benchmarks demonstrate that MoRA achieves an average performance improvement in missing-modality scenarios by 5.24% and uses only 25.90% of the inference time compared to the SOTA method while requiring only 0.11% of trainable parameters compared to full fine-tuning. The code is available at https://anonymous.4open.science/r/mora-20667.

\begin{abstract} Low-rank and sparse composite approximation is a natural idea to compress Large Language Models (LLMs). However, such an idea faces two primary challenges that adversely affect the performance of existing methods. The first challenge relates to the interaction and cooperation between low-rank and sparse matrices, while the second involves determining weight allocation across different layers, as redundancy varies considerably among them. To address these challenges, we propose a novel two-stage LLM compression method with the capability of global resource allocation for rank and sparsity. It is noteworthy that the overall optimization space is vast, making comprehensive optimization computationally prohibitive. Therefore, to reduce the optimization space, our first stage utilizes robust principal component analysis to decompose the weight matrices of LLMs into low-rank and sparse components, which span the low dimensional and sparse spaces containing the resultant low-rank and sparse matrices, respectively. In the second stage, we propose a probabilistic global allocation strategy to jointly identify the low-rank and sparse structures within the above two spaces. The appealing feature of our approach is its ability to automatically detect the redundancy across different layers and to manage the interaction between the sparse and low-rank components. Extensive experimental results indicate that our method significantly surpasses state-of-the-art techniques for sparsification and composite approximation.

A systematic, comparative investigation into the effects of low-quality data reveals a stark spectrum of robustness across modern probabilistic models. We find that autoregressive language models, from token prediction to sequence-to-sequence tasks, are remarkably resilient (for GPT-2, test NLL increases modestly from 2.87 to 3.59 despite 50\% token corruption). By contrast, under the same levels of data corruption, class-conditional diffusion models degrade catastrophically (image-label consistency plummets by 56.81\% relative to baseline), while classifiers show a moderate impact that diminishes with dataset scale. To explain these discrepancies, we analyze the results through a multi-perspective lens, integrating information theory, PAC learning, and gradient dynamics. \textcolor{blue}{These analyses suggest that robustness is heavily influenced by two key principles}: the \textbf{richness of conditioning information}, which constrains the learning problem, and the \textbf{absolute information content} of the training data, which allows the signal from correct information to dominate statistical noise.

The autoregressive nature of large language models (LLMs) fundamentally limits inference speed, as each forward pass generates only a single token and is often bottlenecked by memory bandwidth. Speculative decoding has emerged as a promising solution, adopting a draft-then-verify strategy to accelerate token generation. While the EAGLE series achieves strong acceleration, its requirement of training a separate draft head for each target model introduces substantial adaptation costs. In this work, we propose \textbf{PARD (PARallel Draft)}, a novel speculative decoding method featuring \textit{target-independence} and \textit{parallel token prediction}. Specifically, PARD enables a single draft model to be applied across an entire family of target models without requiring separate training for each variant, thereby minimizing adaptation costs. Meanwhile, PARD substantially accelerates inference by predicting multiple future tokens within a single forward pass of the draft phase. To further reduce the training adaptation cost of PARD, we propose a COnditional Drop-token (COD) mechanism based on the integrity of prefix key-value states, enabling autoregressive draft models to be adapted into parallel draft models at low-cost. Our experiments show that the proposed COD method improves draft model training efficiency by \textbf{3$\times$} compared with traditional masked prediction training. On the \texttt{vLLM} inference framework, PARD achieves up to \textbf{3.67$\times$} speedup on LLaMA3.1-8B, reaching \textbf{264.88} tokens per second, which is \textbf{1.15$\times$} faster than EAGLE-3.

This work presents a novel, fully Riemannian framework for Low-Rank Adaptation (LoRA) that geometrically treats low-rank adapters by optimizing them directly on the fixed-rank manifold. This formulation eliminates the parametrization ambiguity present in standard Euclidean optimizers. Our framework integrates three key components to achieve this: (1) we derive **Riemannion**, a new Riemannian optimizer on the fixed-rank matrix manifold that generalizes the recently proposed Muon optimizer; (2) we develop a Riemannian gradient-informed LoRA initialization, and (3) we provide an efficient implementation without prominent overhead that uses automatic differentiation to compute arising geometric operations while adhering to best practices in numerical linear algebra. Comprehensive experimental results on both LLM and diffusion model architectures demonstrate that our approach yields consistent and noticeable improvements in convergence speed and final task performance over both standard LoRA and its state-of-the-art modifications.

