Daily Papers

Depth-recurrence facilitates latent reasoning by sharing parameters across depths. However, prior work lacks combined FLOP-, parameter-, and memory-matched baselines, underutilizes depth-recurrence due to partially fixed layer stacks, and ignores the bottleneck of constant hidden-sizes that restricts many-step latent reasoning. To address this, we introduce a modular framework of depth-recurrent attention mixtures (Dreamer), combining sequence attention, depth attention, and sparse expert attention. It alleviates the hidden-size bottleneck through attention along depth, decouples scaling dimensions, and allows depth-recurrent models to scale efficiently and effectively. Across language reasoning benchmarks, our models require 2 to 8x fewer training tokens for the same accuracy as FLOP-, parameter-, and memory-matched SOTA, and outperform ca. 2x larger SOTA models with the same training tokens. We further present insights into knowledge usage across depths, e.g., showing 2 to 11x larger expert selection diversity than SOTA MoEs.

Recent advances in large language models highlighted the excessive quadratic cost of self-attention. Despite the significant research efforts, subquadratic attention methods still suffer from inferior performance in practice. We hypothesize that dynamic, learned content-based sparsity can lead to more efficient attention mechanisms. We present Mixture of Sparse Attention (MoSA), a novel approach inspired by Mixture of Experts (MoE) with expert choice routing. MoSA dynamically selects tokens for each attention head, allowing arbitrary sparse attention patterns. By selecting k tokens from a sequence of length T, MoSA reduces the computational complexity of each attention head from O(T^2) to O(k^2 + T). This enables using more heads within the same computational budget, allowing higher specialization. We show that among the tested sparse attention variants, MoSA is the only one that can outperform the dense baseline, sometimes with up to 27% better perplexity for an identical compute budget. MoSA can also reduce the resource usage compared to dense self-attention. Despite using torch implementation without an optimized kernel, perplexity-matched MoSA models are simultaneously faster in wall-clock time, require less memory for training, and drastically reduce the size of the KV-cache compared to the dense transformer baselines.

Hepatocellular Carcinoma diagnosis relies heavily on the interpretation of gigapixel Whole Slide Images. However, current computational approaches are constrained by fixed-resolution processing mechanisms and inefficient feature aggregation, which inevitably lead to either severe information loss or high feature redundancy. To address these challenges, we propose Hepato-LLaVA, a specialized Multi-modal Large Language Model designed for fine-grained hepatocellular pathology analysis. We introduce a novel Sparse Topo-Pack Attention mechanism that explicitly models 2D tissue topology. This mechanism effectively aggregates local diagnostic evidence into semantic summary tokens while preserving global context. Furthermore, to overcome the lack of multi-scale data, we present HepatoPathoVQA, a clinically grounded dataset comprising 33K hierarchically structured question-answer pairs validated by expert pathologists. Our experiments demonstrate that Hepato-LLaVA achieves state-of-the-art performance on HCC diagnosis and captioning tasks, significantly outperforming existing methods. Our code and implementation details are available at https://pris-cv.github.io/Hepto-LLaVA/.

We present Hydra as an architectural proposal for hybrid long-context language models that combine conditional computation, long-context memory mechanisms, and sparse mixture-of-experts within an approximately 1.6B parameter design envelope. Hydra integrates a Mamba-style Structured State Space Model (SSM) backbone with intermittent sparse global attention, chunk-level MoE feed-forward routing, and dual (workspace plus factual PKM) memories. We formalize the component interfaces, give transparent parameter and complexity accounting, and outline a staged curriculum intended to stably activate the parts. We accompany the specification with illustrative toy-scale prototype measurements (tens of millions of parameters on synthetic data) whose sole purpose is to demonstrate implementation feasibility and qualitative scaling behaviors (for example, long-context throughput crossover and controllable expert routing), not to claim competitive full-scale performance. We explicitly delineate assumptions and open risks (training complexity, memory utilization, specialization dynamics) and position Hydra as a blueprint to stimulate empirical follow-up rather than a finished system. By combining SSM efficiency, selective sparse attention, MoE capacity, and learnable memory, Hydra sketches a path toward modular, input-adaptive long-context language models; validating end-task gains at target scale remains future work.

Sparse Mixture of Experts (SMoE) has emerged as a key to achieving unprecedented scalability in deep learning. By activating only a small subset of parameters per sample, SMoE achieves an exponential increase in parameter counts while maintaining a constant computational overhead. However, SMoE models are susceptible to routing fluctuations--changes in the routing of a given input to its target expert--at the late stage of model training, leading to model non-robustness. In this work, we unveil the limitation of SMoE through the perspective of the probabilistic graphical model (PGM). Through this PGM framework, we highlight the independence in the expert-selection of tokens, which exposes the model to routing fluctuation and non-robustness. Alleviating this independence, we propose the novel Similarity-Aware (S)MoE, which considers interactions between tokens during expert selection. We then derive a new PGM underlying an (S)MoE-Attention block, going beyond just a single (S)MoE layer. Leveraging the token similarities captured by the attention matrix, we propose the innovative Attention-Aware (S)MoE, which employs the attention matrix to guide the routing of tokens to appropriate experts in (S)MoE. We theoretically prove that Similarity/Attention-Aware routing help reduce the entropy of expert selection, resulting in more stable token routing mechanisms. We empirically validate our models on various tasks and domains, showing significant improvements in reducing routing fluctuations, enhancing accuracy, and increasing model robustness over the baseline MoE-Transformer with token routing via softmax gating.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

Mixture-of-Experts (MoE) architectures are often considered a natural fit for continual learning because sparse routing should localize updates and reduce interference, yet MoE Transformers still forget substantially even with sparse, well-balanced expert utilization. We attribute this gap to a pre-routing bottleneck: multi-head attention concatenates head-specific signals into a single post-attention router input, forcing routing to act on co-occurring feature compositions rather than separable head channels. We show that this router input simultaneously encodes multiple separately decodable semantic and structural factors with uneven head support, and that different feature compositions induce weakly aligned parameter-gradient directions; as a result, routing maps many distinct compositions to the same route. We quantify this collision effect via a route-wise effective composition number N_{eff} and find that higher N_{eff} is associated with larger old-task loss increases after continual training. Motivated by these findings, we propose MH-MoE, which performs head-wise routing over sub-representations to increase routing granularity and reduce composition collisions. On TRACE with Qwen3-0.6B/8B, MH-MoE effectively mitigates forgetting, reducing BWT on Qwen3-0.6B from 11.2% (LoRAMoE) to 4.5%.

Recent advancements in large language model (LLM) unlearning have shown remarkable success in removing unwanted data-model influences while preserving the model's utility for legitimate knowledge. However, despite these strides, sparse Mixture-of-Experts (MoE) LLMs--a key subset of the LLM family--have received little attention and remain largely unexplored in the context of unlearning. As MoE LLMs are celebrated for their exceptional performance and highly efficient inference processes, we ask: How can unlearning be performed effectively and efficiently on MoE LLMs? And will traditional unlearning methods be applicable to MoE architectures? Our pilot study shows that the dynamic routing nature of MoE LLMs introduces unique challenges, leading to substantial utility drops when existing unlearning methods are applied. Specifically, unlearning disrupts the router's expert selection, causing significant selection shift from the most unlearning target-related experts to irrelevant ones. As a result, more experts than necessary are affected, leading to excessive forgetting and loss of control over which knowledge is erased. To address this, we propose a novel single-expert unlearning framework, referred to as UOE, for MoE LLMs. Through expert attribution, unlearning is concentrated on the most actively engaged expert for the specified knowledge. Concurrently, an anchor loss is applied to the router to stabilize the active state of this targeted expert, ensuring focused and controlled unlearning that preserves model utility. The proposed UOE framework is also compatible with various unlearning algorithms. Extensive experiments demonstrate that UOE enhances both forget quality up to 5% and model utility by 35% on MoE LLMs across various benchmarks, LLM architectures, while only unlearning 0.06% of the model parameters.

While Transformer architectures have demonstrated impressive scalability across domains, they continue to face challenges in long-context reasoning, computational efficiency, and structural generalization - largely due to rigid layer stacking, dense attention, and reliance on positional encodings. We present ReSSFormer, a Recursive Sparse Structured Transformer that integrates three complementary innovations: Recurrent Reasoning & Memory Unit (R2MU) for iterative reasoning with bounded depth, Adaptive Sparse Attention Module (ASAM) for efficient and focused context selection, and Self-Organizing Encoder Structure (SOES) for position-free structure induction. ReSSFormer replaces conventional depth stacking with recurrent inference, substitutes full attention with token- and expert-level sparsity, and models latent token topology directly from content. Across language modeling, multi-hop QA, and structure-sensitive tasks, ReSSFormer consistently outperforms strong baselines under comparable FLOPs and parameter budgets, highlighting its scalability, efficiency, and structural flexibility.

This paper systematically diagnoses the training failure modes of Token-Choice sparse Mixture-of-Experts (MoE) on video Diffusion Transformers. Starting from a pretrained dense model of about 5 billion parameters, we convert it into an MoE architecture following three laws: routed experts exactly clone the original FFN weights, shared experts are initialized to zero for verification and then to extremely small non-zero noise for actual training, while only the gating networks start from random initialization. Experiments reveal a hierarchy of five failure modes: (1) linear routers suffer global soft saturation with complete expert homogenization; (2) MLP routers introduce selective deadlock, where roughly one-third of layers degenerate into a single-expert mode that cannot be prevented by increasing the auxiliary loss; (3) cross-attention routers exhibit preliminary self-recovery, yet about nine layers remain stubbornly deadlocked; (4) deadlocked layers display a U-shaped distribution, concentrated in shallow visual processing layers and deep semantic integration layers; (5) bfloat16 mixed precision causes tiny weight updates to be truncated to zero by hardware. Based on routing decision time series over 65 million tokens across 5,000 training steps, we propose the Functional Redundancy Hypothesis: deadlock is a rational waiting strategy before the shared expert matures within the gate-shared expert-routed expert triadic system. This hypothesis is supported by the theory of functional redundancy in systems biology. On the engineering side, we summarize the Three Laws of dense-to-MoE conversion and provide a complete solution for the bfloat16 precision trap. We calibrate the current capability boundary of the Token-Choice paradigm and outline a three-step evolutionary roadmap from visual unification to a world model.

Conditional computation can decouple language model quality from per-token inference cost, yet leading techniques act on a single axis in isolation: Mixture-of-Experts (MoE) sparsifies the FFN, Mixture-of-Depths (MoD) skips whole transformer blocks, and KV-cache quantization compresses attention memory. We argue these three decisions (attention resolution, expert selection, and cache bit-width) are strongly coupled and should be made jointly: a token rare enough to warrant full attention may also need high-precision caching regardless of which expert processes it. We introduce TriRoute, a single lightweight controller shared across all three axes that, for every token at every layer, emits a coordinated policy: (i) an attention mode (skip/local/full), (ii) a sparse set of FFN experts (with a null expert recovering MoD), and (iii) a KV-cache bit-width. The controller trains end-to-end via a heterogeneous relaxation (Gumbel-Softmax with straight-through estimation for categorical decisions and load-balanced top-k gating for experts) under a Lagrangian budget constraint that turns the average compute and memory cost into a controllable knob. We identify a cross-axis routing-collapse cascade in naive joint training, where collapse on one axis propagates to the others, and address it with per-axis normalization and a coupling-aware balancing loss. On decoder-only models from 160M to 1.3B parameters at compute-optimal token counts, TriRoute Pareto-dominates the best independent MoD+MoE+KV-quantization combination at matched inference FLOPs and memory, while better preserving tail-case robustness on rare entities, code, and arithmetic that pure perplexity optimization erodes. Post-hoc analysis reveals interpretable structure: the controller allocates full attention and high-precision cache to sentence-initial positions, rare subwords, and named entities, while cheaply routing function words.

Looped Transformers reuse one block of layers several times: by spending extra computation they push a model of fixed size further, and so use its parameters more fully; while sparse mixture-of-experts (MoE) models activate only a few of many experts for each token. Looped MoE bridges these two design philosophies and gives MoE models new potential for better expert usage, but it raises a question: how to loop a MoE? We answer it with Foil. With the expert parameters and the expert compute per token held fixed, Foil (1) flattens the experts, halving the expert layers, doubling the experts per layer and doubling the passes, so that every routing decision chooses from a larger pool, and (2) unties the attention, giving each pass its own attention parameters while the experts and routers stay shared. Experiments show that Foil clearly outperforms the unflattened looped baseline: at 20B tokens every Foil model has lower pretraining loss than the baseline; at 100B tokens the loss improves monotonically with the degree of flattening, the most flattened Foil ending 0.012 nat below the baseline at equal parameters and compute, with downstream accuracy on par or better; untying the attention also yields more balanced and more confident routing at equal shape. Our ablations analyse why Foil works and turn the findings into design guidance for looped MoE: the returns of looping and of widening the expert layers amplify each other, routing confidence tracks healthy expert use better than load balance, and a sparse looped MoE should therefore use more experts per layer and more passes. Code and configurations are available at https://github.com/SR-A-W/how-to-loop-moe.

