Daily Papers

In retrieval-augmented generation and agentic LLM serving, prompts are assembled from independent segments into long contexts, making the prefill stage dominate per-request cost. Two directions have emerged to reduce this cost: position-independent caching (PIC) admits KV reuse for non-contiguous segments shared across requests, while hybrid-attention models cut computation by replacing most full-attention layers with linear attention. However, they cannot coexist: applying existing PIC methods to hybrid-attention models breaks down because per-token KV-cache reuse primitives do not transfer to the per-request recurrent state. We present Hypic, the first system to accelerate hybrid-attention LLM serving with position-independent caching. For linear-attention layers, we identify the segment-cumulative transition operator as the missing algebraic primitive and cache it alongside each segment's zero-start end-state, enabling near-exact and constant-time composition of independently cached segments. For the remaining full-attention layers, existing PIC methods also fail because linear layers do not expose the per-token hidden states needed for selective recomputation. We show that the largest deviations concentrate at segment beginnings and construct a small seam window that propagates hidden states through the hybrid-attention stack to repair cross-segment attention. Finally, Hypic introduces segment parallelism, which exploits PIC's segment-level self-containment to parallelize cache-miss prefill across instances, turning long cold requests into an accelerable workload. Evaluated across four hybrid-attention models and five workloads, Hypic reduces time-to-first-token by 3.25times on average and improves QPS by 1.66times over Prefix Cache, while preserving task quality with a 1.71-point gap from Full Recompute.

We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.

Scene Graph Generation, which generally follows a regular encoder-decoder pipeline, aims to first encode the visual contents within the given image and then parse them into a compact summary graph. Existing SGG approaches generally not only neglect the insufficient modality fusion between vision and language, but also fail to provide informative predicates due to the biased relationship predictions, leading SGG far from practical. Towards this end, in this paper, we first present a novel Stacked Hybrid-Attention network, which facilitates the intra-modal refinement as well as the inter-modal interaction, to serve as the encoder. We then devise an innovative Group Collaborative Learning strategy to optimize the decoder. Particularly, based upon the observation that the recognition capability of one classifier is limited towards an extremely unbalanced dataset, we first deploy a group of classifiers that are expert in distinguishing different subsets of classes, and then cooperatively optimize them from two aspects to promote the unbiased SGG. Experiments conducted on VG and GQA datasets demonstrate that, we not only establish a new state-of-the-art in the unbiased metric, but also nearly double the performance compared with two baselines.

Video diffusion models repeatedly process long spatiotemporal token sequences during denoising, making attention a major computational bottleneck. Linear attention offers an appealing alternative and has been widely adopted in recent large language models, but directly applying it to video models often fails to preserve the fine-grained interactions required for high-quality generation. We present Video DeltaNet (VDN), which combines local Softmax attention with bidirectional linear memory for long-range video context. Its linear branch introduces Video Delta Attention (VDA), which updates memory once per frame by jointly incorporating its spatial tokens. Separate output projections and learnable gates calibrate the two branches, while a staged teacher-alignment recipe progressively introduces the new pathway into pretrained models. We instantiate VDN on MiniMax H3, applying the hybrid to video-to-video interactions while retaining Softmax for interactions involving text or audio. With eight-step distillation and an optimized SGLang serving stack, VDN-H3 completes DiT denoising for a 14.3-second, 768p video in 6.70 seconds on eight NVIDIA B200 GPUs, corresponding to a 14.5x speedup over the 50-step dense H3 baseline on the same GPU count.

