Attention Mechanisms

Partial differential equations (PDEs) are fundamental for modeling complex physical systems, yet classical numerical solvers face prohibitive computational costs in high-dimensional and multi-scale regimes. While Transformer-based neural operators have emerged as powerful data-driven alternatives, they conventionally treat all discretized spatial points as uniform, independent tokens. This monolithic approach ignores the intrinsic scale separation of physical fields, applying computationally prohibitive global attention that redundantly mixes smooth large-scale dynamics with high-frequency fluctuations. Rethinking Transformers through the lens of complex dynamics, we propose DynFormer, a novel dynamics-informed neural operator. Rather than applying a uniform attention mechanism across all scales, DynFormer explicitly assigns specialized network modules to distinct physical scales. It leverages a Spectral Embedding to isolate low-frequency modes, enabling a Kronecker-structured attention mechanism to efficiently capture large-scale global interactions with reduced complexity. Concurrently, we introduce a Local-Global-Mixing transformation. This module utilizes nonlinear multiplicative frequency mixing to implicitly reconstruct the small-scale, fast-varying turbulent cascades that are slaved to the macroscopic state, without incurring the cost of global attention. Integrating these modules into a hybrid evolutionary architecture ensures robust long-term temporal stability. Extensive memory-aligned evaluations across four PDE benchmarks demonstrate that DynFormer achieves up to a 95% reduction in relative error compared to state-of-the-art baselines, while significantly reducing GPU memory consumption. Our results establish that embedding first-principles physical dynamics into Transformer architectures yields a highly scalable, theoretically grounded blueprint for PDE surrogate modeling.

Anticipating future actions in videos is challenging, as the observed frames provide only evidence of past activities, requiring the inference of latent intentions to predict upcoming actions. Existing transformer-based approaches, which rely on dot-product attention over pixel representations, often lack the high-level semantics necessary to model video sequences for effective action anticipation. As a result, these methods tend to overfit to explicit visual cues present in the past frames, limiting their ability to capture underlying intentions and degrading generalization to unseen samples. To address this, we propose Action-Guided Attention (AGA), an attention mechanism that explicitly leverages predicted action sequences as queries and keys to guide sequence modeling. Our approach fosters the attention module to emphasize relevant moments from the past based on the upcoming activity and combine this information with the current frame embedding via a dedicated gating function. The design of AGA enables post-training analysis of the knowledge discovered from the training set. Experiments on the widely adopted EPIC-Kitchens-100 benchmark demonstrate that AGA generalizes well from validation to unseen test sets. Post-training analysis can further examine the action dependencies captured by the model and the counterfactual evidence it has internalized, offering transparent and interpretable insights into its anticipative predictions.

We propose Chunk-wise Attention Transducer (CHAT), a novel extension to RNN-T models that processes audio in fixed-size chunks while employing cross-attention within each chunk. This hybrid approach maintains RNN-T's streaming capability while introducing controlled flexibility for local alignment modeling. CHAT significantly reduces the temporal dimension that RNN-T must handle, yielding substantial efficiency improvements: up to 46.2% reduction in peak training memory, up to 1.36X faster training, and up to 1.69X faster inference. Alongside these efficiency gains, CHAT achieves consistent accuracy improvements over RNN-T across multiple languages and tasks -- up to 6.3% relative WER reduction for speech recognition and up to 18.0% BLEU improvement for speech translation. The method proves particularly effective for speech translation, where RNN-T's strict monotonic alignment hurts performance. Our results demonstrate that the CHAT model offers a practical solution for deploying more capable streaming speech models without sacrificing real-time constraints.