Unified diffusion editors often rely on a fixed, shared backbone for diverse tasks, suffering from task interference and poor adaptation to heterogeneous demands (e.g., local vs global, semantic vs photometric). In particular, prevalent ControlNet and OmniControl variants combine multiple conditioning signals (e.g., text, mask, reference) via static concatenation or additive adapters which cannot dynamically prioritize or suppress conflicting modalities, thus resulting in artifacts like color bleeding across mask boundaries, identity or style drift, and unpredictable behavior under multi-condition inputs. To address this, we propose Condition-Aware Routing of Experts (CARE-Edit) that aligns model computation with specific editing competencies. At its core, a lightweight latent-attention router assigns encoded diffusion tokens to four specialized experts--Text, Mask, Reference, and Base--based on multi-modal conditions and diffusion timesteps: (i) a Mask Repaint module first refines coarse user-defined masks for precise spatial guidance; (ii) the router applies sparse top-K selection to dynamically allocate computation to the most relevant experts; (iii) a Latent Mixture module subsequently fuses expert outputs, coherently integrating semantic, spatial, and stylistic information to the base images. Experiments validate CARE-Edit's strong performance on contextual editing tasks, including erasure, replacement, text-driven edits, and style transfer. Empirical analysis further reveals task-specific behavior of specialized experts, showcasing the importance of dynamic, condition-aware processing to mitigate multi-condition conflicts.

We introduce Xiaomi-TabLDM, a tabular large data foundation model for classification and regression via in-context learning, which delivers superior prediction accuracy without requiring task-specific fine-tuning. Pretrained exclusively on synthetic data generated from structural causal models (SCMs), our model enables more flexible context utilization and more efficient capacity scaling. i) A new performance standard. Strong regression performance across benchmarks: Xiaomi-TabLDM ranks 1st on OpenML-CTR23 and 2nd on regression across TALENT, TabArena, and BCCO, demonstrating consistently strong regression performance across four complementary benchmark suites. Favorable performance--efficiency trade-off: Xiaomi-TabLDM combines strong predictive performance with substantially lower computational cost. For example, on TabArena regression, it achieves the second-highest Elo while using 82% less training time and 68% less prediction time than the top-ranked TabFM. ii) Large-scale synthetic pretraining. Xiaomi-TabLDM expands the coverage and diversity of synthetic tabular data used for pretraining. We also adopt a three-stage training strategy together with dual-stream feature grouping, lightweight Attention Residual, and sparse Mixture-of-Experts, enabling Xiaomi-TabLDM to learn richer feature interactions and expert specialization across diverse tabular tasks. iii) Test-time scaling. Xiaomi-TabLDM further extends tabular prediction through test-time compute scaling, where allocating additional computation at inference time consistently improves predictive performance over the base model.

Scientific literature is growing exponentially, creating a critical bottleneck for researchers to efficiently synthesize knowledge. While general-purpose Large Language Models (LLMs) show potential in text processing, they often fail to capture scientific domain-specific nuances (e.g., technical jargon, methodological rigor) and struggle with complex scientific tasks, limiting their utility for interdisciplinary research. To address these gaps, this paper presents SciGPT, a domain-adapted foundation model for scientific literature understanding and ScienceBench, an open source benchmark tailored to evaluate scientific LLMs. Built on the Qwen3 architecture, SciGPT incorporates three key innovations: (1) low-cost domain distillation via a two-stage pipeline to balance performance and efficiency; (2) a Sparse Mixture-of-Experts (SMoE) attention mechanism that cuts memory consumption by 55\% for 32,000-token long-document reasoning; and (3) knowledge-aware adaptation integrating domain ontologies to bridge interdisciplinary knowledge gaps. Experimental results on ScienceBench show that SciGPT outperforms GPT-4o in core scientific tasks including sequence labeling, generation, and inference. It also exhibits strong robustness in unseen scientific tasks, validating its potential to facilitate AI-augmented scientific discovery.

Attention Model has now become an important concept in neural networks that has been researched within diverse application domains. This survey provides a structured and comprehensive overview of the developments in modeling attention. In particular, we propose a taxonomy which groups existing techniques into coherent categories. We review salient neural architectures in which attention has been incorporated, and discuss applications in which modeling attention has shown a significant impact. We also describe how attention has been used to improve the interpretability of neural networks. Finally, we discuss some future research directions in attention. We hope this survey will provide a succinct introduction to attention models and guide practitioners while developing approaches for their applications.

Sparse expert models are a thirty-year old concept re-emerging as a popular architecture in deep learning. This class of architecture encompasses Mixture-of-Experts, Switch Transformers, Routing Networks, BASE layers, and others, all with the unifying idea that each example is acted on by a subset of the parameters. By doing so, the degree of sparsity decouples the parameter count from the compute per example allowing for extremely large, but efficient models. The resulting models have demonstrated significant improvements across diverse domains such as natural language processing, computer vision, and speech recognition. We review the concept of sparse expert models, provide a basic description of the common algorithms, contextualize the advances in the deep learning era, and conclude by highlighting areas for future work.

The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Block-wise sparse attention offers significant efficiency gains for long-context modeling, yet existing methods often suffer from low selection fidelity and cumulative contextual loss by completely discarding unselected blocks. To address these limitations, we introduce Sparse Plus Linear Attention (SPLA), a framework that utilizes a selection metric derived from second-order Taylor expansions to accurately identify relevant blocks for exact attention. Instead of discarding the remaining "long tail," SPLA compresses unselected blocks into a compact recurrent state via a residual linear attention (RLA) module. Crucially, to avoid IO overhead, we derive an optimized subtraction-based formulation for RLA -- calculating the residual as the difference between global and selected linear attention -- ensuring that unselected blocks are never explicitly accessed during inference. Our experiments demonstrate that SPLA closes the performance gap in continual pretraining, surpassing dense attention models on long-context benchmarks like RULER while maintaining competitive general knowledge and reasoning capabilities.

The transformer architecture has driven breakthroughs in recent years on tasks which require modeling pairwise relationships between sequential elements, as is the case in natural language understanding. However, long seqeuences pose a problem due to the quadratic complexity of the attention operation. Previous research has aimed to lower the complexity by sparsifying or linearly approximating the attention matrix. Yet, these approaches cannot straightforwardly distill knowledge from a teacher's attention matrix and often require complete retraining from scratch. Furthermore, previous sparse and linear approaches lose interpretability if they cannot produce full attention matrices. To address these challenges, we propose SEA: Sparse linear attention with an Estimated Attention mask. SEA estimates the attention matrix with linear complexity via kernel-based linear attention, then subsequently creates a sparse attention matrix with a top-k selection to perform a sparse attention operation. For language modeling tasks (Wikitext2), previous linear and sparse attention methods show roughly two-fold worse perplexity scores over the quadratic OPT-1.3B baseline, while SEA achieves better perplexity than OPT-1.3B, using roughly half the memory of OPT-1.3B, providing interpretable attention matrix. We believe that our work will have a large practical impact, as it opens the possibility of running large transformers on resource-limited devices with less memory.

Sparse attention offers a promising strategy to extend long-context capabilities in Transformer LLMs, yet its viability, its efficiency-accuracy trade-offs, and systematic scaling studies remain unexplored. To address this gap, we perform a careful comparison of training-free sparse attention methods at varying model scales, sequence lengths, and sparsity levels on a diverse collection of long-sequence tasks-including novel ones that rely on natural language while remaining controllable and easy to evaluate. Based on our experiments, we report a series of key findings: 1) an isoFLOPS analysis reveals that for very long sequences, larger and highly sparse models are preferable to smaller and dense ones. 2) The level of sparsity attainable while statistically guaranteeing accuracy preservation is higher during decoding than prefilling, and correlates with model size in the former. 3) There is no clear strategy that performs best across tasks and phases, with different units of sparsification or budget adaptivity needed for different scenarios. Even moderate sparsity levels often result in significant performance degradation on at least one task, highlighting that sparse attention is not a universal solution. 4) We introduce and validate novel scaling laws specifically tailored for sparse attention, providing evidence that our findings are likely to hold true beyond our range of experiments. Through these insights, we demonstrate that sparse attention is a key tool to enhance the capabilities of Transformer LLMs for processing longer sequences, but requires careful evaluation of trade-offs for performance-sensitive applications.

We introduce three new attention mechanisms that outperform standard multi-head attention in terms of efficiency and learning capabilities, thereby improving the performance and broader deployability of Transformer models. Our first contribution is Optimised Attention, which performs similarly to standard attention, but has 3/4 as many parameters and one matrix multiplication fewer per head. Next, we introduce Efficient Attention, which performs on par with standard attention with only 1/2 as many parameters as many parameters and two matrix multiplications fewer per head and is up to twice as fast as standard attention. Lastly, we introduce Super Attention, which surpasses standard attention by a significant margin in both vision and natural language processing tasks while having fewer parameters and matrix multiplications. In addition to providing rigorous mathematical comparisons, we evaluate the presented attention mechanisms on MNIST, CIFAR100, IMDB Movie Reviews, and Amazon Reviews datasets.

Sparse attention visits few keys, yet finding them can still require scoring every key with every indexer head. We introduce DoSA (Dots Sparse Attention), a lossless method for making this computation conditional while preserving exact token selection. Its central observation is that Top-K selection requires sufficient evidence to reject a candidate, not necessarily its full score. A seed set supplies a lower bound on the winning threshold. Nonnegative head projections and joint residual bounds upper-bound positive contributions, while activation-informed aggregation bounds the magnitude subtracted by negative heads. The certificate directs full scoring toward candidates that remain eligible; exact-score filtering further reduces writeback and selection work. We establish the certificate and its exact-selection conditions, and analyze how geometry and selection margins govern its effectiveness. Our GPU implementation accelerates the complete Indexer by up to 2.74 times over SGLang and 1.69 times over LiteTopK across English, Chinese, and code workloads. End-to-end evaluation on GLM-5.2 and DeepSeek-V3.2 achieves up to 1.94 times and 1.50 times time-to-first-token speedups over the two baselines, respectively.

Sparse Mixture of Experts (MoE) architectures have emerged as a promising approach for scaling Transformer models. While initial works primarily incorporated MoE into feed-forward network (FFN) layers, recent studies have explored extending the MoE paradigm to attention layers to enhance model performance. However, existing attention-based MoE layers require specialized implementations and demonstrate suboptimal performance compared to their FFN-based counterparts. In this paper, we aim to unify the MoE designs in attention and FFN layers by introducing a novel reformulation of the attention mechanism, revealing an underlying FFN-like structure within attention modules. Our proposed architecture, UMoE, achieves superior performance through attention-based MoE layers while enabling efficient parameter sharing between FFN and attention components.

Mixture-of-Experts (MoE) architectures have become the dominant choice for scaling Large Language Models (LLMs), activating only a subset of parameters per token. While MoE architectures are primarily adopted for computational efficiency, it remains an open question whether their sparsity makes them inherently easier to interpret than dense feed-forward networks (FFNs). We compare MoE experts and dense FFNs using k-sparse probing and find that expert neurons are consistently less polysemantic, with the gap widening as routing becomes sparser. This suggests that sparsity pressures both individual neurons and entire experts toward monosemanticity. Leveraging this finding, we zoom out from the neuron to the expert level as a more effective unit of analysis. We validate this approach by automatically interpreting hundreds of experts. This analysis allows us to resolve the debate on specialization: experts are neither broad domain specialists (e.g., biology) nor simple token-level processors. Instead, they function as fine-grained task experts, specializing in linguistic operations or semantic tasks (e.g., closing brackets in LaTeX). Our findings suggest that MoEs are inherently interpretable at the expert level, providing a clearer path toward large-scale model interpretability. Code is available at: https://github.com/jerryy33/MoE_analysis

Sparsely activated Mixture-of-Experts (MoE) models have shown promise in enhancing the learning capacity of large language models (LLMs). Leveraging the intrinsic importance differences among experts, recent research has explored expert-level compression techniques to improve the efficiency of MoE LLMs. However, existing approaches often rely on empirical criteria to identify critical experts, lacking a deeper exploration and understanding of the heterogeneous importance of experts. In this study, we present the first discovery and investigation of a distinct subset of experts that play a crucial role in the underlying mechanisms during the model's forward inference. These experts are prevalent in open-source MoE LLMs, and despite their limited number, pruning them leads to a significant decline in model performance (e.g., pruning three causes Qwen3-30B-A3B to produce repetitive and uninformative outputs). We refer to these experts as Super Experts (SEs). Our comprehensive analysis provides progressively deeper insights into SEs. (i) SEs are characterized by rare but extreme activation outliers in the output of the down_proj, which give rise to massive activations in the hidden states between decoder layers. Moreover, the distribution of SEs remains model-specific and is unaffected by post-training processes. (ii) By pruning SEs, we assess their significance across a variety of tasks, revealing their considerable impact on the model's overall performance, particularly in mathematical reasoning. (iii) We further enhance our understanding of the influence of SEs compression. Our findings confirm that MoE LLMs rely on SEs to induce attention sinks, which are crucial for the distribution of attention scores but are significantly disrupted by SE pruning. The code is available at https://github.com/ZunhaiSu/Super-Experts-Profilling.

Self-attention has recently been adopted for a wide range of sequence modeling problems. Despite its effectiveness, self-attention suffers from quadratic compute and memory requirements with respect to sequence length. Successful approaches to reduce this complexity focused on attending to local sliding windows or a small set of locations independent of content. Our work proposes to learn dynamic sparse attention patterns that avoid allocating computation and memory to attend to content unrelated to the query of interest. This work builds upon two lines of research: it combines the modeling flexibility of prior work on content-based sparse attention with the efficiency gains from approaches based on local, temporal sparse attention. Our model, the Routing Transformer, endows self-attention with a sparse routing module based on online k-means while reducing the overall complexity of attention to Oleft(n^{1.5}dright) from Oleft(n^2dright) for sequence length n and hidden dimension d. We show that our model outperforms comparable sparse attention models on language modeling on Wikitext-103 (15.8 vs 18.3 perplexity) as well as on image generation on ImageNet-64 (3.43 vs 3.44 bits/dim) while using fewer self-attention layers. Additionally, we set a new state-of-the-art on the newly released PG-19 data-set, obtaining a test perplexity of 33.2 with a 22 layer Routing Transformer model trained on sequences of length 8192.