Continual Instruction Tuning (CIT) enables Large Multimodal Models (LMMs) to rapidly adapt to new tasks without retraining, but it suffers from the catastrophic forgetting problem. By adding new branches, model extension provides a great idea to accommodate novel knowledge while causing huge memory consumption. To jointly address forgetting and memory explosion, we propose the Compression–Integration–Learning (CIL) pipeline, which draws on the memory consolidation processes during human sleep. Compression streamlines old parameters to release capacity. Integration merges knowledge from similar tasks to restore the performance loss due to compression. For example, based on LLaVA-7B, the forgetting is reduced from 11.29 to 5.09. Learning reallocates released capacity for new task-relevant parameters. Next, based on the characteristics of LMMs at different learning stages, we establish the progressive learning process, further reducing forgetting from 5.09 to 3.39. Moreover, to adapt this process, we decompose LoRA into a set of rank vectors and introduce an extremely fine-grained architecture, LoRA Rank Pool (LRP), with the goal of flexible knowledge employment and editing. Finally, we combine all components, and yield **P**rogressively **C**ompressed **L**o**R**A (PCLR). Extensive experiments demonstrate that PCLR owns a memory budget close to non-extension methods while outperforming extension methods in performance. The implementation code is available at https://github.com/SII-HITclearlove777/PCLR.

Test-time prompt tuning (TPT) has emerged as a promising technique for adapting large vision-language models (VLMs) to unseen tasks without relying on labeled data. However, the lack of dispersion between textual features can hurt calibration performance, which raises concerns about VLMs' reliability, trustworthiness, and safety. Current TPT approaches primarily focus on improving prompt calibration by either maximizing average textual feature dispersion or enforcing orthogonality constraints to encourage angular separation. However, these methods may not always have optimal angular separation between class-wise textual features, which implies overlooking the critical role of angular diversity. To address this, we propose A-TPT, a novel TPT framework that introduces angular diversity to encourage uniformity in the distribution of normalized textual features induced by corresponding learnable prompts. This uniformity is achieved by maximizing the minimum pairwise angular distance between features on the unit hypersphere. We show that our approach consistently surpasses state-of-the-art TPT methods in reducing the aggregate average calibration error while maintaining comparable accuracy through extensive experiments with various backbones on different datasets. Notably, our approach exhibits superior zero-shot calibration performance on natural distribution shifts and generalizes well to medical datasets. We provide extensive analyses, including theoretical aspects, to establish the grounding of A-TPT. These results highlight the potency of promoting angular diversity to achieve well-dispersed textual features, significantly improving VLM calibration during test-time adaptation. Our code will be made publicly available.

Low-rank architectures have become increasingly important for efficient large language model (LLM) pre-training, providing substantial reductions in both parameter complexity and memory/computational demands. Despite these advantages, current low-rank methods face three critical shortcomings: (1) compromised model performance, (2) considerable computational overhead, and (3) limited activation memory savings. To address these limitations, we propose **C**ross-layer Low-**R**ank residual **Net**work (**CR-Net**), an innovative parameter-efficient framework inspired by our discovery that inter-layer activation residuals possess low-rank properties. CR-Net implements this insight through a dual-path architecture that efficiently reconstructs layer activations by combining previous-layer outputs with their low-rank differences, thereby maintaining high-rank information with minimal parameters. We further develop a specialized activation recomputation strategy tailored for CR-Net that dramatically reduces memory requirements. Extensive pre-training experiments across model scales from 60M to 7B parameters demonstrate that CR-Net consistently outperforms state-of-the-art low-rank frameworks while requiring fewer computational resources and less memory.