In 3D point cloud object tracking, the motion-centric methods have emerged as a promising avenue due to its superior performance in modeling inter-frame motion. However, existing two-stage motion-based approaches suffer from fundamental limitations: (1) error accumulation due to decoupled optimization caused by explicit foreground segmentation prior to motion estimation, and (2) computational bottlenecks from sequential processing. To address these challenges, we propose FocusTrack, a novel one-stage paradigms tracking framework that unifies motion-semantics co-modeling through two core innovations: Inter-frame Motion Modeling (IMM) and Focus-and-Suppress Attention. The IMM module employs a temp-oral-difference siamese encoder to capture global motion patterns between adjacent frames. The Focus-and-Suppress attention that enhance the foreground semantics via motion-salient feature gating and suppress the background noise based on the temporal-aware motion context from IMM without explicit segmentation. Based on above two designs, FocusTrack enables end-to-end training with compact one-stage pipeline. Extensive experiments on prominent 3D tracking benchmarks, such as KITTI, nuScenes, and Waymo, demonstrate that the FocusTrack achieves new SOTA performance while running at a high speed with 105 FPS.

How natural speech is represented in the brain constitutes a major challenge for cognitive neuroscience. Reconstructing the speech envelope and Mel spectrogram from EEG and MEG provides a time-resolved way to study its temporal and spectral structure. Speech-related neural activity spans sensors and temporal scales; extracting these representations while adapting the use of context to each acoustic target is a central problem in speech reconstruction. We propose SHINE, a Sequential Hierarchical Integration Network for EEG and MEG. A residual sensor adapter unifies input dimensions, intermediate dilated-block states retain temporal depth, and a target- and time-dependent gate fuses local hierarchical and attention-enhanced context predictions. Across two EEG and two MEG datasets, SHINE has the highest mean envelope and mean-Mel Pearson correlations among nine local baseline implementations on all eight dataset-metric combinations. SHINE also placed second in the speech-detection Extended Track of the NeurIPS 2025 PNPL Competition. Code will be released at https://github.com/xuxiran/SHINE.

In this paper, we propose ReSeg-CLIP, a new training-free Open-Vocabulary Semantic Segmentation method for remote sensing data. To compensate for the problems of vision language models, such as CLIP in semantic segmentation caused by inappropriate interactions within the self-attention layers, we introduce a hierarchical scheme utilizing masks generated by SAM to constrain the interactions at multiple scales. We also present a model composition approach that averages the parameters of multiple RS-specific CLIP variants, taking advantage of a new weighting scheme that evaluates representational quality using varying text prompts. Our method achieves state-of-the-art results across three RS benchmarks without additional training.

Transformer attention is typically implemented using softmax normalization, which enforces attention weights with unit sum normalization. While effective in many settings, this constraint can limit flexibility in controlling attention magnitudes and may contribute to overly concentrated or unstable attention patterns during training. Prior work has explored modifications such as attention sinks or gating mechanisms, but these approaches provide only limited or indirect control over attention reweighting. We propose Affine-Scaled Attention, a simple extension to standard attention that introduces input-dependent scaling and a corresponding bias term applied to softmax-normalized attention weights. This design relaxes the strict normalization constraint while maintaining aggregation of value representations, allowing the model to adjust both the relative distribution and the scale of attention in a controlled manner. We empirically evaluate Affine-Scaled Attention in large-scale language model pretraining across multiple model sizes. Experimental results show consistent improvements in training stability, optimization behavior, and downstream task performance compared to standard softmax attention and attention sink baselines. These findings suggest that modest reweighting of attention outputs provides a practical and effective way to improve attention behavior in Transformer models.

Retrieval heads, a subset of attention heads in Transformers, were studied in English, showing its crucial role in retrieving information from the context. We expand the study of retrieval heads to multilingual context and find that while nearly half of all retrieval heads are often shared across multiple languages, language-specific retrieval heads also emerge. We further design a cross-lingual needle-in-the-haystack task, to identify $\textit{Retrieval Transition Heads (RTH)}$ that retrieve key information and further map it to target-language output. Our experiments reveal that RTHs do not always overlap with RH, and can be more vital for Chain-of-Thought reasoning in multilingual LLMs. Across four multilingual benchmarks (MMLU-ProX, MGSM, MLQA, and XQuAD) and two model families (Qwen2.5-7B Instruct and Llama3.1-8B Instruct), we demonstrate that masking RTH induces bigger performance drop than masking Retrieval Heads (RH). Specifically, for Llama3.1-8B Instruct, masking the top-25 RTHs results in an average 37.3-point drop (against 19.5-point drop from RH masking) in reasoning accuracy (MMLU-ProX, MGSM) and 9.0 F1-score reduction (against 5.1 drop for RH) in extractive QA (MLQA, XQuAD). Our work advances the understanding of multilingual LMs by isolating the attention heads responsible for mapping to target languages.