We introduce SANA-Video 2.0, a hybrid video diffusion transformer instantiated at 5B and 14B scales under a unified architecture. Designed to generate high-quality video up to 720p on a single GPU, SANA-Video 2.0 matches full-softmax video DiTs in quality while retaining the favorable long-sequence scaling of linear attention. To avoid quadratic attention throughout, Hybrid Linear-Softmax Attention combines gated linear attention for O(N)-dominated mixing with periodic gated-softmax anchors at a 3:1 ratio, restoring the full-rank token interactions that pure linear attention lacks. To propagate these refreshed representations across depth, Block Attention Residuals (AttnRes) route completed block summaries into later linear layers, enabling anchor-feature reuse and boosting deep-layer effective rank by ~12%. Through from-scratch training, SANA-Video 2.0 learns the complete hybrid directly rather than linearizing pretrained models, with reduced-resolution proxy studies establishing 25% softmax as the optimal quality-efficiency trade-off. With 40-step sampling, SANA-Video 2.0 achieves a VBench score of 84.30 in 13.2s at 480p on a single H100, remaining competitive with far larger softmax video DiTs at a fraction of the latency. Its compiled DiT forward pass is 3.2x faster than a matched full-softmax baseline at 720p/60s, a gap that expands with video duration. Furthermore, full-stack Sol-Engine optimization (kernel fusion, caching, and sparse attention) accelerates this hardware-friendly backbone by a further 3.58x, bringing the 5B pipeline to 13.06s at 720p/5s and making it 120x faster than Wan 2.2-A14B on one H100. Overall, our hybrid design recovers softmax-level expressiveness at substantially reduced cost, unlocking scalable long, high resolution video generation.

World models are transitioning from passive visual generators to foundational, operational infrastructure for Physical AI: they must natively acquire world knowledge from heterogeneous experience, maintain persistent states over long horizons, and execute efficiently within real deployment constraints. We introduce Kairos, a native world model stack designed around these requirements. (1) Kairos learns the world by pioneering a Native Pre-training Paradigm governed by a Cross-Embodiment Data Curriculum, which organizes open-world videos, human behavioral data, and robot interactions into a progressive developmental pathway. (2) Kairos maintains the world by unified world understanding, generation, and prediction within a Native Unified Architecture equipped with Hybrid Linear Temporal Attention, where sliding-window attention captures local dynamics, dilated sliding windows capture mid-range dependencies, and gated linear attention maintains persistent global memory. We establish formal theoretical bounds demonstrating that this temporal factorization strictly limits error accumulation, mathematically guaranteeing state propagation across extended horizons. (3) Kairos runs the world by incorporating a Deployment-Aware System Co-Design to support low-latency rollout generation on server and consumer-grade hardware for real-world observation-action-feedback loops. Experiments on embodied world-model, long-horizon, and action-policy benchmarks show that Kairos achieves top level performance while offering a strong efficiency-capability trade-off. Together, these results position Kairos as a cohesive operational foundation for future self-evolving physical intelligence.

Transformer-based methods have demonstrated excellent performance on super-resolution visual tasks, surpassing conventional convolutional neural networks. However, existing work typically restricts self-attention computation to non-overlapping windows to save computational costs. This means that Transformer-based networks can only use input information from a limited spatial range. Therefore, a novel Hybrid Multi-Axis Aggregation network (HMA) is proposed in this paper to exploit feature potential information better. HMA is constructed by stacking Residual Hybrid Transformer Blocks(RHTB) and Grid Attention Blocks(GAB). On the one side, RHTB combines channel attention and self-attention to enhance non-local feature fusion and produce more attractive visual results. Conversely, GAB is used in cross-domain information interaction to jointly model similar features and obtain a larger perceptual field. For the super-resolution task in the training phase, a novel pre-training method is designed to enhance the model representation capabilities further and validate the proposed model's effectiveness through many experiments. The experimental results show that HMA outperforms the state-of-the-art methods on the benchmark dataset. We provide code and models at https://github.com/korouuuuu/HMA.