Generalization in EEG-based motor imagery (MI) brain-computer interfaces (BCIs) is hampered by cross-subject and cross-session variability. Although large-scale EEG pretraining has advanced representation learning, their practical deployment is hindered by the need for costly fine-tuning to overcome significant domain shifts. Test-time adaptation (TTA) methods that adapt models during inference offer a promising solution. However, existing EEG-TTA methods either rely on gradient-based fine-tuning (suffering from high computational cost and catastrophic forgetting) or data alignment strategies (failing to capture shifts in temporal predictive embeddings). To address these limitations, we propose BTTA-DG, a novel Bayesian Test-Time Adaptation framework that performs efficient, gradient-free adaptation by modeling the distribution of temporal predictive embeddings. Our approach first employs a lightweight SincAdaptNet with learnable filters to extract task-specific frequency bands. We then introduce a novel Dirichlet feature projection that maps temporal embeddings onto a compact and interpretable parameter space, effectively capturing the concentration of time-varying predictive evidence. Adaptation is achieved via a GMM-driven Bayesian inference mechanism, which models the historical distribution of these Dirichlet parameters and fuses this evidence with the model's prior predictions to calibrate outputs for the target domain. Extensive experiments show that BTTA‑DG significantly outperforms previous EEG‑TTA methods, achieving state‑of‑the‑art accuracy while running at real‑time speed. Furthermore, visualizations confirm the physiological interpretability of our learned filters and the robust class separability of our Dirichlet feature space.

Efficient deployment of deep neural networks increasingly relies on Post-Training Quantization (PTQ). Logarithmic PTQ, in particular, promises multiplier-free hardware efficiency, but its performance is often limited by the nonlinear and symmetric quantization grid and standard rounding-to-nearest (RTN) approach. While learnable rounding has significantly advanced linear PTQ, its application to the non-linear and often discrete nature of logarithmic domain remains unexplored. This paper introduces learnable Logarithmic Adaptive Rounding Techniques (LogART) that pioneer task-aware learnable rounding specifically for the logarithmic domain. LogART further extends the learnable rounding strategy to flexibly support outlier-aware, asymmetric, and hardware-friendly dynamic logarithmic bases, determined in a distribution-aware manner using an efficient search strategy. Extensive experiments demonstrate that LogART achieves state-of-the-art accuracy while maintaining efficiency in quantizing models across various architectures and ultra-low bitwidths, outperforming existing logarithmic PTQ methods and paving the way for more effective hardware deployment. The code is available at https://github.com/logart-lab/logart.

Video generative models achieve high-quality synthesis from natural-language prompts by leveraging large-scale web data. However, this training paradigm inherently exposes them to unsafe biases and harmful concepts, introducing the risk of generating undesirable or illicit content. To mitigate unsafe generations, existing machine unlearning approaches either rely on filtering, and can therefore be bypassed, or they update model weights, but with costly fine-tuning or training-free closed-form edits. We propose the first training-free weight update framework for concept removal in video diffusion models. From five paired safe/unsafe prompts, our method estimates a refusal vector and integrates it into the model weights as a closed-form update. A contrastive low-rank factorization further disentangles the target concept from unrelated semantics, it ensures a selective concept suppression and it does not harm generation quality. Our approach reduces unsafe generations on the Open-Sora and ZeroScopeT2V models across the T2VSafetyBench and SafeSora benchmarks, with average reductions of 36.3% and 58.2% respectively, while preserving prompt alignment and video quality. This establishes an efficient and scalable solution for safe video generation without retraining nor any inference overhead.

Low-rank optimization has emerged as a promising direction in training large language models (LLMs) to improve running time and reduce the memory usage of adaptive optimizers by constraining learning to a lower-dimensional space. Prior work typically projects gradients of linear layers using approaches based on Singular Value Decomposition (SVD) or QR-decomposition. Applying these techniques individually to each layer in large models is computationally expensive and incurs additional memory costs due to storing the projection matrices. In this work, we propose a computationally efficient and conceptually simple, two-step procedure to approximate SVD/QR-based gradient projections into lower-dimensional spaces by using a predefined orthogonal matrix of the Discrete Cosine Transform (DCT). We dynamically select columns from the DCT matrix based on their alignment with the gradient of each layer. The effective projection matrices are obtained via a simple matmul with the DCT matrix in $O(n^3)$ time, followed by a lightweight sorting step to identify the most relevant basis vectors. For large layers, DCT can be computed via Makhoul's $N$-point algorithm based on Fast Fourier Transform (FFT) in $O(n^2 \log(n))$ time. Due to the predefined nature of the orthogonal bases, they are computed once at the start of training. Our numerical experiments on both pre-training and fine-tuning tasks demonstrate the effectiveness of our dual strategy in approximating optimal low-rank projections, obtaining an approach with rank-independent running time that matches the performance of costly SVD/QR-based methods while achieving faster runtime and reduced memory usage by up to $25\\%$ across different model sizes.