The pursuit of world model based artificial intelligence has predominantly relied on projecting high-dimensional observations into parameterized latent spaces, wherein transition dynamics are subsequently learned. However, this conventional paradigm is mathematically flawed: it merely displaces the manifold learning problem into the latent space. When the underlying data distribution shifts, the latent manifold shifts accordingly, forcing the predictive operator to implicitly relearn the new topological structure. Furthermore, by classical approximation theory, positive operators like dot product attention inevitably suffer from the saturation phenomenon, permanently bottlenecking their predictive capacity and leaving them vulnerable to the curse of dimensionality. In this paper, we formulate a mathematically rigorous paradigm for world model construction by redefining the core predictive mechanism. Inspired by Ryan O'Dowd's foundational work we introduce Spherical Kernel Operator (SKO), a framework that replaces standard attention. By projecting the unknown data manifold onto a unified ambient hypersphere and utilizing a localized sequence of ultraspherical (Gegenbauer) polynomials, SKO performs direct integral reconstruction of the target function. Because this localized spherical polynomial kernel is not strictly positive, it bypasses the saturation phenomenon, yielding approximation error bounds that depend strictly on the intrinsic manifold dimension q, rather than the ambient dimension. Furthermore, by formalizing its unnormalized output as an authentic measure support estimator, SKO mathematically decouples the true environmental transition dynamics from the biased observation frequency of the agent. Empirical evaluations confirm that SKO significantly accelerates convergence and outperforms standard attention baselines in autoregressive language modeling.

Accurate prediction of malignancy in renal tumors is crucial for informing clinical decisions and optimizing treatment strategies. However, existing imaging modalities lack the necessary accuracy to reliably predict malignancy before surgical intervention. While deep learning has shown promise in malignancy prediction using 3D CT images, traditional approaches often rely on manual segmentation to isolate the tumor region and reduce noise, which enhances predictive performance. Manual segmentation, however, is labor-intensive, costly, and dependent on expert knowledge. In this study, a deep learning framework was developed utilizing an Organ Focused Attention (OFA) loss function to modify the attention of image patches so that organ patches attend only to other organ patches. Hence, no segmentation of 3D renal CT images is required at deployment time for malignancy prediction. The proposed framework achieved an AUC of 0.685 and an F1-score of 0.872 on a private dataset from the UF Integrated Data Repository (IDR), and an AUC of 0.760 and an F1-score of 0.852 on the publicly available KiTS21 dataset. These results surpass the performance of conventional models that rely on segmentation-based cropping for noise reduction, demonstrating the frameworks ability to enhance predictive accuracy without explicit segmentation input. The findings suggest that this approach offers a more efficient and reliable method for malignancy prediction, thereby enhancing clinical decision-making in renal cancer diagnosis.

Sequential self-attention models usually rely on additive positional embeddings, which inject positional information into item representations at the input. In the absence of positional signals, the attention block is permutation-equivariant over sequence positions and thus has no intrinsic notion of temporal order beyond causal masking. We argue that additive positional embeddings make the attention mechanism only superficially sensitive to sequence order: positional information is entangled with item embedding semantics, propagates weakly in deep architectures, and limits the ability to capture rich sequential patterns. To address these limitations, we introduce a kernelized self-attention mechanism, where a learnable positional kernel operates purely in the position space, disentangled from semantic similarity, and directly modulates attention weights. When applied per attention block, this kernel enables adaptive multi-scale sequential modeling. Experiments on standard next-item prediction benchmarks show that our positional kernel attention consistently improves over strong competing baselines.