Convolutional networks, transformers, hybrid models, and Mamba-based architectures have demonstrated strong performance across various medical image classification tasks. However, these methods were primarily designed to classify clean images using labeled data. In contrast, real-world clinical data often involve image corruptions that are unique to multi-center studies and stem from variations in imaging equipment across manufacturers. In this paper, we introduce the Medical Vision Transformer (MedViTV2), a novel architecture incorporating Kolmogorov-Arnold Network (KAN) layers into the transformer architecture for the first time, aiming for generalized medical image classification. We have developed an efficient KAN block to reduce computational load while enhancing the accuracy of the original MedViT. Additionally, to counteract the fragility of our MedViT when scaled up, we propose an enhanced Dilated Neighborhood Attention (DiNA), an adaptation of the efficient fused dot-product attention kernel capable of capturing global context and expanding receptive fields to scale the model effectively and addressing feature collapse issues. Moreover, a hierarchical hybrid strategy is introduced to stack our Local Feature Perception and Global Feature Perception blocks in an efficient manner, which balances local and global feature perceptions to boost performance. Extensive experiments on 17 medical image classification datasets and 12 corrupted medical image datasets demonstrate that MedViTV2 achieved state-of-the-art results in 27 out of 29 experiments with reduced computational complexity. MedViTV2 is 44\% more computationally efficient than the previous version and significantly enhances accuracy, achieving improvements of 4.6\% on MedMNIST, 5.8\% on NonMNIST, and 13.4\% on the MedMNIST-C benchmark.

Self-attention lets each token retrieve information from the full context, but its quadratic cost in sequence length limits training and inference at long context. This paper presents a comparative study of softmax attention and four recent recurrent linear-attention architectures: DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2. We express these mechanisms in a common recurrent-memory notation, making explicit how they differ in expressivity, memory decay, erase and write control, training throughput, and implementation complexity. Our experiments center on 350M-parameter models trained for 15B tokens, and include optimizer and learning-rate comparisons, hybrid-versus-pure stack comparisons, sequence-length runtime measurements, larger DeltaNet runs at 1.3B and 3B parameters, and a small set of downstream evaluations. The reported speed results measure training throughput and iteration time; we do not provide an empirical inference-speed benchmark. Within the reported 350M-parameter, 15B-token sweep, Kimi Delta Attention with Muon reaches the lowest final validation loss, a pure Gated DeltaNet stack trained with AdamW has the highest normalized training throughput, hybrid stacks generally improve loss at a throughput cost, and Muon consistently lowers final validation loss relative to AdamW in the matched architecture settings we evaluate. We introduce and evaluate lightweight cross-layer routing mechanisms for DeltaNet-style memories. The most natural DeltaNet-inspired formulation, forwarding a lower layer's delta-rule write error into the next layer's value target, does not improve over matched baselines. Routing into the aligned hidden stream and forwarding the write value instead yields a modest improvement in the matched runs we report: Cross-Layer Value Routing (CLVR) lowers final validation loss for both DeltaNet and Gated DeltaNet.

Transformers have attracted increasing interests in computer vision, but they still fall behind state-of-the-art convolutional networks. In this work, we show that while Transformers tend to have larger model capacity, their generalization can be worse than convolutional networks due to the lack of the right inductive bias. To effectively combine the strengths from both architectures, we present CoAtNets(pronounced "coat" nets), a family of hybrid models built from two key insights: (1) depthwise Convolution and self-Attention can be naturally unified via simple relative attention; (2) vertically stacking convolution layers and attention layers in a principled way is surprisingly effective in improving generalization, capacity and efficiency. Experiments show that our CoAtNets achieve state-of-the-art performance under different resource constraints across various datasets: Without extra data, CoAtNet achieves 86.0% ImageNet top-1 accuracy; When pre-trained with 13M images from ImageNet-21K, our CoAtNet achieves 88.56% top-1 accuracy, matching ViT-huge pre-trained with 300M images from JFT-300M while using 23x less data; Notably, when we further scale up CoAtNet with JFT-3B, it achieves 90.88% top-1 accuracy on ImageNet, establishing a new state-of-the-art result.

Recent advancements in areas such as natural language processing and computer vision rely on intricate and massive models that have been trained using vast amounts of unlabelled or partly labeled data and training or deploying these state-of-the-art methods to resource constraint environments has been a challenge. Galaxy morphologies are crucial to understanding the processes by which galaxies form and evolve. Efficient methods to classify galaxy morphologies are required to extract physical information from modern-day astronomy surveys. In this paper, we introduce Astroformer, a method to learn from less amount of data. We propose using a hybrid transformer-convolutional architecture drawing much inspiration from the success of CoAtNet and MaxViT. Concretely, we use the transformer-convolutional hybrid with a new stack design for the network, a different way of creating a relative self-attention layer, and pair it with a careful selection of data augmentation and regularization techniques. Our approach sets a new state-of-the-art on predicting galaxy morphologies from images on the Galaxy10 DECals dataset, a science objective, which consists of 17736 labeled images achieving 94.86% top-1 accuracy, beating the current state-of-the-art for this task by 4.62%. Furthermore, this approach also sets a new state-of-the-art on CIFAR-100 and Tiny ImageNet. We also find that models and training methods used for larger datasets would often not work very well in the low-data regime.