Sparsely Mixture of Experts (MoE) has received great interest due to its promising scaling capability with affordable computational overhead. MoE converts dense layers into sparse experts, and utilizes a gated routing network to make experts conditionally activated. However, as the number of experts grows, MoE with outrageous parameters suffers from overfitting and sparse data allocation. Such problems are especially severe on tasks with limited data, thus hindering the progress for MoE models to improve performance by scaling up. In this work, we propose Mixture of Expert Clusters - a general approach to enable expert layers to learn more diverse and appropriate knowledge by imposing variance-based constraints on the routing stage. We further propose a cluster-level expert dropout strategy specifically designed for the expert cluster structure. Our experiments reveal that MoEC could improve performance on machine translation and natural language understanding tasks, and raise the performance upper bound for scaling up experts under limited data. We also verify that MoEC plays a positive role in mitigating overfitting and sparse data allocation.

Many machine learning tasks such as multiple instance learning, 3D shape recognition, and few-shot image classification are defined on sets of instances. Since solutions to such problems do not depend on the order of elements of the set, models used to address them should be permutation invariant. We present an attention-based neural network module, the Set Transformer, specifically designed to model interactions among elements in the input set. The model consists of an encoder and a decoder, both of which rely on attention mechanisms. In an effort to reduce computational complexity, we introduce an attention scheme inspired by inducing point methods from sparse Gaussian process literature. It reduces the computation time of self-attention from quadratic to linear in the number of elements in the set. We show that our model is theoretically attractive and we evaluate it on a range of tasks, demonstrating the state-of-the-art performance compared to recent methods for set-structured data.

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

Sparsely-gated Mixture of Expert (MoE) layers have been recently successfully applied for scaling large transformers, especially for language modeling tasks. An intriguing side effect of sparse MoE layers is that they convey inherent interpretability to a model via natural expert specialization. In this work, we apply sparse MoE layers to CNNs for computer vision tasks and analyze the resulting effect on model interpretability. To stabilize MoE training, we present both soft and hard constraint-based approaches. With hard constraints, the weights of certain experts are allowed to become zero, while soft constraints balance the contribution of experts with an additional auxiliary loss. As a result, soft constraints handle expert utilization better and support the expert specialization process, while hard constraints maintain more generalized experts and increase overall model performance. Our findings demonstrate that experts can implicitly focus on individual sub-domains of the input space. For example, experts trained for CIFAR-100 image classification specialize in recognizing different domains such as flowers or animals without previous data clustering. Experiments with RetinaNet and the COCO dataset further indicate that object detection experts can also specialize in detecting objects of distinct sizes.

Large Language Models (LLMs) face efficiency bottlenecks due to the quadratic complexity of the attention mechanism when processing long contexts. Sparse attention methods offer a promising solution, but existing approaches often suffer from incomplete effective context and/or require complex implementation of pipeline. We present a comprehensive analysis of sparse attention for autoregressive LLMs from the respective of receptive field, recognize the suboptimal nature of existing methods for expanding the receptive field, and introduce PowerAttention, a novel sparse attention design that facilitates effective and complete context extension through the theoretical analysis. PowerAttention achieves exponential receptive field growth in d-layer LLMs, allowing each output token to attend to 2^d tokens, ensuring completeness and continuity of the receptive field. Experiments demonstrate that PowerAttention outperforms existing static sparse attention methods by 5sim 40%, especially on tasks demanding long-range dependencies like Passkey Retrieval and RULER, while maintaining a comparable time complexity to sliding window attention. Efficiency evaluations further highlight PowerAttention's superior speedup in both prefilling and decoding phases compared with dynamic sparse attentions and full attention (3.0times faster on 128K context), making it a highly effective and user-friendly solution for processing long sequences in LLMs.

Large language models (LLMs) have grown in their usage to provide support for question answering across numerous disciplines. The models on their own have already shown promise for answering basic questions, however fail quickly where expert domain knowledge is required or the question is nuanced. Scientific research often involves searching for relevant literature, distilling pertinent information from that literature and analysing how the findings support or contradict one another. The information is often encapsulated in the full text body of research articles, rather than just in the abstracts. Statements within these articles frequently require the wider article context to be fully understood. We have built an LLM-based system that performs such search and distillation of information encapsulated in scientific literature, and we evaluate our keyword based search and information distillation system against a set of biology related questions from previously released literature benchmarks. We demonstrate sparse retrieval methods exhibit results close to state of the art without the need for dense retrieval, with its associated infrastructure and complexity overhead. We also show how to increase the coverage of relevant documents for literature review generation.

Vision large language models (VLLMs) are focusing primarily on handling complex and fine-grained visual information by incorporating advanced vision encoders and scaling up visual models. However, these approaches face high training and inference costs, as well as challenges in extracting visual details, effectively bridging across modalities. In this work, we propose a novel visual framework, MoCHA, to address these issues. Our framework integrates four vision backbones (i.e., CLIP, SigLIP, DINOv2 and ConvNeXt) to extract complementary visual features and is equipped with a sparse Mixture of Experts Connectors (MoECs) module to dynamically select experts tailored to different visual dimensions. To mitigate redundant or insufficient use of the visual information encoded by the MoECs module, we further design a Hierarchical Group Attention (HGA) with intra- and inter-group operations and an adaptive gating strategy for encoded visual features. We train MoCHA on two mainstream LLMs (e.g., Phi2-2.7B and Vicuna-7B) and evaluate their performance across various benchmarks. Notably, MoCHA outperforms state-of-the-art open-weight models on various tasks. For example, compared to CuMo (Mistral-7B), our MoCHA (Phi2-2.7B) presents outstanding abilities to mitigate hallucination by showing improvements of 3.25% in POPE and to follow visual instructions by raising 153 points on MME. Finally, ablation studies further confirm the effectiveness and robustness of the proposed MoECs and HGA in improving the overall performance of MoCHA.

Multi-head attentive neural architectures have achieved state-of-the-art results on a variety of natural language processing tasks. Evidence has shown that they are overparameterized; attention heads can be pruned without significant performance loss. In this work, we instead "reallocate" them -- the model learns to activate different heads on different inputs. Drawing connections between multi-head attention and mixture of experts, we propose the mixture of attentive experts model (MAE). MAE is trained using a block coordinate descent algorithm that alternates between updating (1) the responsibilities of the experts and (2) their parameters. Experiments on machine translation and language modeling show that MAE outperforms strong baselines on both tasks. Particularly, on the WMT14 English to German translation dataset, MAE improves over "transformer-base" by 0.8 BLEU, with a comparable number of parameters. Our analysis shows that our model learns to specialize different experts to different inputs.

Mixture-of-Experts (MoE) models achieve a favorable trade-off between performance and inference efficiency by activating only a subset of experts. However, the memory overhead of storing all experts remains a major limitation, especially in large-scale MoE models such as DeepSeek-R1(671B). In this study, we investigate domain specialization and expert redundancy in large-scale MoE models and uncover a consistent behavior we term few-shot expert localization, with only a few in-domain demonstrations, the model consistently activates a sparse and stable subset of experts on tasks within the same domain. Building on this observation, we propose a simple yet effective pruning framework, EASY-EP, that leverages a few domain-specific demonstrations to identify and retain only the most relevant experts. EASY-EP comprises two key components: output-aware expert importance assessment and expert-level token contribution estimation. The former evaluates the importance of each expert for the current token by considering the gating scores and L2 norm of the outputs of activated experts, while the latter assesses the contribution of tokens based on representation similarities before and after routed experts. Experiments on DeepSeek-R1 and DeepSeek-V3-0324 show that our method can achieve comparable performances and 2.99times throughput under the same memory budget with full model with only half the experts.

Dual encoders perform retrieval by encoding documents and queries into dense lowdimensional vectors, scoring each document by its inner product with the query. We investigate the capacity of this architecture relative to sparse bag-of-words models and attentional neural networks. Using both theoretical and empirical analysis, we establish connections between the encoding dimension, the margin between gold and lower-ranked documents, and the document length, suggesting limitations in the capacity of fixed-length encodings to support precise retrieval of long documents. Building on these insights, we propose a simple neural model that combines the efficiency of dual encoders with some of the expressiveness of more costly attentional architectures, and explore sparse-dense hybrids to capitalize on the precision of sparse retrieval. These models outperform strong alternatives in large-scale retrieval.

While diffusion language models (DLMs) offer a promising alternative to autoregressive models (ARs), existing open-source DLMs suffer from high inference latency. This bottleneck is mainly due to the attention's quadratic complexity with respect to context length in computing all query-key pairs. Intuitively, to reduce this complexity, a natural strategy is to restrict attention to sparse patterns that retain only the most relevant connections. Such approaches are well-established in ARs, where attention follows fixed and clearly defined sparse patterns. However, in DLMs, we observe distinct sparsity behaviors: (1) attention patterns vary across heads, (2) attention patterns in each head remain highly similar across denoising steps, and (3) early denoising steps are critical for generation. These findings render sparse attention methods designed for ARs largely incompatible with DLMs, as they fail to capture head-specific structures and risk degrading generation when applied in early denoising steps. To address these challenges, we propose SparseD, a novel sparse attention method for DLMs. Leveraging the observations, SparseD only requires pre-computing head-specific sparse patterns one time, and reuses them across all steps. This prevents recomputing sparse patterns at each denoising step. Meanwhile, SparseD uses full attention in the early steps, then switches to sparse attention later to maintain generation quality. Together, these establish SparseD as a practical and efficient solution for deploying DLMs in long-context applications. Experimental results demonstrate that SparseD achieves lossless acceleration, delivering up to 1.50times speedup over FlashAttention at a 64k context length with 1,024 denoising steps.

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical token for each query using a lightweight indexer, and then computing attention only over the selected subset. While the downstream sparse attention scales efficiently, the indexer still scans the entire prefix for every query, introducing an O(L^2) per-layer bottleneck that becomes prohibitive as context length grows. We propose HISA (Hierarchical Indexed Sparse Attention), a drop-in replacement for the indexer that transforms the search process from a flat token scan into a two-stage hierarchical procedure. First, a block-level coarse filter scores pooled block representatives to prune irrelevant regions. Then, a token-level refinement applies the original indexer only within the remaining candidate blocks. HISA preserves the exact token-level top-k sparsity pattern required by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves a 2times speedup at 32K context length and 4times at 128K. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 with HISA, without any fine-tuning. HISA closely matches the original DSA in quality while significantly outperforming block-sparse baselines. Moreover, the token selection sets produced by HISA and the original DSA exhibit a mean IoU greater than 99%, indicating that the efficiency gains come with virtually no impact on selection fidelity.

Following the success of dot-product attention in Transformers, numerous approximations have been recently proposed to address its quadratic complexity with respect to the input length. However, all approximations thus far have ignored the contribution of the value vectors to the quality of approximation. In this work, we argue that research efforts should be directed towards approximating the true output of the attention sub-layer, which includes the value vectors. We propose a value-aware objective, and show theoretically and empirically that an optimal approximation of a value-aware objective substantially outperforms an optimal approximation that ignores values, in the context of language modeling. Moreover, we show that the choice of kernel function for computing attention similarity can substantially affect the quality of sparse approximations, where kernel functions that are less skewed are more affected by the value vectors.

Transformer-based architectures have become the prevailing backbone of large language models. However, the quadratic time and memory complexity of self-attention remains a fundamental obstacle to efficient long-context modeling. To address this limitation, recent research has introduced two principal categories of efficient attention mechanisms. Linear attention methods achieve linear complexity through kernel approximations, recurrent formulations, or fastweight dynamics, thereby enabling scalable inference with reduced computational overhead. Sparse attention techniques, in contrast, limit attention computation to selected subsets of tokens based on fixed patterns, block-wise routing, or clustering strategies, enhancing efficiency while preserving contextual coverage. This survey provides a systematic and comprehensive overview of these developments, integrating both algorithmic innovations and hardware-level considerations. In addition, we analyze the incorporation of efficient attention into largescale pre-trained language models, including both architectures built entirely on efficient attention and hybrid designs that combine local and global components. By aligning theoretical foundations with practical deployment strategies, this work aims to serve as a foundational reference for advancing the design of scalable and efficient language models.

The sparse Mixture-of-Experts (MoE) model is powerful for large-scale pre-training and has achieved promising results due to its model capacity. However, with trillions of parameters, MoE is hard to be deployed on cloud or mobile environment. The inference of MoE requires expert parallelism, which is not hardware-friendly and communication expensive. Especially for resource-limited downstream tasks, such sparse structure has to sacrifice a lot of computing efficiency for limited performance gains. In this work, we observe most experts contribute scarcely little to the MoE fine-tuning and inference. We further propose a general method to progressively drop the non-professional experts for the target downstream task, which preserves the benefits of MoE while reducing the MoE model into one single-expert dense model. Our experiments reveal that the fine-tuned single-expert model could preserve 99.3% benefits from MoE across six different types of tasks while enjoying 2x inference speed with free communication cost.

Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance. As shown in Figure 1, experiments show the model pretrained with NSA maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. Meanwhile, NSA achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation, validating its efficiency throughout the model lifecycle.