Attention learners, neural networks built on the attention mechanism, e.g., transformers, excel at learning the implicit relationships that relate sequences to their corresponding properties, e.g., mapping a given sequence of tokens to the probability of the next token. However, the learning process tends to be costly. To address this, we present a novel paradigm named Attention Neural Teaching (AtteNT) that reinterprets the learning process through a nonparametric teaching perspective. Specifically, the latter provides a theoretical framework for teaching mappings that are implicitly defined (i.e., nonparametric) via example selection. Such an implicit mapping is embodied through a dense set of sequence-property pairs, with the AtteNT teacher selecting a subset to accelerate convergence in attention learner training. By analytically investigating the role of attention on parameter-based gradient descent during training, and recasting the evolution of attention learners, shaped by parameter updates, through functional gradient descent in nonparametric teaching, we show for the first time that teaching attention learners is consistent with teaching importance-adaptive nonparametric learners. These new findings readily commit AtteNT to enhancing learning efficiency of attention learners. Specifically, we observe training time reductions of 13.01% for LLMs and 20.58% for ViTs, spanning both fine-tuning and training-from-scratch regimes. Crucially, these gains are achieved without compromising accuracy; in fact, performance is consistently preserved and often enhanced across a diverse set of downstream tasks.

The strong capabilities of recent Large Language Models (LLMs) have made them highly effective for zero-shot re-ranking task. Attention-based re-ranking methods, which derive relevance scores directly from attention weights, offer an efficient and interpretable alternative to generation-based re-ranking methods. However, they still face two major limitations. First, attention signals are highly concentrated a small subset of tokens within a few documents, making others indistinguishable. Second, attention often overemphasizes phrases lexically similar to the query, yielding biased rankings that irrelevant documents with mere lexical resemblance are regarded as relevant. In this paper, we propose ReAttn, a post-hoc re-weighting strategy for attention-based re-ranking methods. It first compute the cross-document IDF weighting to down-weight attention on query-overlapping tokens that frequently appear across the candidate documents, reducing lexical bias and emphasizing distinctive terms. It then employs entropy-based regularization to mitigate over-concentrated attention, encouraging a more balanced distribution across informative tokens. Both adjustments operate directly on existing attention weights without additional training or supervision. Extensive experiments demonstrate the effectiveness of our method.

This paper studies simple transformers trained on a high-order Markov chain, where the model must incorporate information from multiple past positions, each with different statistical importance. We show that transformers learn the task incrementally, with each stage corresponding to learning how to copy information from a subset of positions via a sparse attention pattern. Notably, the learning dynamics transition from a competitive phase, where all heads focus on the statistically most important positions, to a cooperative phase, where different heads specialize in different patterns. We model these dynamics with simplified differential equations and prove stage-wise convergence of the resulting system. Functionally, these stages correspond to a sequence of increasingly expressive misspecified models, with the full model class reached only at the end. Overall, we give a theoretical account of how structured attention patterns and head specialization emerge in stages without an explicit curriculum, with implications for generalization in sequential tasks.

Federated learning has emerged as a privacy-preserving and efficient approach for deploying intelligent agricultural solutions. Accurate edge-based diagnosis across geographically dispersed farms is crucial for recognising tomato diseases in sustainable farming. Traditional centralised training aggregates raw data on a central server, leading to communication overhead, privacy risks and latency. Meanwhile, edge devices require lightweight networks to operate effectively within limited resources. In this paper, we propose U-FedTomAtt, an ultra-lightweight federated learning framework with attention for tomato disease recognition in resource-constrained and distributed environments. The model comprises only 245.34K parameters and 71.41 MFLOPS. First, we propose an ultra-lightweight neural network with dilated bottleneck (DBNeck) modules and a linear transformer to minimise computational and memory overhead. To mitigate potential accuracy loss, a novel local-global residual attention (LoGRA) module is incorporated. Second, we propose the federated dual adaptive weight aggregation (FedDAWA) algorithm that enhances global model accuracy. Third, our framework is validated using three benchmark datasets for tomato diseases under simulated federated settings. Experimental results show that the proposed method achieves 0.9910% and 0.9915% Top-1 accuracy and 0.9923% and 0.9897% F1-scores on SLIF-Tomato and PlantVillage tomato datasets, respectively.