The quadratic complexity of attention poses a critical bottleneck for long-context processing, spurring interest in hybrid attention designs. Most open-source hybrid models adopt a layer-wise strategy. Yet, prior work has noted the inherent difficulty of integrating Linear Attention (LA) with Full Attention (FA), suggesting that the design space of attention hybridization remains underexplored. To probe this space, we conduct interpretability analysis and observe that layers exhibit block-wise functional similarity, while individual heads within the same layer display distinct functional specialization despite sharing input features. This head-level heterogeneity suggests that the head dimension provides a natural and principled granularity for fusing heterogeneous attention signals. Building on this insight, we introduce HydraHead, a novel architecture that hybridizes FA and LA along the head axis. HydraHead features two key innovations: (1) an interpretability-driven selection strategy that identifies retrieval-critical heads and preserves FA only for them, and (2) a scale-normalized fusion module that reconciles the distributional gap between FA and LA head outputs. By leveraging a three-stage transfer pipeline with parameter reuse and distillation, we achieve high-performance hybrid models with minimal training overhead. Under a unified training setup, HydraHead outperforms other hybrid designs in long-context tasks while maintaining strong general reasoning. With interpretability-driven head selection, it matches a 3:1 layer-wise hybrid's long-context performance at a 7:1 LA-to-FA ratio. Crucially, trained on only 15B tokens, HydraHead achieves over 69% improvement over the baseline at 512K context length, approaching Qwen3.5, a leading model of comparable size with a native context length of 256K. This highlights the significant scaling potential of head-level hybridization.

Modern language models increasingly adopt hybrid architectures that combine full attention with efficient attention modules, such as sliding-window attention (SWA) and recurrent sequence mixers. However, how these efficient modules shape model capabilities remains poorly understood. To address this gap, we conduct a systematic analysis across hybrid architectures from three perspectives: scaling behavior, mechanism analysis, and architecture design. First, from a scaling perspective, we find that efficient-attention design primarily affects how fast long-context capability emerges, while different hybrids eventually converge to comparable long-context performance under sufficient training. Second, mechanistically, we show that long-range retrieval is mainly carried by full attention, whereas efficient attention shapes its optimization trajectory. This explains a counter-intuitive phenomenon we call Large-Window Laziness: larger SWA windows can delay the formation of retrieval heads in full-attention layers. Third, guided by this mechanism, we show that applying NoPE to only the full-attention layers of a small-window SWA hybrid substantially improves long-context performance with negligible impact on short-context performance.

Transformers excel at sequence modeling but face quadratic complexity, while linear attention offers improved efficiency but often compromises recall accuracy over long contexts. In this work, we introduce Native Hybrid Attention (NHA), a novel hybrid architecture of linear and full attention that integrates both intra \& inter-layer hybridization into a unified layer design. NHA maintains long-term context in key-value slots updated by a linear RNN, and augments them with short-term tokens from a sliding window. A single softmax attention operation is then applied over all keys and values, enabling per-token and per-head context-dependent weighting without requiring additional fusion parameters. The inter-layer behavior is controlled through a single hyperparameter, the sliding window size, which allows smooth adjustment between purely linear and full attention while keeping all layers structurally uniform. Experimental results show that NHA surpasses Transformers and other hybrid baselines on recall-intensive and commonsense reasoning tasks. Furthermore, pretrained LLMs can be structurally hybridized with NHA, achieving competitive accuracy while delivering significant efficiency gains. Code is available at https://github.com/JusenD/NHA.