In this work, we conduct a systematic analysis of Native Sparse Attention (NSA) and propose targeted improvements that enhance long-context modeling. A key insight is that alternating between local (sliding-window) and global (compression, selective) attention across layers, rather than using fixed patterns, enables more effective propagation of long-range dependencies and substantially boosts performance on long-sequence tasks. Meanwhile, we further refine NSA's branches with Latent Attention that the sliding-window branch is enhanced with Multi-head Latent Attention (MLA) while compression and selective branches adopt Group-head Latent Attention (GLA). These changes reduce KV-cache memory by 50\% versus NSA while improving the model's common-sense reasoning and long-text understanding capabilities. Experiments on models from 340M to 1.3B parameters (trained on 15B and 100B tokens) show our method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks.

The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

Soft attention is a critical mechanism powering LLMs to locate relevant parts within a given context. However, individual attention weights are determined by the similarity of only a single query and key token vector. This "single token attention" bottlenecks the amount of information used in distinguishing a relevant part from the rest of the context. To address this issue, we propose a new attention method, Multi-Token Attention (MTA), which allows LLMs to condition their attention weights on multiple query and key vectors simultaneously. This is achieved by applying convolution operations over queries, keys and heads, allowing nearby queries and keys to affect each other's attention weights for more precise attention. As a result, our method can locate relevant context using richer, more nuanced information that can exceed a single vector's capacity. Through extensive evaluations, we demonstrate that MTA achieves enhanced performance on a range of popular benchmarks. Notably, it outperforms Transformer baseline models on standard language modeling tasks, and on tasks that require searching for information within long contexts, where our method's ability to leverage richer information proves particularly beneficial.

The quadratic computational complexity of the attention mechanism in current Large Language Models (LLMs) renders inference with long contexts prohibitively expensive. To address this challenge, various approaches aim to retain critical portions of the context to optimally approximate Full Attention (FA) through Key-Value (KV) compression or Sparse Attention (SA), enabling the processing of virtually unlimited text lengths in a streaming manner. However, these methods struggle to achieve performance levels comparable to FA, particularly in retrieval tasks. In this paper, our analysis of attention head patterns reveals that LLMs' attention distributions show strong local correlations, naturally reflecting a chunking mechanism for input context. We propose Ltri-LLM framework, which divides KVs into spans, stores them in an offline index, and retrieves the relevant KVs into memory for various queries. Experimental results on popular long text benchmarks show that Ltri-LLM can achieve performance close to FA while maintaining efficient, streaming-based inference.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external context, but retrieved passages are often lengthy, noisy, or exceed input limits. Existing compression methods typically require supervised training of dedicated compression models, increasing cost and reducing portability. We propose Sentinel, a lightweight sentence-level compression framework that reframes context filtering as an attention-based understanding task. Rather than training a compression model, Sentinel probes decoder attention from an off-the-shelf 0.5B proxy LLM using a lightweight classifier to identify sentence relevance. Empirically, we find that query-context relevance estimation is consistent across model scales, with 0.5B proxies closely matching the behaviors of larger models. On the LongBench benchmark, Sentinel achieves up to 5times compression while matching the QA performance of 7B-scale compression systems. Our results suggest that probing native attention signals enables fast, effective, and question-aware context compression. Code available at: https://github.com/yzhangchuck/Sentinel.

As model parameters scale up, sparsely activated Mixture-of-Experts (MoE) approaches have become essential for model scaling. However, mainstream models currently adopt hierarchical architectures, where experts in different layers, although topologically independent, exhibit substantial similarity and redundancy in their stored content. Compared to the hierarchically arranged Local Expert in existing mainstream models, we hypothesize that inter-layer shared Meta Expert hold greater potential. To this end, we conduct a multi-dimensional empirical comparison between Local and Meta Experts, and find that models incorporating Meta Experts exhibit lower parameter redundancy, together with lower cross-entropy loss and higher downstream task accuracy. With the activated quantity and the expert granularity held fixed, our model matches the performance of Vanilla MoE using only 70--80\% of the experts, substantially improving the model's scaling potential. Beyond removing inter-layer redundancy among experts and improving information storage efficiency, our theoretical analysis further shows that, compared with Local Experts, Meta Experts provide hidden states with more flexible residual pathways, thereby improving information retrieval efficiency.

Sparse Mixtures of Experts (SMoE) scales model capacity without significant increases in training and inference costs, but exhibits the following two issues: (1) Low expert activation, where only a small subset of experts are activated for optimization. (2) Lacking fine-grained analytical capabilities for multiple semantic concepts within individual tokens. We propose Multi-Head Mixture-of-Experts (MH-MoE), which employs a multi-head mechanism to split each token into multiple sub-tokens. These sub-tokens are then assigned to and processed by a diverse set of experts in parallel, and seamlessly reintegrated into the original token form. The multi-head mechanism enables the model to collectively attend to information from various representation spaces within different experts, while significantly enhances expert activation, thus deepens context understanding and alleviate overfitting. Moreover, our MH-MoE is straightforward to implement and decouples from other SMoE optimization methods, making it easy to integrate with other SMoE models for enhanced performance. Extensive experimental results across three tasks: English-focused language modeling, Multi-lingual language modeling and Masked multi-modality modeling tasks, demonstrate the effectiveness of MH-MoE.

Diffusion and flow matching models have made significant progress in text-to-image generation, yet high computation, quadratic complexity, and large memory footprint hinder high-resolution synthesis and edge deployment. We propose Nexus, which integrates sparse architecture, linear complexity, and low-bit quantization. It combines MoE feed-forward layers, gated DeltaNet attention, and per-expert low-bit training to reduce computation and memory. Their joint optimization allows Nexus to achieve generation quality comparable to mainstream models such as SDXL and SD3 while delivering markedly higher inference efficiency. Experiments on COCO and LAION validate its effectiveness.

We present MiMo-V2-Flash, a Mixture-of-Experts (MoE) model with 309B total parameters and 15B active parameters, designed for fast, strong reasoning and agentic capabilities. MiMo-V2-Flash adopts a hybrid attention architecture that interleaves Sliding Window Attention (SWA) with global attention, with a 128-token sliding window under a 5:1 hybrid ratio. The model is pre-trained on 27 trillion tokens with Multi-Token Prediction (MTP), employing a native 32k context length and subsequently extended to 256k. To efficiently scale post-training compute, MiMo-V2-Flash introduces a novel Multi-Teacher On-Policy Distillation (MOPD) paradigm. In this framework, domain-specialized teachers (e.g., trained via large-scale reinforcement learning) provide dense and token-level reward, enabling the student model to perfectly master teacher expertise. MiMo-V2-Flash rivals top-tier open-weight models such as DeepSeek-V3.2 and Kimi-K2, despite using only 1/2 and 1/3 of their total parameters, respectively. During inference, by repurposing MTP as a draft model for speculative decoding, MiMo-V2-Flash achieves up to 3.6 acceptance length and 2.6x decoding speedup with three MTP layers. We open-source both the model weights and the three-layer MTP weights to foster open research and community collaboration.

Small language models trained on a single local machine leave little room for architectural trial and error. Kairos combines bidirectional gated-DeltaNet linear attention and sliding-window attention over shared projections (LiZAttention2), block-wise Attention-Residual aggregation, a sparse Mixture-of-Experts (MoE) feed-forward layer and a multi-scale byte-level codec, trained with a uniform-noise denoising objective over one byte space shared by text, image, audio, video, lidar and control streams; this report asks whether the combination optimizes at all before asking whether it is useful. We give code-derived algorithms for each component, including a cross-session recurrent-state memory gate, then run convergence sanity checks: overfittinga single 440-byte sentence at 1--15M parameters, text only, with three seeds for codec and attention choices and single-run sweeps for seven further factors. The two attention variants tested at 3\,000 steps (LiZAttention2 and window attention alone) reach a minimum loss near 10^{-3}. After 600 steps, the mean minimum loss ranges from 0.018 (patch codec) to 0.354 (depthwise convolutional codec) with an identical ranking on three seeds. An apparent advantage of shared over separate attention projections closes with more steps or more diverse data and reverses when the projections are frozen: it is a convergence-speed effect, not a capacity effect. These checks suggest that the implementation optimizes; they say nothing about generalization, multimodal training or comparison with existing models, which a planned pretraining run is meant to address. The Discussion sizes a 200M-total/25M-active target analytically; none of its throughput figures is measured. Code is available at https://github.com/fabienfrfr/Kairos . Python package at https://pypi.org/project/kairos-fm/ .

We introduce Motif 3, a decoder-only Mixture-of-Experts language model with 314 billion total parameters and 13.2 billion activated per token. Each sparse MoE layer contains 384 routed experts, with eight selected per token. This fine-grained sparsity provides substantial expert capacity while limiting computation. Motif 3 is built around Grouped Differential Latent Attention (GDLA), which integrates grouped differential attention with the compressed key-value representation of Multi-head Latent Attention. The architecture further incorporates modified manifold-constrained hyper-connections, Expert Specific PolyNorm activations, and multi-token prediction to improve optimization stability, expert specialization, and inference efficiency. We pretrain Motif 3 on approximately 12.5 trillion tokens spanning web documents, STEM, code, mathematics, multilingual content, and domain-specialized corpora. Expert-balancing and numerical-stabilization techniques support stable training at scale, while selective MXFP8 computation and communication, memory-efficient fused kernels, and window-aware context parallelism enable training with context lengths up to 256K tokens. Our post-training pipeline combines general supervised fine-tuning, six specialist teachers trained with reinforcement learning, a software-engineering teacher trained with supervised fine-tuning, and Multi-teacher On-Policy Distillation. The resulting unified model consolidates complementary capabilities in reasoning, coding, tool use, professional work, long-context understanding, calibrated abstention, and instruction following. Across a broad evaluation suite, Motif 3 demonstrates competitive performance against leading open weight models, including strong results on long-horizon agentic tasks, mathematical reasoning, scientific knowledge, and hallucination-sensitive evaluation.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

We present the technical report for Arcee Trinity Large, a sparse Mixture-of-Experts model with 400B total parameters and 13B activated per token. Additionally, we report on Trinity Nano and Trinity Mini, with Trinity Nano having 6B total parameters with 1B activated per token, Trinity Mini having 26B total parameters with 3B activated per token. The models' modern architecture includes interleaved local and global attention, gated attention, depth-scaled sandwich norm, and sigmoid routing for Mixture-of-Experts. For Trinity Large, we also introduce a new MoE load balancing strategy titled Soft-clamped Momentum Expert Bias Updates (SMEBU). We train the models using the Muon optimizer. All three models completed training with zero loss spikes. Trinity Nano and Trinity Mini were pre-trained on 10 trillion tokens, and Trinity Large was pre-trained on 17 trillion tokens. The model checkpoints are available at https://huggingface.co/arcee-ai.

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-p selection more suitable than fixed top-k sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36times prefill speedup at 1M context and about a 2.01times decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.

Mixtral, a representative sparse mixture of experts (SMoE) language model, has received significant attention due to its unique model design and superior performance. Based on Mixtral-8x7B-v0.1, in this paper, we propose Chinese-Mixtral and Chinese-Mixtral-Instruct with improved Chinese language abilities by adopting further pre-training and instruction fine-tuning. Experimental results show that our Chinese-Mixtral and Chinese-Mixtral-Instruct successfully improve Chinese understanding and generation performance while retaining the original English abilities. Then, we discuss several key questions when performing language adaptation on large language models, including the necessity of extending the language-specific vocabulary and the choice of the initialization model (foundation model v.s. instruction model), by providing empirical results and analysis. We also present the visualizations of each expert to examine their importance on downstream tasks. Our resources are publicly available through https://github.com/ymcui/Chinese-Mixtral.

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.

Large Language Models (LLMs) have delivered impressive results in language understanding, generation, reasoning, and pushes the ability boundary of multimodal models. Transformer models, as the foundation of modern LLMs, offer a strong baseline with excellent scaling properties. However, the traditional transformer architecture requires substantial computations and poses significant obstacles for large-scale training and practical deployment. In this survey, we offer a systematic examination of innovative LLM architectures that address the inherent limitations of transformers and boost the efficiency. Starting from language modeling, this survey covers the background and technical details of linear and sparse sequence modeling methods, efficient full attention variants, sparse mixture-of-experts, hybrid model architectures incorporating the above techniques, and emerging diffusion LLMs. Additionally, we discuss applications of these techniques to other modalities and consider their wider implications for developing scalable, resource-aware foundation models. By grouping recent studies into the above category, this survey presents a blueprint of modern efficient LLM architectures, and we hope this could help motivate future research toward more efficient, versatile AI systems.

We present Top-Theta (Top-theta) Attention, a training-free method for sparsifying transformer attention during inference. Our key insight is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row. This approach enables content-based sparsity without retraining, and it remains robust across data domains. We further introduce compensation techniques to preserve accuracy under aggressive sparsification, establishing attention thresholding as a practical and principled alternative to top-k attention. We provide extensive evaluation on natural language processing tasks, showing that Top-theta achieves 3-10x reduction in V-cache usage and up to 10x fewer attention elements during inference while degrading no more than 1% in accuracy.