The prefill stage in long-context LLM inference remains a computational bottleneck. Recent token-ranking heuristics accelerate inference by selectively processing a subset of semantically relevant tokens. However, existing methods suffer from unstable token importance estimation, often varying between layers. Evaluating token-ranking quality independently from heuristic-specific architectures is challenging. To address this, we introduce an Answer-Informed Oracle, which defines ground-truth token importance by measuring attention from generated answers back to the prompt. This oracle reveals that existing heuristics exhibit high variance across layers: rankings can degrade sharply at specific layers, a failure mode invisible to end-to-end benchmarks. The diagnosis suggests a simple fix: aggregate scores across layers rather than relying on any single one. We implement this as Cross-Layer Attention Aggregation (CLAA), which closes the gap to the oracle upper bound and reduces Time-to-First-Token (TTFT) by up to 39% compared to the Full KV Cache baseline.

General-purpose super-resolution models, particularly Vision Transformers, have achieved remarkable success but exhibit fundamental inefficiencies in common infrared imaging scenarios like surveillance and autonomous driving, which operate from fixed or nearly-static viewpoints. These models fail to exploit the strong, persistent spatial priors inherent in such scenes, leading to redundant learning and suboptimal performance. To address this, we propose the Regional Prior attention Transformer for infrared image Super-Resolution (RPT-SR), a novel architecture that explicitly encodes scene layout information into the attention mechanism. Our core contribution is a dual-token framework that fuses (1) learnable, regional prior tokens, which act as a persistent memory for the scene's global structure, with (2) local tokens that capture the frame-specific content of the current input. By utilizing these tokens into an attention, our model allows the priors to dynamically modulate the local reconstruction process. Extensive experiments validate our approach. While most prior works focus on a single infrared band, we demonstrate the broad applicability and versatility of RPT-SR by establishing new state-of-the-art performance across diverse datasets covering both Long-Wave (LWIR) and Short-Wave (SWIR) spectra

Multimodal methods are widely used in rice deterioration detection, but they exhibit limited capability in representing and extracting fine-grained abnormal features. Moreover, these methods rely on devices such as hyperspectral cameras and mass spectrometers, which increase detection costs and prolong data acquisition time. To address these issues, we propose a feature recalibration based olfactory-visual multimodal model for enhanced rice deterioration detection. A fine-grained deterioration embedding constructor (FDEC) is proposed to reconstruct the labeled multimodal embedded feature dataset, thereby enhancing sample representation. A fine-grained deterioration recalibration attention network (FDRA-Net) is proposed to emphasize signal variations and improve sensitivity to fine-grained deterioration on the rice surface. Compared with SS-Net, the proposed method improves classification accuracy by 8.67%, with an average improvement of 11.51% over other traditional baseline models, while simultaneously simplifying the detection procedure. Furthermore, field detection results demonstrate advantages in both accuracy and operational simplicity. The proposed method can also be extended to other agrifood applications in agriculture and the food industry.

More than half the global population now carries devices that can run ChatGPT-like language models with no Internet connection and minimal safety oversight -- and hence the potential to promote self-harm, financial losses and extremism among other dangers. Existing safety tools either require cloud connectivity or discover failures only after harm has occurred. Here we show that a large class of potentially dangerous tipping originates at the atomistic scale in such edge AI due to competition for the machinery's attention. This yields a mathematical formula for the dynamical tipping point n*, governed by dot-product competition for attention between the conversation's context and competing output basins, that reveals new control levers. Validated against multiple AI models, the mechanism can be instantiated for different definitions of 'good' and 'bad' and hence in principle applies across domains (e.g. health, law, finance, defense), changing legal landscapes (e.g. EU, UK, US and state level), languages, and cultural settings.

Large language models (LLMs) exhibit a wide range of capabilities, including mathematical reasoning, code generation, and linguistic behaviors. We show that Transformer architectures contain small subsets of attention heads that are necessary for certain capabilities. Zeroing out as few as five task-specific heads can degrade performance by up to $60\%$ on standard benchmarks measuring the capability of interest, while largely preserving performance on unrelated tasks. We introduce a compressed sensing-based method that exploits the sparsity of these heads to identify them via strategic knockouts and a small number of model evaluations. We validate these findings across Llama and Qwen models ranging from 1B to 14B parameters and a diverse set of capabilities including mathematical abilities and code generation, revealing a modular organization in which specialized capabilities are dependent on sparse, functionally distinct components. Overall, our results suggest that capability localization is a general organizational principle of Transformer language models, with implications for interpretability, model editing, and AI safety. Code is released at https://github.com/locuslab/llm-components.