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Attention, specifically scaled dot-product attention, has proven effective for natural language, but it does not have a mechanism for handling hierarchical patterns of arbitrary nesting depth, which limits its ability to recognize certain syntactic structures. To address this shortcoming, we propose stack attention: an attention operator that incorporates stacks, inspired by their theoretical connections to context-free languages (CFLs). We show that stack attention is analogous to standard attention, but with a latent model of syntax that requires no syntactic supervision. We propose two variants: one related to deterministic pushdown automata (PDAs) and one based on nondeterministic PDAs, which allows transformers to recognize arbitrary CFLs. We show that transformers with stack attention are very effective at learning CFLs that standard transformers struggle on, achieving strong results on a CFL with theoretically maximal parsing difficulty. We also show that stack attention is more effective at natural language modeling under a constrained parameter budget, and we include results on machine translation.

Hybrid attention models improve long-context efficiency by retaining only a subset of full-attention layers and replacing the remaining layers with linear attention. However, the effectiveness of Transformer-to-hybrid conversion critically depends on which layers preserve full attention. Existing hybrid layer selection methods typically rely on heuristic strategies such as fixed placement patterns or layerwise scoring, implicitly treating layer importance as isolated and overlooking the interdependent layer effect under a global hybrid configuration. In this work, we formulate hybrid layer selection as a budget-constrained subset optimization problem. We further propose FlashMorph (Fast LAyer Selection for Hybrid MORPHing), an effective, efficient and scalable layer selection method for Transformer-to-hybrid conversion. FlashMorph first constructs a morphable model by equipping each full-attention layer with a converted linear-attention branch. It then freezes all model weights and jointly optimizes layerwise gates on synthetic long-context retrieval data, with a linearization regularization that encourages the model to rely on linear attention for efficiency. The learned gates are discretized under a preset full-attention budget to instantiate the hybrid architecture, followed by standard logits distillation and long-context finetuning. Extensive experiments show that FlashMorph discovers more effective hybrid configurations, preserves strong long-context recall and general benchmark performance while substantially reducing layer selection cost compared with existing layer selection methods, demonstrating its effectiveness, efficiency, and scalability.

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.

Hybrid attention architectures are becoming an increasingly important paradigm for improving LLM inference efficiency while preserving model quality, making hybrid architecture design a central problem. Existing designs often rely on manual empirical rules or proxy-based selector signals for layer-wise operator allocation. Recent NAS-style systems such as Jet-Nemotron demonstrate the promise of automated hybrid architecture search. However, Jet-Nemotron's PostNAS search stages alone use 200B tokens, making such search pipelines difficult to use as routine methods for hybrid architecture design. We introduce DASH, a fast differentiable search framework for hybrid attention architecture design, which relaxes discrete layer-wise attention operator placement into continuous architecture logits, prepares reusable teacher-aligned linear candidates, and performs architecture-only search with model and operator weights frozen to significantly enhance search efficiency. On Qwen2.5-3B-Instruct, DASH consistently outperforms a comprehensive suite of existing selector-style hybrid attention design baselines, showing that direct differentiable search can discover stronger hybrid architectures. Moreover, DASH achieves stronger RULER performance than released Jet-Nemotron models while remaining competitive on overlapping short-context and general benchmarks. Notably, each DASH search run uses only 12.3M tokens and takes about 20 minutes on a single RTX Pro 6000 GPU, corresponding to merely 0.006% of the PostNAS search tokens reported by Jet-Nemotron. These results suggest that high-quality hybrid attention architectures can be obtained through minutes-level differentiable search, providing a promising direction for hybrid architecture design.

Hybrid language models that interleave attention with recurrent components are increasingly competitive with pure Transformers, yet standard LoRA practice applies adapters uniformly without considering the distinct functional roles of each component type. We systematically study component-type LoRA placement across two hybrid architectures -- Qwen3.5-0.8B (sequential, GatedDeltaNet + softmax attention) and Falcon-H1-0.5B (parallel, Mamba-2 SSM + attention) -- fine-tuned on three domains and evaluated on five benchmarks. We find that the attention pathway -- despite being the minority component -- consistently outperforms full-model adaptation with 5-10x fewer trainable parameters. Crucially, adapting the recurrent backbone is destructive in sequential hybrids (-14.8 pp on GSM8K) but constructive in parallel ones (+8.6 pp). We further document a transfer asymmetry: parallel hybrids exhibit positive cross-task transfer while sequential hybrids suffer catastrophic forgetting. These results establish that hybrid topology fundamentally determines adaptation response, and that component-aware LoRA placement is a necessary design dimension for hybrid architectures.