Large Language Models (LLMs) have driven significant progress, yet their growing parameter counts and context windows incur prohibitive compute, energy, and monetary costs. We introduce EfficientLLM, a novel benchmark and the first comprehensive empirical study evaluating efficiency techniques for LLMs at scale. Conducted on a production-class cluster (48xGH200, 8xH200 GPUs), our study systematically explores three key axes: (1) architecture pretraining (efficient attention variants: MQA, GQA, MLA, NSA; sparse Mixture-of-Experts (MoE)), (2) fine-tuning (parameter-efficient methods: LoRA, RSLoRA, DoRA), and (3) inference (quantization methods: int4, float16). We define six fine-grained metrics (Memory Utilization, Compute Utilization, Latency, Throughput, Energy Consumption, Compression Rate) to capture hardware saturation, latency-throughput balance, and carbon cost. Evaluating over 100 model-technique pairs (0.5B-72B parameters), we derive three core insights: (i) Efficiency involves quantifiable trade-offs: no single method is universally optimal; e.g., MoE reduces FLOPs and improves accuracy but increases VRAM by 40%, while int4 quantization cuts memory/energy by up to 3.9x at a 3-5% accuracy drop. (ii) Optima are task- and scale-dependent: MQA offers optimal memory-latency trade-offs for constrained devices, MLA achieves lowest perplexity for quality-critical tasks, and RSLoRA surpasses LoRA efficiency only beyond 14B parameters. (iii) Techniques generalize across modalities: we extend evaluations to Large Vision Models (Stable Diffusion 3.5, Wan 2.1) and Vision-Language Models (Qwen2.5-VL), confirming effective transferability. By open-sourcing datasets, evaluation pipelines, and leaderboards, EfficientLLM provides essential guidance for researchers and engineers navigating the efficiency-performance landscape of next-generation foundation models.

Current language models often fail to incorporate long contexts efficiently during generation. We show that a major contributor to this issue are attention priors that are likely learned during pre-training: relevant information located earlier in context is attended to less on average. Yet even when models fail to use the information from a relevant document in their response, they still pay preferential attention to that document compared to an irrelevant document at the same position. We leverage this fact to introduce ``attention sorting'': perform one step of decoding, sort documents by the attention they receive (highest attention going last), repeat the process, generate the answer with the newly sorted context. We find that attention sorting improves performance of long context models. Our findings highlight some challenges in using off-the-shelf language models for retrieval augmented generation.

This article does not propose any novel algorithm or new hardware for sparsity. Instead, it aims to serve the "common good" for the increasingly prosperous Sparse Neural Network (SNN) research community. We attempt to summarize some most common confusions in SNNs, that one may come across in various scenarios such as paper review/rebuttal and talks - many drawn from the authors' own bittersweet experiences! We feel that doing so is meaningful and timely, since the focus of SNN research is notably shifting from traditional pruning to more diverse and profound forms of sparsity before, during, and after training. The intricate relationships between their scopes, assumptions, and approaches lead to misunderstandings, for non-experts or even experts in SNNs. In response, we summarize ten Q\&As of SNNs from many key aspects, including dense vs. sparse, unstructured sparse vs. structured sparse, pruning vs. sparse training, dense-to-sparse training vs. sparse-to-sparse training, static sparsity vs. dynamic sparsity, before-training/during-training vs. post-training sparsity, and many more. We strive to provide proper and generically applicable answers to clarify those confusions to the best extent possible. We hope our summary provides useful general knowledge for people who want to enter and engage with this exciting community; and also provides some "mind of ease" convenience for SNN researchers to explain their work in the right contexts. At the very least (and perhaps as this article's most insignificant target functionality), if you are writing/planning to write a paper or rebuttal in the field of SNNs, we hope some of our answers could help you!

Block-sparse attention is promising for accelerating long-context LLM pre-filling, yet identifying relevant blocks efficiently remains a bottleneck. Existing methods typically employ coarse-grained attention as a proxy for block importance estimation, but often resort to expensive token-level searching or scoring, resulting in significant selection overhead. In this work, we trace the inaccuracy of standard coarse-grained attention via mean pooling to a theoretical root cause: the interaction between mean pooling and Rotary Positional Embeddings (RoPE). We prove that mean pooling acts as a low-pass filter that induces destructive interference in high-frequency dimensions, effectively creating a "blind spot" for local positional information (e.g., slash patterns). To address this, we introduce Prism, a training-free spectral-aware approach that decomposes block selection into high-frequency and low-frequency branches. By applying energy-based temperature calibration, Prism restores the attenuated positional signals directly from pooled representations, enabling block importance estimation using purely block-level operations, thereby improving efficiency. Extensive evaluations confirm that Prism maintains accuracy parity with full attention while delivering up to 5.1times speedup.

We rethink test-time scaling laws from a practical efficiency perspective, revealing that the effectiveness of smaller models is significantly overestimated. Prior work, grounded in compute-optimality, overlooks critical memory access bottlenecks introduced by inference-time strategies (e.g., Best-of-N, long CoTs). Our holistic analysis, spanning models from 0.6B to 32B parameters, reveals a new Kinetics Scaling Law that better guides resource allocation by incorporating both computation and memory access costs. Kinetics Scaling Law suggests that test-time compute is more effective when used on models above a threshold than smaller ones. A key reason is that in TTS, attention, rather than parameter count, emerges as the dominant cost factor. Motivated by this, we propose a new scaling paradigm centered on sparse attention, which lowers per-token cost and enables longer generations and more parallel samples within the same resource budget. Empirically, we show that sparse attention models consistently outperform dense counterparts, achieving over 60 points gains in low-cost regimes and over 5 points gains in high-cost regimes for problem-solving accuracy on AIME, encompassing evaluations on state-of-the-art MoEs. These results suggest that sparse attention is essential for realizing the full potential of test-time scaling because, unlike training, where parameter scaling saturates, test-time accuracy continues to improve through increased generation. The code is available at https://github.com/Infini-AI-Lab/Kinetics.

Inference on large language models can be expensive in terms of the compute and memory costs involved, especially when long sequence lengths are used. In particular, the self-attention mechanism used in such models contributes significantly to these costs, which has resulted in several recent works that propose sparse attention approximations for inference. In this work, we propose to approximate the self-attention computation by focusing on the dimensionality of key vectors computed in the attention block. Our analysis reveals that the key vectors lie in a significantly lower-dimensional space, consistently across several datasets and models. Exploiting this observation, we propose Loki, a novel sparse attention method that ranks and selects tokens in the KV-cache based on attention scores computed in low-dimensional space. Our evaluations show that Loki is able to maintain the efficacy of the models better than other popular approximation methods, while speeding up the attention computation due to reduced data movement (load/store) and compute costs.

The quadratic cost of attention hinders the scalability of long-context LLMs, especially in resource-constrained settings. Existing static sparse methods such as sliding windows or global tokens utilizes the sparsity of attention to reduce the cost of attention, but poorly adapts to the content-dependent variations in attention due to their staticity. While previous work has proposed several dynamic approaches to improve flexibility, they still depend on predefined templates or heuristic mechanisms. Such strategies reduce generality and prune tokens that remain contextually important, limiting their accuracy across diverse tasks. To tackle these bottlenecks of existing methods for long-context modeling, we introduce Dynamic Hierarchical Sparse Attention (DHSA), a data-driven framework that dynamically predicts attention sparsity online without retraining. Our proposed DHSA adaptively segments sequences into variable-length chunks, then computes chunk representations by aggregating the token embeddings within each chunk. To avoid the bias introduced by varying chunk lengths, we apply length-normalized aggregation that scales the averaged embeddings by the square root of the chunk size. Finally, DHSA upsamples the chunk-level similarity scores to token level similarities to calculate importance scores that determine which token-level interactions should be preserved. Our experiments on Gemma2 with Needle-in-a-Haystack Test and LongBench show that DHSA matches dense attention in accuracy, while reducing prefill latency by 20-60% and peak memory usage by 35%. Compared to other representative baselines such as block sparse attention, DHSA achieves consistently higher accuracy (6-18% relative gains) with comparable or lower cost, offering an efficient and adaptable solution for long-context on-device LLMs.

The Mixture of Experts (MoE) framework has become a popular architecture for large language models due to its superior performance over dense models. However, training MoEs from scratch in a large-scale regime is prohibitively expensive. Existing methods mitigate this by pre-training multiple dense expert models independently and using them to initialize an MoE. This is done by using experts' feed-forward network (FFN) to initialize the MoE's experts while merging other parameters. However, this method limits the reuse of dense model parameters to only the FFN layers, thereby constraining the advantages when "upcycling" these models into MoEs. We propose BAM (Branch-Attend-Mix), a simple yet effective method that addresses this shortcoming. BAM makes full use of specialized dense models by not only using their FFN to initialize the MoE layers but also leveraging experts' attention parameters fully by initializing them into a soft-variant of Mixture of Attention (MoA) layers. We explore two methods for upcycling attention parameters: 1) initializing separate attention experts from dense models including all attention parameters for the best model performance; and 2) sharing key and value parameters across all experts to facilitate for better inference efficiency. To further improve efficiency, we adopt a parallel attention transformer architecture to MoEs, which allows the attention experts and FFN experts to be computed concurrently. Our experiments on seed models ranging from 590 million to 2 billion parameters demonstrate that BAM surpasses baselines in both perplexity and downstream task performance, within the same computational and data constraints.

Accommodating long sequences efficiently in autoregressive Transformers, especially within an extended context window, poses significant challenges due to the quadratic computational complexity and substantial KV memory requirements inherent in self-attention mechanisms. In this work, we introduce SPARSEK Attention, a novel sparse attention mechanism designed to overcome these computational and memory obstacles while maintaining performance. Our approach integrates a scoring network and a differentiable top-k mask operator, SPARSEK, to select a constant number of KV pairs for each query, thereby enabling gradient-based optimization. As a result, SPARSEK Attention offers linear time complexity and constant memory footprint during generation. Experimental results reveal that SPARSEK Attention outperforms previous sparse attention methods and provides significant speed improvements during both training and inference, particularly in language modeling and downstream tasks. Furthermore, our method can be seamlessly integrated into pre-trained Large Language Models (LLMs) with minimal fine-tuning, offering a practical solution for effectively managing long-range dependencies in diverse applications.

The quadratic complexity of self-attention during the prefill phase impedes long-context inference in large language models. Existing sparse attention methods face a trade-off among context adaptivity, sampling overhead, and fine-tuning costs. We propose VSPrefill, a mechanism requiring lightweight training that uses the vertical-slash structural pattern in attention distributions. Our compact VSIndexer module predicts context-aware importance scores for vertical columns and slash diagonals from key-value representations augmented with RoPE. This approach constructs sparse masks with linear complexity without modifying the backbone parameters. During inference, an adaptive cumulative-threshold strategy allocates sparsity budgets per layer, while a fused kernel executes attention with on-the-fly index merging. Evaluated on Qwen3-4B-Instruct and LLaMA-3.1-8B-Instruct across the LongBench and RULER benchmarks, VSPrefill preserves 98.35% of the full attention accuracy while delivering a 4.95x average speedup at a context length of 128k. These results establish a new Pareto frontier in the trade-off between accuracy and efficiency.

Long-context modeling is a pivotal capability for Large Language Models, yet the quadratic complexity of attention remains a critical bottleneck, particularly during the compute-intensive prefilling phase. While various sparse attention mechanisms have been explored, they typically suffer from either significant search latency or insufficient sparsity. In this paper, we propose FlashPrefill, a framework enabling ultra-fast prefilling via instantaneous pattern discovery and thresholding. FlashPrefill leverages a fast block-searching technique to simultaneously locate dynamic vertical, slash, and block-sparse attention patterns. Crucially, it introduces a dynamic thresholding mechanism that bypasses the prohibitive overhead of sorting or accumulating attention scores while effectively eliminating the long-tail distribution to enhance sparsity. Extensive evaluations demonstrate that FlashPrefill achieves a substantial leap in efficiency, delivering an unprecedented 27.78x speedup on 256K sequences. Notably, unlike existing methods that incur efficiency degradation on shorter contexts, FlashPrefill maintains a 1.71x speedup even at a 4K context length, demonstrating its robustness and practical utility across varying sequence scales.

Transformers are powerful sequence models, but require time and memory that grows quadratically with the sequence length. In this paper we introduce sparse factorizations of the attention matrix which reduce this to O(n n). We also introduce a) a variation on architecture and initialization to train deeper networks, b) the recomputation of attention matrices to save memory, and c) fast attention kernels for training. We call networks with these changes Sparse Transformers, and show they can model sequences tens of thousands of timesteps long using hundreds of layers. We use the same architecture to model images, audio, and text from raw bytes, setting a new state of the art for density modeling of Enwik8, CIFAR-10, and ImageNet-64. We generate unconditional samples that demonstrate global coherence and great diversity, and show it is possible in principle to use self-attention to model sequences of length one million or more.

Attention is the cornerstone of modern Large Language Models (LLMs). Yet its quadratic complexity limits the efficiency and scalability of LLMs, especially for those with a long-context window. A promising approach addressing this limitation is to leverage the sparsity in attention. However, existing sparsity-based solutions predominantly rely on predefined patterns or heuristics to approximate sparsity. This practice falls short to fully capture the dynamic nature of attention sparsity in language-based tasks. This paper argues that attention sparsity should be learned rather than predefined. To this end, we design SeerAttention, a new Attention mechanism that augments the conventional attention with a learnable gate that adaptively selects significant blocks in an attention map and deems the rest blocks sparse. Such block-level sparsity effectively balances accuracy and speedup. To enable efficient learning of the gating network, we develop a customized FlashAttention implementation that extracts the block-level ground truth of attention map with minimum overhead. SeerAttention not only applies to post-training, but also excels in long-context fine-tuning. Our results show that at post-training stages, SeerAttention significantly outperforms state-of-the-art static or heuristic-based sparse attention methods, while also being more versatile and flexible to adapt to varying context lengths and sparsity ratios. When applied to long-context fine-tuning with YaRN, SeerAttention can achieve a remarkable 90% sparsity ratio at a 32k context length with minimal perplexity loss, offering a 5.67x speedup over FlashAttention-2.