Learned gates can approximate sparse attention patterns on frozen transformers, yet provide limited benefit over random gates when trained jointly with the model. We investigate this difference in a controlled 31M-parameter transformer and attribute it to routing absorption: model representations co-adapt to the imposed mask, reducing the incremental benefit of learned routing. Four experiments characterize the phenomenon. Differentiable soft gating yields perplexities of 48.73 plus or minus 0.60 with learned gates and 49.83 plus or minus 0.04 with frozen random gates over three seeds. Hard top-k masking provides no gradient path to the gate scores in the tested implementation. Gates distilled onto co-adapted and dense-trained Q/K/V both achieve high F1 against oracle masks, but hard-mask deployment yields perplexities of 601.6 and 48.6, respectively. Stochastic mask training also leaves a substantial deployment penalty: dense evaluation yields 78.2 perplexity, compared with 37.3 for the dense baseline. Experiments on Qwen3-1.7B show that increasing the number of trainable attention layers reduces the gap between learned and random gates. We relate these results to co-adaptation in Mixture-of-Experts and propose parameter asymmetry between the gate and the model as a contributing mechanism. For the tested per-query, token-level gates, freezing the model provides stable routing targets and enables effective post-hoc sparsification. The results motivate random-routing controls and separate evaluation of routing quality and model adaptation in sparse attention methods.

With the increasing availability of high-resolution remote sensing and aerial imagery, oriented object detection has become a key capability for geographic information updating, maritime surveillance, and disaster response. However, it remains challenging due to cluttered backgrounds, severe scale variation, and large orientation changes. Existing approaches largely improve performance through multi-scale feature fusion with feature pyramid networks or contextual modeling with attention, but they often lack explicit foreground modeling and do not leverage geometric orientation priors, which limits feature discriminability. To overcome these limitations, we propose FGAA-FPN, a Foreground-Guided Angle-Aware Feature Pyramid Network for oriented object detection. FGAA-FPN is built on a hierarchical functional decomposition that accounts for the distinct spatial resolution and semantic abstraction across pyramid levels, thereby strengthening multi-scale representations. Concretely, a Foreground-Guided Feature Modulation module learns foreground saliency under weak supervision to enhance object regions and suppress background interference in low-level features. In parallel, an Angle-Aware Multi-Head Attention module encodes relative orientation relationships to guide global interactions among high-level semantic features. Extensive experiments on DOTA v1.0 and DOTA v1.5 demonstrate that FGAA-FPN achieves state-of-the-art results, reaching 75.5% and 68.3% mAP, respectively.

A core bottleneck in large language model (LLM) inference is the cost of attending over the ever-growing key-value (KV) cache. Although near-oracle top-k KV selection can preserve the quality of dense attention while sharply reducing computation and bandwidth, existing sparse methods generally rely on posterior heuristics, i.e., selectors conditioned on observed attention or proxy scores. Such conditioning introduces posterior bias: it tends to distort true token importance and miss salient tokens, thereby impairing long-range reasoning. To tackle this problem, we propose Pre-hoc Sparsity (PrHS), which selects KV entries before attention scoring and provides explicit accuracy control. Let the attention mass of discarded entries be delta (the dropped mass). Through a marginal-to-mutual-information analysis, we derive an upper bound on the mutual-information loss that depends only on the dropped mass. This relation explains failure modes of posterior heuristics and enables verifiable guarantees by controlling the dropped mass in advance. Within PrHS, we instantiate three orthogonal pre-hoc selectors along the axes of time, depth, and layer. Extensive experiments on LLaMA and Mistral families validate PrHS. Across GSM8K and CoQA, PrHS reduces retrieval overhead by over 90%, achieving 3x higher retrieval sparsity than HShare at matched or better accuracy. It incurs under 1% average degradation on LongBench, lowers attention FLOPs by about 15% versus prior sparse baselines, and yields a 9.9x speedup in attention-operator latency and 2.8x higher throughput on NVIDIA A100-80GB GPUs than the dense baseline.