Some attention heads learn similar patterns across inputs. Reusing these patterns could reduce training cost by avoiding repeated query-key score computation and softmax. Through controlled pretraining comparisons, we identify Selective Attention Freezing (SAF), which selects heads with low attention-pattern variance and replaces their attention weights with fitted post-softmax means halfway through training. We represent these fixed patterns with absolute-position and relative-distance preferences, reducing storage from quadratic to linear in sequence length. A fused kernel reconstructs the patterns and executes ordinary-attention and replaced heads together. At matched training-token budgets, replacing 25% of attention heads gives 1.056x faster post-replacement optimiser updates at 124M parameters and 4K context, with a 0.77% perplexity increase. At 1B and 8K context, post-replacement updates are 1.068x faster on four GPUs including communication, with a 0.51% perplexity increase. The resulting models also accelerate long-input finetuning and causal prefill. After associative-recall adaptation, the 124M model with 25% replacement generalises to more key-value pairs at a fixed length better than ordinary attention and two pruning controls.

Long-context understanding is crucial for many NLP applications, yet transformers struggle with efficiency due to the quadratic complexity of self-attention. Sparse attention methods alleviate this cost but often impose static, predefined masks, failing to capture heterogeneous attention patterns. This results in suboptimal token interactions, limiting adaptability and retrieval accuracy in long-sequence tasks. This work introduces a dynamic sparse attention mechanism that assigns adaptive masks at the attention-map level, preserving heterogeneous patterns across layers and heads. Unlike existing approaches, our method eliminates the need for fine-tuning and predefined mask structures while maintaining computational efficiency. By learning context-aware attention structures, it achieves high alignment with full-attention models, ensuring minimal performance degradation while reducing memory and compute overhead. This approach provides a scalable alternative to full attention, enabling the practical deployment of large-scale Large Language Models (LLMs) without sacrificing retrieval performance. DAM is available at: https://github.com/HanzhiZhang-Ulrica/DAM.

The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the standard attention mechanism. To address this challenge, we introduce BLASST, a drop-in sparse attention method that dynamically prunes the attention matrix without any pre-computation or proxy scores. Our method uses a fixed threshold and existing information from online softmax to identify negligible attention scores, skipping softmax computation, Value block loading, and the subsequent matrix multiplication. This fits seamlessly into existing FlashAttention kernel designs with negligible latency overhead. The approach is applicable to both prefill and decode stages across all attention variants (MHA, GQA, MQA, and MLA), providing a unified solution for accelerating long-context inference. We develop an automated calibration procedure that reveals a simple inverse relationship between optimal threshold and context length, enabling robust deployment across diverse scenarios. Maintaining high accuracy, we demonstrate a 1.62x speedup for prefill at 74.7% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs. Furthermore, we explore sparsity-aware training as a natural extension, showing that models can be trained to be inherently more robust to sparse attention patterns, pushing the accuracy-sparsity frontier even further.

Standard attention scales quadratically with sequence length. Efficient attention methods reduce this O(n^2) cost, but when retrofitted into pretrained models, they often degrade perplexity, downstream accuracy, or both. We introduce Focus, a method that learns which token pairs matter. Focus adds a small set of learnable centroids--as few as 148K parameters per layer--that act as gates: only token pairs belonging to the same centroid group attend to each other over long ranges. Focus is composable: it can be added to any pretrained model by training only the centroids while keeping all original weights frozen. Experiments show that composing Focus onto pretrained models yields zero degradation on downstream benchmarks across model sizes from 124M to 70B parameters and five attention architectures. Surprisingly, sparse Focus attention outperforms full attention at 124M scale (30.3 vs. 31.4 perplexity) and matches full attention when trained from scratch at 7B scale (13.82 vs. 13.89). Focus is also fast: top-k group membership gives a 2x speedup with better quality than the original pretrained model. Using our FlashAttention decomposition, Focus achieves an 8.6x speedup at 1M tokens without custom kernels.

Block sparse attention is a hardware friendly way to alleviate the key-value (KV) cache read bottleneck in large language models (LLMs). However, it is not prevalent among leading open-weight LLMs, which rely instead on dense attention or fine-grained selection, thereby motivating our analysis. We study DeepSeek's Native Sparse Attention (NSA) as a representative method, whose three-branch design lets us isolate block selection, the most challenging and consequential stage. We formalize selection and reduce it to ranking blocks by a single quantity, the attention mass: the sum of a block's attention scores. We show that if selection retrieves the blocks with the largest attention mass, block sparse attention can match the quality of dense attention. However, computing the exact attention mass requires reading every key, so the problem of block selection ultimately reduces to approximating this mass from a compact summary instead of the full keys. Via a cumulant expansion, we show why existing methods falter: their selection strategies attempt to estimate the attention mass, but are confined to a first-order approximation. Therefore, we propose COBS (Cumulant Order Block Sparse Attention), an attention method that builds on NSA, incorporating a novel selector that stores a compressed second-order statistic per block. On the 32k RULER long-context retrieval benchmark, COBS raises the NSA baseline's mean score from 0.2999 to 0.8195, approaching dense attention at 0.9040 and closing about 86% of the gap, while using only 1.21x the KV cache read traffic of the NSA baseline and 15.15x less read traffic than dense. The same model preserves short-context behavior and attains lower position-wise negative log-likelihood (NLL) than dense attention in our comparison.

Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold. FlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference.

Large reasoning models achieve strong performance through test-time scaling but incur substantial computational overhead, particularly from excessive token generation when processing short input prompts. While sparse attention mechanisms can reduce latency and memory usage, existing approaches suffer from significant accuracy degradation due to accumulated errors during long-generation reasoning. These methods generally require either high token retention rates or expensive retraining. We introduce LessIsMore, a training-free sparse attention mechanism for reasoning tasks, which leverages global attention patterns rather than relying on traditional head-specific local optimizations. LessIsMore aggregates token selections from local attention heads with recent contextual information, enabling unified cross-head token ranking for future decoding layers. This unified selection improves generalization and efficiency by avoiding the need to maintain separate token subsets per head. Evaluation across diverse reasoning tasks and benchmarks shows that LessIsMore preserves -- and in some cases improves -- accuracy while achieving a 1.1times average decoding speed-up compared to full attention. Moreover, LessIsMore attends to 2times fewer tokens without accuracy loss, achieving a 1.13times end-to-end speed-up compared to existing sparse attention methods.

Attention-based mechanisms are widely used in machine learning, most prominently in transformers. However, hyperparameters such as the rank of the attention matrices and the number of heads are scaled nearly the same way in all realizations of this architecture, without theoretical justification. In this work we show that there are dramatic trade-offs between the rank and number of heads of the attention mechanism. Specifically, we present a simple and natural target function that can be represented using a single full-rank attention head for any context length, but that cannot be approximated by low-rank attention unless the number of heads is exponential in the embedding dimension, even for short context lengths. Moreover, we prove that, for short context lengths, adding depth allows the target to be approximated by low-rank attention. For long contexts, we conjecture that full-rank attention is necessary. Finally, we present experiments with off-the-shelf transformers that validate our theoretical findings.

Long-context LLM inference is bottlenecked by quadratic attention computation and growing KV-cache costs. Existing sparse attention and KV-compression methods typically decide which tokens or heads to preserve from runtime attention scores, observation windows, calibration prompts, or learned gates, making head diagnosis input-dependent and costly to deploy. We propose Autonomy-of-Heads (AoH), a data-free method that identifies retrieval and streaming heads from the spectral geometry of query-key projections. AoH defines the kernel attention operator M_h = W_K^{htop}W_Q^h and uses its effective-rank as a weight-space measure of head function: concentrated spectra indicate a small number of dominant query-key matching directions and are associated with retrieval heads, whereas diffuse spectra indicate the absence of a dominant global matching direction and are associated with streaming heads. We further derive an efficient d_head-dimensional computation that avoids constructing the full d_modeltimes d_model matrix. We conducted extensive experiments across models demonstrating that at 50\% sparsity, AoH retains 96.5\% of Full Attention performance on average while reducing prefill and decode latency by up to 41.4\% and 66.0\%, respectively, and KV-cache memory by 50.0\% at 256K tokens.

As large language models are increasingly deployed in retrieval-augmented generation and agentic systems that accumulate extensive context, understanding how distracting information affects long-context performance becomes critical. Prior work shows that semantically relevant yet misleading documents degrade performance, but the quantitative relationship between the proportion of distractors and performance remains unstudied. In this work, we systematically vary the hard-distractor proportion in fixed-length contexts, revealing a striking nonlinear pattern: as the proportion of hard distractors increases, performance drops sharply within the first small fraction, while the remainder of the range yields only marginal additional decline. We term this ''The First Drop of Ink'' effect, analogous to how a single drop of ink contaminates water. Our theoretical and empirical analyses grounded in attention mechanics show that hard distractors capture disproportionate attention even at small proportions, with diminishing marginal impact as their proportion grows. Controlled experiments further show that filtering gains mainly come from context-length reduction rather than distractor removal; substantial recovery requires reducing the hard-distractor proportion to near zero, highlighting the importance of upstream retrieval precision.

Motivated by the factorization inherent in the original fast multipole method and the improved fast Gauss transform we introduce a factorable form of attention that operates efficiently in high dimensions. This approach reduces the computational and memory complexity of the attention mechanism in transformers from O(N^2) to O(N). In comparison to previous attempts, our work presents a linearly scaled attention mechanism that maintains the full representation of the attention matrix without compromising on sparsification and incorporates the all-to-all relationship between tokens. We explore the properties of our new attention metric and conduct tests in various standard settings. Results indicate that our attention mechanism has a robust performance and holds significant promise for diverse applications where self-attention is used.

Large language models have demonstrated exceptional performance across a wide range of tasks. However, dense models usually suffer from sparse activation, where many activation values tend towards zero (i.e., being inactivated). We argue that this could restrict the efficient exploration of model representation space. To mitigate this issue, we propose Finedeep, a deep-layered fine-grained expert architecture for dense models. Our framework partitions the feed-forward neural network layers of traditional dense models into small experts, arranges them across multiple sub-layers. A novel routing mechanism is proposed to determine each expert's contribution. We conduct extensive experiments across various model sizes, demonstrating that our approach significantly outperforms traditional dense architectures in terms of perplexity and benchmark performance while maintaining a comparable number of parameters and floating-point operations. Moreover, we find that Finedeep achieves optimal results when balancing depth and width, specifically by adjusting the number of expert sub-layers and the number of experts per sub-layer. Empirical results confirm that Finedeep effectively alleviates sparse activation and efficiently utilizes representation capacity in dense models.

Attention is a powerful and ubiquitous mechanism for allowing neural models to focus on particular salient pieces of information by taking their weighted average when making predictions. In particular, multi-headed attention is a driving force behind many recent state-of-the-art NLP models such as Transformer-based MT models and BERT. These models apply multiple attention mechanisms in parallel, with each attention "head" potentially focusing on different parts of the input, which makes it possible to express sophisticated functions beyond the simple weighted average. In this paper we make the surprising observation that even if models have been trained using multiple heads, in practice, a large percentage of attention heads can be removed at test time without significantly impacting performance. In fact, some layers can even be reduced to a single head. We further examine greedy algorithms for pruning down models, and the potential speed, memory efficiency, and accuracy improvements obtainable therefrom. Finally, we analyze the results with respect to which parts of the model are more reliant on having multiple heads, and provide precursory evidence that training dynamics play a role in the gains provided by multi-head attention.

Transformer encoders are widely deployed in large-scale web services for natural language understanding tasks such as text classification, semantic retrieval, and content ranking. However, their high inference latency and memory consumption pose significant challenges for real-time serving and scalability. These limitations stem largely from architectural redundancy, particularly in the attention module. The inherent parameter redundancy of the attention mechanism, coupled with the fact that its attention heads operate with a degree of independence, makes it particularly amenable to structured model compression. In this paper, we propose SHRP (Specialized Head Routing and Pruning), a novel structured pruning framework that automatically identifies and removes redundant attention heads while preserving most of the model's accuracy and compatibility. SHRP introduces Expert Attention, a modular design that treats each attention head as an independent expert, followed by a lightweight shared expander feed-forward network that refines their outputs. The framework employs a unified Top-1 usage-driven mechanism to jointly perform dynamic routing during training and deterministic pruning at deployment. Experimental results on the GLUE benchmark using a BERT-base encoder show that SHRP achieves 93% of the original model accuracy while reducing parameters by 48 percent. Under an extreme compression scenario where 11/12 of the layers are pruned, the model still maintains 84% accuracy and delivers a 4.2x throughput gain while reducing computation to as low as 11.5 percent of the original FLOPs, demonstrating its practical utility for large-scale and latency-sensitive web deployments.