Efficient long-context inference faces two coupled bottlenecks: KV-cache memory grows linearly with context length, while attention computation grows quadratically. Existing approaches typically address one at the expense of irreversible token eviction, full-cache retention, or full-history reconstruction, limiting their effectiveness for multi-turn interaction and long-form reasoning. Motivated by two empirical properties, Long-Range Inter-Token Similarity and Smooth Residual Distribution, we propose ResidualKV, which factorizes the KV cache into a sparse set of globally retrieved references and compact, quantized residual codes for the remaining tokens. This representation preserves token-specific information without permanent eviction and, when combined with sparse attention, reconstructs only the selected states on demand. Dynamic-stride scheduling further reduces reference growth from linear to approximately logarithmic at ultra-long contexts. Across Llama, Qwen, LLaVA-OV, and Qwen3-VL backbones, ResidualKV maintains near-full-cache performance using only 13%-16% KV storage and 30% attention computation on LongBench, and 8%-10% storage and 10% computation in matched-budget multimodal evaluation. It also accelerates decoding by up to $1.5\times$ with KV-cache quantization and $3.4\times$ without it. These results show that global cross-token redundancy supports accurate, memory-efficient, and computation-efficient long-context inference. The source code is available at https://github.com/CURRENTF/ResidualKV.

We propose a variational quantum implementation of self-attention (QSA)-the core operation in transformers and large language models-which predicts future elements of a sequence by forming overlap-weighted combinations of past data. At variance with previous approaches, our QSA realizes the required nonlinearity through interference of state overlaps and a degree-$k$ polynomial kernel, and estimates a loss based on Rényi-$1/2$ entropic functionals via two observables' expectation values, avoiding the decoding of amplitude-encoded predictions into classical probabilities. QSA also accommodates a constrained, trainable data-embedding tying state overlaps to data-level similarities. Its dominant end-to-end training complexity scales as $O\left(μ^{-1}k^2Td\right)$, versus $O\left(T d^{k+1}\right)$ of the fairest classical comparison, with $μ$ a training signal; we show numerically that this allows a complexity advantage in the regime where sequence length $T$ dominates the embedding size $d$. In simulations, our QSA-based quantum transformer learns sequence prediction on classical data and on many-body transverse-field Ising trajectories-establishing trainable attention as a practical primitive for quantum dynamical modeling.

Generating long-form content, such as minute-long videos and extended texts, is increasingly important for modern generative models. Block diffusion improves inference efficiency via KV caching and block-wise causal inference and has been widely adopted in diffusion language models and video generation. However, in long-context settings, block diffusion still incurs substantial overhead from repeatedly computing attention over a growing KV cache. We identify an underexplored property of block diffusion: cross-step redundancy of attention within a block. Our analysis shows that attention outputs from tokens outside the current block remain largely stable across diffusion steps, while block-internal attention varies significantly. Based on this observation, we propose FlashBlock, a cached block-external attention mechanism that reuses stable attention output, reducing attention computation and KV cache access without modifying the diffusion process. Moreover, FlashBlock is orthogonal to sparse attention and can be combined as a complementary residual reuse strategy, substantially improving model accuracy under aggressive sparsification. Experiments on diffusion language models and video generation demonstrate up to 1.44$\times$ higher token throughput and up to 1.6$\times$ reduction in attention time, with negligible impact on generation quality. Project page: https://caesarhhh.github.io/FlashBlock/.

The self-attention mechanism, at the heart of the Transformer model, is able to effectively model pairwise interactions between tokens. However, numerous recent works have shown that it is unable to perform basic tasks involving detecting triples of correlated tokens, or compositional tasks where multiple input tokens need to be referenced to generate a result. Some higher-dimensional alternatives to self-attention have been proposed to address this, including higher-order attention and Strassen attention, which can perform some of these polyadic tasks in exchange for slower, superquadratic running times. In this work, we define a vast class of generalizations of self-attention, which we call poly-attention mechanisms. Our mechanisms can incorporate arbitrary higher-order (tensor) computations as well as arbitrary relationship structures between the input tokens, and they include the aforementioned alternatives as special cases. We then systematically study their computational complexity and representational strength, including giving new algorithms and matching complexity-theoretic lower bounds on the time complexity of computing the attention matrix exactly as well as approximately, and tightly determining which polyadic tasks they can each perform. Our results give interesting trade-offs between different desiderata for these mechanisms, including a tight relationship between how expressive a mechanism is, and how large the coefficients in the model may be so that the mechanism can be approximated in almost-linear time. Notably, we give a new attention mechanism which can be computed exactly in quadratic time, and which can perform function composition for any fixed number of functions. Prior mechanisms, even for just composing two functions, could only be computed in superquadratic time, and our new lower bounds show that faster algorithms for them are not possible.

Linearizing pretrained large language models (LLMs) primarily relies on intra-layer hybrid attention mechanisms to alleviate the quadratic complexity of standard softmax attention. Existing methods perform token routing based on sliding-window partitions, resulting in position-based selection and fails to capture token-specific global importance. Meanwhile, linear attention further suffers from distribution shift caused by learnable feature maps that distort pretrained feature magnitudes. Motivated by these limitations, we propose STILL, an intra-layer hybrid linearization framework for efficiently linearizing LLMs. STILL introduces a Self-Saliency Score with strong local-global consistency, enabling accurate token selection using sliding-window computation, and retains salient tokens for sparse softmax attention while summarizing the remaining context via linear attention. To preserve pretrained representations, we design a Norm-Preserved Feature Map (NP-Map) that decouples feature direction from magnitude and reinjects pretrained norms. We further adopt a unified training-inference architecture with chunk-wise parallelization and delayed selection to improve hardware efficiency. Experiments show that STILL matches or surpasses the original pretrained model on commonsense and general reasoning tasks, and achieves up to a 86.2% relative improvement over prior linearized attention methods on long-context benchmarks. Code is available at this URL.

Neural Controlled Differential Equations (Neural CDEs) provide a powerful continuous-time framework for sequence modeling, yet the roughness of the driving control path often restricts their efficiency. Standard splines introduce high-frequency variations that force adaptive solvers to take excessively small steps, driving up the Number of Function Evaluations (NFE). We propose a novel approach to Neural CDE path construction that replaces exact interpolation with Kernel and Gaussian Process (GP) smoothing, enabling explicit control over trajectory regularity. To recover details lost during smoothing, we propose an attention-based Multi-View CDE (MV-CDE) and its convolutional extension (MVC-CDE), which employ learnable queries to inform path reconstruction. This framework allows the model to distribute representational capacity across multiple trajectories, each capturing distinct temporal patterns. Empirical results demonstrate that our method, MVC-CDE with GP, achieves state-of-the-art accuracy while significantly reducing NFEs and total inference time compared to spline-based baselines.

Diffusion language models (DLMs) generate text through iterative denoising, but inference requires full-sequence attention at every iteration, resulting in substantial redundant computation on masked tokens. Block-wise diffusion can reduce this cost, yet it typically relies on retraining and constrained update orders, limiting its direct applicability to pretrained DLMs. Our token-level analysis reveals pronounced structural locality in DLM inference. Decoding is driven by a small set of prefix-localized active tokens; the influence of distant undecoded context diminishes rapidly, and decoded tokens exhibit stage-wise temporal stability, enabling reuse of intermediate representations except for a brief post-decode transient. Motivated by these observations, we propose \placeholder\footnote{The source code is available at https://github.com/vhicrgit/Window-Diffusion.}, a window-based token pruning and caching method for inference. We maintain a local computation window that slides rightward as denoising progresses, and partition undecoded tokens into: (i) active tokens that are computed online, (ii) buffer tokens whose KV states are cached and periodically refreshed, and (iii) far-field tokens that are pruned outside the window. Computation is restricted to active and buffer tokens within the window, while far-field tokens are omitted at each stage. Experiments on LLaDA and Dream show that, under matched compute budgets, our method achieves up to $99\times$ inference speedup while largely preserving generation performance.