Sparse Mixture-of-Experts (MoE) is a neural architecture design that can be utilized to add learnable parameters to Large Language Models (LLMs) without increasing inference cost. Instruction tuning is a technique for training LLMs to follow instructions. We advocate combining these two approaches, as we find that MoE models benefit more from instruction tuning than dense models. In particular, we conduct empirical studies across three experimental setups: (i) Direct finetuning on individual downstream tasks devoid of instruction tuning; (ii) Instructiontuning followed by in-context few-shot or zero-shot generalization on downstream tasks; and (iii) Instruction tuning supplemented by further finetuning on individual downstream tasks. In the first scenario, MoE models overall underperform dense models of identical computational capacity. This narrative, however, dramatically changes with the introduction of instruction tuning (second and third scenario), used independently or in conjunction with task-specific finetuning. Our most powerful model, FLAN-MOE-32B, surpasses the performance of FLAN-PALM-62B on four benchmark tasks, while using only a third of the FLOPs. The advancements embodied byFLAN-MOE inspire a reevaluation of the design principles of large-scale, high-performance language models in the framework of task-agnostic learning.

Large reasoning models (LRMs) achieve state-of-the-art performance on challenging benchmarks by generating long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence. Existing sparse attention methods reduce computation by pruning the key-value (KV) cache, yet they suffer from severe accuracy degradation on reasoning tasks due to cumulative selection errors and the dynamic importance of tokens over long derivations. We present DELTA, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy. DELTA partitions transformer layers into three groups: initial layers that use full attention, a small set of selection layers that identify salient tokens via aggregated head-level attention scores, and subsequent sparse-attention layers that attend only to the selected subset. This design preserves the full KV cache in GPU memory for accuracy, while avoiding expensive full-attention computation over many layers. On reasoning benchmarks such as AIME and GPQA-Diamond, DELTA matches or surpasses full attention in accuracy, while reducing the number of attended tokens by up to 5times and delivering 1.5times end-to-end speedup. Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.

This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.

Transformer-based architectures traditionally employ softmax to compute attention weights, which produces dense distributions over all tokens in a sequence. While effective in many settings, this density has been shown to be detrimental for tasks that demand precise focus on fixed-size patterns: as sequence length increases, non-informative tokens accumulate attention probability mass, leading to dispersion and representational collapse. We show in this paper that dynamically sparse attention mechanisms using α-entmax can avoid these issues, due to their ability to assign exact zeros to irrelevant tokens. Furthermore, we introduce Adaptive-Scalable Entmax (ASEntmax), which endows α-entmax with a learnable temperature parameter, allowing the attention distribution to interpolate between sparse (pattern-focused) and dense (softmax-like) regimes. Our empirical evaluation on synthetic tasks and language modeling demonstrates that ASEntmax substantially outperforms softmax, scalable softmax, and fixed-temperature α-entmax baselines, achieving up to 1000times length extrapolation on synthetic benchmarks and superior long-context generalization on language modeling while preserving short-context performance, including better perplexity trends and higher retrieval accuracies at 8times training length.

Retrieval Augmented Generation faces a trade-off: concatenating documents in a long prompt enables multi-document reasoning but creates prefill bottlenecks, while encoding document KV caches separately offers speed but breaks cross-document interaction. We propose Parallel Context-of-Experts Decoding (Pced), a training-free framework that shifts evidence aggregation from the attention mechanism to the decoding. Pced treats retrieved documents as isolated "experts", synchronizing their predictions via a novel retrieval-aware contrastive decoding rule that weighs expert logits against the model prior. This approach recovers cross-document reasoning capabilities without constructing a shared attention across documents.

Sparsity allows scaling model parameters without proportionally increasing computational cost. While mixture of experts (MoE) models are made increasingly sparse, individual experts typically remain large and dense. Here, we demonstrate that further increasing sparsity by shrinking each expert to consist of a single neuron and selecting a tiny fraction of many available neurons can improve compute efficiency and interpretability. Counterintuitively, the key to achieving both is removing the nonlinearity typically applied to the experts, resulting in a network of sparsely gated linear neurons (sgatlin). In an isoflop comparison, we find that replacing all transformer feedforward layers with sgatlin improves perplexity in language models across different compute budgets. At the same time, the sparsity and linearity of the resulting feedforward circuits present new opportunities for model interpretability. In a small-scale case study, we demonstrate that feedforward circuits in sgatlin can be interpreted without having to train additional replacement models. We find that they form semantically structured clusters and are causally implicated in factual recall. Our findings paint a possible path towards compute-efficient and interpretable transformer feedforward layers.

Large language models (LLMs) now support extremely long context windows, but the quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. Existing approaches to address this complexity require additional pretraining or finetuning, and often sacrifice model accuracy. In this paper, we first provide both theoretical and empirical foundations for near-lossless sparse attention. We find dynamically capturing head-specific sparse patterns at runtime with low overhead is crucial. To address this, we propose SampleAttention, an adaptive structured and near-lossless sparse attention. Leveraging observed significant sparse patterns, SampleAttention attends to a fixed percentage of adjacent tokens to capture local window patterns, and employs a two-stage query-guided key-value filtering approach, which adaptively select a minimum set of key-values with low overhead, to capture column stripe patterns. Comprehensive evaluations show that SampleAttention can seamlessly replace vanilla attention in off-the-shelf LLMs with nearly no accuracy loss, and reduces TTFT by up to 2.42times compared with FlashAttention.

The Mixture-of-Experts (MoE) architecture enables a significant increase in the total number of model parameters with minimal computational overhead. However, it is not clear what performance tradeoffs, if any, exist between MoEs and standard dense transformers. In this paper, we show that as we increase the number of experts (while fixing the number of active parameters), the memorization performance consistently increases while the reasoning capabilities saturate. We begin by analyzing the theoretical limitations of MoEs at reasoning. We prove that there exist graph problems that cannot be solved by any number of experts of a certain width; however, the same task can be easily solved by a dense model with a slightly larger width. On the other hand, we find that on memory-intensive tasks, MoEs can effectively leverage a small number of active parameters with a large number of experts to memorize the data. We empirically validate these findings on synthetic graph problems and memory-intensive closed book retrieval tasks. Lastly, we pre-train a series of MoEs and dense transformers and evaluate them on commonly used benchmarks in math and natural language. We find that increasing the number of experts helps solve knowledge-intensive tasks, but fails to yield the same benefits for reasoning tasks.

Full self-attention in large language models scales as O(N^2), which limits long-context document analysis to 65,536 tokens and requires costly GPU clusters. The Reduced Interaction Sampling (RIS) inference engine addresses this constraint as a model-agnostic architecture. Without modifying weights, RIS reduces self-attention complexity to O(N log N) using sparse stochastic geometry that fits within commodity memory limits. We validate RIS on Qwen2-1.5B-Instruct across two regimes. In controlled evaluations at 32,768 tokens (where native dense attention serves as the upper bound), RIS-Stochastic at 1% density and 70 ensemble seeds achieves 75.00% accuracy, outperforming the native dense baseline (71.88%), while RIS-Stochastic at 5% density and 10 seeds matches it (71.88%). This demonstrates that sparse attention acts as a regularizer: low density (1%) over multiple seeds filters out sequence-level noise, whereas higher density (5%) reintroduces distractor noise. Under the tightest budget, RIS-Structural reaches 68.75% accuracy at 1% density with just 10 seeds, recovering 75% of the contextual gap relative to the zero-context floor (59.38%). At 65,536 tokens, where dense attention triggers out-of-memory faults, RIS yields retrieval gains of up to 14.06 percentage points over the zero-context floor (51.56%), which is confirmed as marginally significant under McNemar's paired test (p = 0.078 < 0.10). All evaluations run on commodity, unaccelerated CPU servers (16-128 GB of RAM), demonstrating that long-context LLM inference is feasible on standard academic hardware without GPU acceleration.

The feedforward (FFW) layers in standard transformer architectures incur a linear increase in computational costs and activation memory as the hidden layer width grows. Sparse mixture-of-experts (MoE) architectures have emerged as a viable approach to address this issue by decoupling model size from computational cost. The recent discovery of the fine-grained MoE scaling law shows that higher granularity leads to better performance. However, existing MoE models are limited to a small number of experts due to computational and optimization challenges. This paper introduces PEER (parameter efficient expert retrieval), a novel layer design that utilizes the product key technique for sparse retrieval from a vast pool of tiny experts (over a million). Experiments on language modeling tasks demonstrate that PEER layers outperform dense FFWs and coarse-grained MoEs in terms of performance-compute trade-off. By enabling efficient utilization of a massive number of experts, PEER unlocks the potential for further scaling of transformer models while maintaining computational efficiency.

Unified image generation and editing models suffer from severe task interference in dense diffusion transformers architectures, where a shared parameter space must compromise between conflicting objectives (e.g., local editing v.s. subject-driven generation). While the sparse Mixture-of-Experts (MoE) paradigm is a promising solution, its gating networks remain task-agnostic, operating based on local features, unaware of global task intent. This task-agnostic nature prevents meaningful specialization and fails to resolve the underlying task interference. In this paper, we propose a novel framework to inject semantic intent into MoE routing. We introduce a Hierarchical Task Semantic Annotation scheme to create structured task descriptors (e.g., scope, type, preservation). We then design Predictive Alignment Regularization to align internal routing decisions with the task's high-level semantics. This regularization evolves the gating network from a task-agnostic executor to a dispatch center. Our model effectively mitigates task interference, outperforming dense baselines in fidelity and quality, and our analysis shows that experts naturally develop clear and semantically correlated specializations.

Mixture-of-Experts (MoE) architectures scale model capacity through sparse computation, routing each token through only a small subset of experts. In this work, we explore whether this sparsity gives rise to emergent intrinsic organization in multimodal MoEs. We find that experts develop strong semantic specialization across modalities and domains despite not being explicitly trained for modularity. Building on this structure, we introduce ExpertLens, a data-free method that identifies domain-specialized experts directly from pretrained model weights by decoding router weights into semantically meaningful vocabulary tokens. We leverage this specialization for efficient multimodal adaptation by selectively fine-tuning experts relevant to a target domain. Across math, medical, and remote sensing tasks, ExpertLens matches or surpasses full fine-tuning while updating only 21.7 - 47.0% of model parameters and achieving a 4.0x average training speedup, and outperforms LoRA in both adaptation performance and training efficiency. These results show that sparsity introduced for efficiency can give rise to semantic modularity that is directly useful for efficient adaptation.

Sparse Attention Vectors (SAVs) have emerged as an excellent training-free alternative to supervised finetuning or low-rank adaptation to improve the performance of Vision Language Models (VLMs). At their heart, SAVs select a few accurate attention heads for a task of interest and use them as classifiers, rather than relying on the model's prediction. In a similar spirit, we find that directly probing the raw activations of the VLM, in the form of scalar values, is sufficient to yield accurate classifiers on diverse visually grounded downstream tasks. Shifting focus from attention vectors to scalar activations dramatically increases the search space for accurate parameters, allowing us to find more discriminative neurons immediately from the first generated token. We call such activations Super Neurons (SNs). In this probing setting, we discover that enough SNs appear in the shallower layers of the large language model to allow for extreme early exiting from the first layer of the model at the first generated token. Compared to the original network, SNs robustly improve the classification performance while achieving a speedup of up to 5.10x.

Recent developments in representational learning for information retrieval can be organized in a conceptual framework that establishes two pairs of contrasts: sparse vs. dense representations and unsupervised vs. learned representations. Sparse learned representations can further be decomposed into expansion and term weighting components. This framework allows us to understand the relationship between recently proposed techniques such as DPR, ANCE, DeepCT, DeepImpact, and COIL, and furthermore, gaps revealed by our analysis point to "low hanging fruit" in terms of techniques that have yet to be explored. We present a novel technique dubbed "uniCOIL", a simple extension of COIL that achieves to our knowledge the current state-of-the-art in sparse retrieval on the popular MS MARCO passage ranking dataset. Our implementation using the Anserini IR toolkit is built on the Lucene search library and thus fully compatible with standard inverted indexes.

The Transformer architecture, underpinned by the Multi-Head Attention (MHA) mechanism, has become the de facto standard for state-of-the-art models in artificial intelligence. However, the quadratic computational complexity of MHA with respect to sequence length presents a significant barrier to scaling, particularly for applications involving long contexts. Prevailing solutions, such as Multi-Query Attention (MQA) and Grouped-Query Attention (GQA), have effectively addressed the memory bandwidth bottleneck that dominates autoregressive inference latency by sharing Key and Value projections. While highly successful, these methods do not reduce the fundamental number of floating-point operations (FLOPs) required for the attention score computation, which remains a critical bottleneck for training and full-sequence processing. This paper introduces Sparse Query Attention (SQA), a novel attention architecture that pursues an alternative and complementary optimization path. Instead of reducing Key/Value heads, SQA reduces the number of Query heads. This architectural modification directly decreases the computational complexity of the attention mechanism by a factor proportional to the reduction in query heads, thereby lowering the overall FLOPs. This work presents the theoretical foundation of SQA, its mathematical formulation, and a family of architectural variants. Empirical benchmarks on long sequences (32k-200k tokens) demonstrate that SQA can achieve significant throughput improvements of up to 3x in computation-bound scenarios such as model pre-training, fine-tuning, and encoder-based tasks, with only a minimal impact on model quality in preliminary smallscale experiments. SQA was discovered serendipitously during the development of the upcoming Reactive Transformer architecture, suggesting its potential as a powerful tool for building more efficient and scalable models

The attention mechanism of a transformer has a quadratic complexity, leading to high inference costs and latency for long sequences. However, attention matrices are mostly sparse, which implies that many entries may be omitted from computation for efficient inference. Sparse attention inference methods aim to reduce this computational burden; however, they also come with a troublesome performance degradation. We discover that one reason for this degradation is that the sparse calculation induces a distributional shift in the attention outputs. The distributional shift causes decoding-time queries to fail to align well with the appropriate keys from the prefill stage, leading to a drop in performance. We propose a simple, novel, and effective procedure for correcting this distributional shift, bringing the distribution of sparse attention outputs closer to that of quadratic attention. Our method can be applied on top of any sparse attention method, and results in an average 36%pt performance increase, recovering 88% of quadratic attention accuracy on the 131K RULER benchmark when applied on top of sliding window attention with sink tokens while only adding a small overhead. Our method can maintain approximately 98.5% sparsity over full quadratic attention, making our model 32 times faster than Flash Attention 2 when processing 1M token prefills.

The pursuit of reducing the memory footprint of the self-attention mechanism in multi-headed self attention (MHA) spawned a rich portfolio of methods, e.g., group-query attention (GQA) and multi-head latent attention (MLA). The methods leverage specialized low-rank factorizations across embedding dimensions or attention heads. From the point of view of classical low-rank approximation, these methods are unconventional and raise questions of which objects they really approximate and how to interpret the low-rank behavior of the resulting representations. To answer these questions, this work proposes a generalized view on the weight objects in the self-attention layer and a factorization strategy, which allows us to construct a parameter efficient scheme, called Tucker Attention. Tucker Attention requires an order of magnitude fewer parameters for comparable validation metrics, compared to GQA and MLA, as evaluated in LLM and ViT test cases. Additionally, Tucker Attention~encompasses GQA, MLA, MHA as special cases and is fully compatible with flash-attention and rotary position embeddings (RoPE). This generalization strategy yields insights of the actual ranks achieved by MHA, GQA, and MLA, and further enables simplifications for MLA.

In-Context Learning (ICL) enables large language models (LLMs) to achieve rapid task adaptation by learning from demonstrations. With the increase in available context length of LLMs, recent experiments have shown that the performance of ICL does not necessarily scale well in many-shot (demonstration) settings. We theoretically and experimentally confirm that the reason lies in more demonstrations dispersing the model attention from the query, hindering its understanding of key content. Inspired by how humans learn from examples, we propose a training-free method FocusICL, which conducts triviality filtering to avoid attention being diverted by unimportant contents at token-level and operates hierarchical attention to further ensure sufficient attention towards current query at demonstration-level. We also design an efficient hyperparameter searching strategy for FocusICL based on model perplexity of demonstrations. Comprehensive experiments validate that FocusICL achieves an average performance improvement of 5.2% over vanilla ICL and scales well with many-shot demonstrations.

In modern large language models (LLMs), increasing sequence lengths is a crucial challenge for enhancing their comprehension and coherence in handling complex tasks such as multi-modal question answering. However, handling long context sequences with LLMs is prohibitively costly due to the conventional attention mechanism's quadratic time and space complexity, and the context window size is limited by the GPU memory. Although recent works have proposed linear and sparse attention mechanisms to address this issue, their real-world applicability is often limited by the need to re-train pre-trained models. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which simultaneously reduces the training and inference time complexity from O(T^2) to O(T log T) and the space complexity from O(T^2) to O(T). To this end, we devise a dynamic sparse attention mechanism that generates an attention mask through a novel tree-search-like algorithm for a given query on the fly. HiP is training-free as it only utilizes the pre-trained attention scores to spot the positions of the top-k most significant elements for each query. Moreover, it ensures that no token is overlooked, unlike the sliding window-based sub-quadratic attention methods, such as StreamingLLM. Extensive experiments on diverse real-world benchmarks demonstrate that HiP significantly reduces prompt (i.e., prefill) and decoding latency and memory usage while maintaining high generation performance with little or no degradation. As HiP allows pretrained LLMs to scale to millions of tokens on commodity GPUs with no additional engineering due to its easy plug-and-play deployment, we believe that our work will have a large practical impact, opening up the possibility to many long-context LLM applications previously infeasible.

Neural networks -- especially those that use large, pre-trained language models -- have improved search engines in various ways. Most prominently, they can estimate the relevance of a passage or document to a user's query. In this work, we depart from this direction by exploring whether neural networks can effectively predict which of a document's passages are unlikely to be relevant to any query submitted to the search engine. We refer to this query-agnostic estimation of passage relevance as a passage's quality. We find that our novel methods for estimating passage quality allow passage corpora to be pruned considerably while maintaining statistically equivalent effectiveness; our best methods can consistently prune >25% of passages in a corpora, across various retrieval pipelines. Such substantial pruning reduces the operating costs of neural search engines in terms of computing resources, power usage, and carbon footprint -- both when processing queries (thanks to a smaller index size) and when indexing (lightweight models can prune low-quality passages prior to the costly dense or learned sparse encoding step). This work sets the stage for developing more advanced neural "learning-what-to-index" methods.

The ability to process long contexts is crucial for many natural language processing tasks, yet it remains a significant challenge. While substantial progress has been made in enhancing the efficiency of attention mechanisms, there is still a gap in understanding how attention heads function in long-context settings. In this paper, we observe that while certain heads consistently attend to local information only, others swing between attending to local and long-context information depending on the query. This raises the question: can we identify which heads require long-context information to predict the next token accurately? We demonstrate that it's possible to predict which heads are crucial for long-context processing using only local keys. The core idea here is to exploit a simple model for the long-context scores via second moment approximations. These findings unveil simple properties of attention in the context of long sequences, and open the door to potentially significant gains in efficiency.

Transformer-based Mixture-of-Experts (MoE) models have been driving several recent technological advancements in Natural Language Processing (NLP). These MoE models adopt a router mechanism to determine which experts to activate for routing input tokens. However, existing router mechanisms allocate a fixed number of experts to each token, which neglects the varying importance of different input tokens. In this study, we propose a novel dynamic router mechanism that Dynamically Allocates a variable number of experts for Mixture-of-Experts (DA-MoE) models based on an effective token importance measure. First, we show that the Transformer attention mechanism provides a natural and effective way of calculating token importance. Second, we propose a dynamic router mechanism that effectively decides the optimal number of experts (K) and allocates the top-K experts for each input token. Third, comprehensive experiments on several benchmark datasets demonstrate that our DA-MoE approach consistently outperforms the state-of-the-art Transformer based MoE model on the popular GLUE benchmark.

Long-context large language models (LLMs) are prone to be distracted by irrelevant contexts. The reason for distraction remains poorly understood. In this paper, we first identify the contextual heads, a special group of attention heads that control the overall attention of the LLM. Then, we demonstrate that distraction arises when contextual heads fail to allocate sufficient attention to relevant contexts and can be mitigated by increasing attention to these contexts. We further identify focus directions, located at the key and query activations of these heads, which enable them to allocate more attention to relevant contexts without explicitly specifying which context is relevant. We comprehensively evaluate the effect of focus direction on various long-context tasks and find out focus directions could help to mitigate the poor task alignment of the long-context LLMs. We believe our findings could promote further research on long-context LLM alignment.

Scaling the size of language models usually leads to remarkable advancements in NLP tasks. But it often comes with a price of growing computational cost. Although a sparse Mixture of Experts (MoE) can reduce the cost by activating a small subset of parameters (e.g., one expert) for each input, its computation escalates significantly if increasing the number of activated experts, limiting its practical utility. Can we retain the advantages of adding more experts without substantially increasing the computational costs? In this paper, we first demonstrate the superiority of selecting multiple experts and then propose a computation-efficient approach called \texttt{Merging Experts into One} (MEO), which reduces the computation cost to that of a single expert. Extensive experiments show that MEO significantly improves computational efficiency, e.g., FLOPS drops from 72.0G of vanilla MoE to 28.6G (MEO). Moreover, we propose a token-level attention block that further enhances the efficiency and performance of token-level MEO, e.g., 83.3\% (MEO) vs. 82.6\% (vanilla MoE) average score on the GLUE benchmark. Our code will be released upon acceptance. Code will be released at: https://github.com/Shwai-He/MEO.

Block diffusion LLMs are emerging as a promising next paradigm for language generation, but their use of KV caching makes memory access a dominant bottleneck in long-context settings. While dynamic sparse attention has been actively explored, existing methods designed for autoregressive LLMs rely on approximate importance estimation and perform poorly when adapted to block diffusion. This work identifies a key opportunity unique to block diffusion: attention at the first All-[MASK] denoising step reliably predicts important KV entries and budget requirements, enabling MAGE to perform a single exact attention pass per block and reuse it for training-free sparse denoising. Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse attention baselines. A lightweight fine-tuning strategy further strengthens [MASK]-guided patterns with minimal cost, requiring only a few hours of training on a single NVIDIA H100 GPU for both 1.5B and 7B models.

Sparse Mixture of Experts (sMoE) has become a pivotal approach for scaling large vision-language models, offering substantial capacity while maintaining computational efficiency through dynamic, sparse activation of experts. However, existing routing mechanisms, typically based on similarity scoring, struggle to effectively capture the underlying input structure. This limitation leads to a trade-off between expert specialization and balanced computation, hindering both scalability and performance. We propose Input Domain Aware MoE, a novel routing framework that leverages a probabilistic mixture model to better partition the input space. By modeling routing probabilities as a mixture of distributions, our method enables experts to develop clear specialization boundaries while achieving balanced utilization. Unlike conventional approaches, our routing mechanism is trained independently of task-specific objectives, allowing for stable optimization and decisive expert assignments. Empirical results on vision-language tasks demonstrate that our method consistently outperforms existing sMoE approaches, achieving higher task performance and improved expert utilization balance.

Attention often concentrates on a small subset of tokens in the context, but which subset matters changes from one query to the next. To exploit this changing structure, we introduce SANTA++, a training-free stochastic attention method that uses representative keys for memory-efficient selection without scanning the entire key-value (KV) cache. Cached keys are organized into teams, and the query scores one representative from each team to decide which teams to sample. We compute exact attention scores within the sampled teams and reweight each team's contribution by the inverse of its inclusion probability. This importance sampling correction estimates attention over the full cache, with a sampling budget that lets us trade memory reads for accuracy. Remarkably, with 32 or 64 sampled teams, SANTA++ uses 16% to 22% of dense attention's KV reads and retains 94% to 99% of the dense-attention baseline's scores on LongBench v2 and HELMET's retrieval-augmented generation subset, and 85% to 91% on RULER, with Qwen2.5-7B-Instruct at 32K context. With 31 sampled teams, our GPU implementation delivers a 1.69times attention speedup over the dense FlashAttention baseline at 32K context. By reducing the number of cache entries read, SANTA++ in principle complements architectures with compressed KV representations, such as multi-head latent attention. Our kernels are available at: https://github.com/OPUSLab/santapp-kernel-demo.git.

Parameter Efficient Tuning has been an prominent approach to adapt the Large Language Model to downstream tasks. Most previous works considers adding the dense trainable parameters, where all parameters are used to adapt certain task. We found this less effective empirically using the example of LoRA that introducing more trainable parameters does not help. Motivated by this we investigate the importance of leveraging "sparse" computation and propose SiRA: sparse mixture of low rank adaption. SiRA leverages the Sparse Mixture of Expert(SMoE) to boost the performance of LoRA. Specifically it enforces the top k experts routing with a capacity limit restricting the maximum number of tokens each expert can process. We propose a novel and simple expert dropout on top of gating network to reduce the over-fitting issue. Through extensive experiments, we verify SiRA performs better than LoRA and other mixture of expert approaches across different single tasks and multitask settings.

Existing LLM training and inference frameworks struggle in boosting efficiency with sparsity while maintaining the integrity of context and model architecture. Inspired by the sharding concept in database and the fact that attention parallelizes over heads on accelerators, we propose Sparsely-Sharded (S2) Attention, an attention algorithm that allocates heterogeneous context partitions for different attention heads to divide and conquer. S2-Attention enforces each attention head to only attend to a partition of contexts following a strided sparsity pattern, while the full context is preserved as the union of all the shards. As attention heads are processed in separate thread blocks, the context reduction for each head can thus produce end-to-end speed-up and memory reduction. At inference, LLMs trained with S2-Attention can then take the KV cache reduction as free meals with guaranteed model quality preserve. In experiments, we show S2-Attentioncan provide as much as (1) 25.3X wall-clock attention speed-up over FlashAttention-2, resulting in 6X reduction in end-to-end training time and 10X inference latency, (2) on-par model training quality compared to default attention, (3)perfect needle retrieval accuracy over 32K context window. On top of the algorithm, we build DKernel, an LLM training and inference kernel library that allows users to customize sparsity patterns for their own models. We open-sourced DKerneland make it compatible with Megatron, Pytorch, and vLLM.

Attention mechanisms have become ubiquitous in NLP. Recent architectures, notably the Transformer, learn powerful context-aware word representations through layered, multi-headed attention. The multiple heads learn diverse types of word relationships. However, with standard softmax attention, all attention heads are dense, assigning a non-zero weight to all context words. In this work, we introduce the adaptively sparse Transformer, wherein attention heads have flexible, context-dependent sparsity patterns. This sparsity is accomplished by replacing softmax with alpha-entmax: a differentiable generalization of softmax that allows low-scoring words to receive precisely zero weight. Moreover, we derive a method to automatically learn the alpha parameter -- which controls the shape and sparsity of alpha-entmax -- allowing attention heads to choose between focused or spread-out behavior. Our adaptively sparse Transformer improves interpretability and head diversity when compared to softmax Transformers on machine translation datasets. Findings of the quantitative and qualitative analysis of our approach include that heads in different layers learn different sparsity preferences and tend to be more diverse in their attention distributions than softmax Transformers. Furthermore, at no cost in accuracy, sparsity in attention heads helps to uncover different head specializations.