Recent progress in large language models demonstrates that hybrid architectures--combining self-attention mechanisms with structured state space models like Mamba--can achieve a compelling balance between modeling quality and computational efficiency, particularly for long-context tasks. While these hybrid models show promising performance, systematic comparisons of hybridization strategies and analyses on the key factors behind their effectiveness have not been clearly shared to the community. In this work, we present a holistic evaluation of hybrid architectures based on inter-layer (sequential) or intra-layer (parallel) fusion. We evaluate these designs from a variety of perspectives: language modeling performance, long-context capabilities, scaling analysis, and training and inference efficiency. By investigating the core characteristics of their computational primitive, we identify the most critical elements for each hybridization strategy and further propose optimal design recipes for both hybrid models. Our comprehensive analysis provides practical guidance and valuable insights for developing hybrid language models, facilitating the optimization of architectural configurations.

The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8timesA800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to 2.8times and 2.0times in the prefill and decode stages.

The quadratic complexity of standard attention mechanisms poses a significant scalability bottleneck for large language models (LLMs) in long-context scenarios. While hybrid attention strategies that combine sparse and full attention within a single model offer a viable solution, they typically employ static computation ratios (i.e., fixed proportions of sparse versus full attention) and fail to adapt to the varying sparsity sensitivities of downstream tasks during inference. To address this issue, we propose Elastic Attention, which allows the model to dynamically adjust its overall sparsity based on the input. This is achieved by integrating a lightweight Attention Router into the existing pretrained model, which dynamically assigns each attention head to different computation modes. Within only 12 hours of training on 8xA800 GPUs, our method enables models to achieve both strong performance and efficient inference. Experiments across three long-context benchmarks on widely-used LLMs demonstrate the superiority of our method.

Linear-attention models that compress the entire input sequence into a fixed-size recurrent state offer an efficient alternative to Transformers, but their finite memory induces forgetfulness that harms retrieval-intensive tasks. To mitigate the issue, we explore a series of hybrid models that restore direct access to past tokens. We interleave token mixers with intermediate time and space complexity between linear and full attention, including sparse attention with token eviction, and the query-aware native sparse attention. Particularly, we propose a novel learnable token eviction approach. Combined with sliding-window attention, an end-to-end trainable lightweight CNN aggregates information from both past and future adjacent tokens to adaptively retain a limited set of critical KV-pairs per head, maintaining linear attention's constant time and space complexity. Efficient Triton kernels for the sparse attention mechanisms are provided. Empirical evaluations on retrieval-intensive benchmarks support the effectiveness of our approaches.

Transformers and large language models (LLMs) have revolutionized machine learning, with attention mechanisms at the core of their success. As the landscape of attention variants expands, so too do the challenges of optimizing their performance, particularly across different hardware platforms. Current optimization strategies are often narrowly focused, requiring extensive manual intervention to accommodate changes in model configurations or hardware environments. In this paper, we introduce AttentionEngine, a comprehensive framework designed to streamline the optimization of attention mechanisms across heterogeneous hardware backends. By decomposing attention computation into modular operations with customizable components, AttentionEngine enables flexible adaptation to diverse algorithmic requirements. The framework further automates kernel optimization through a combination of programmable templates and a robust cross-platform scheduling strategy. Empirical results reveal performance gains of up to 10x on configurations beyond the reach of existing methods. AttentionEngine offers a scalable, efficient foundation for developing and deploying attention mechanisms with minimal manual tuning. Our code has been open-sourced and is available at https://github.com/microsoft/AttentionEngine.

Transformers serve as the foundation of most modern large language models. To mitigate the quadratic complexity of standard full attention, various efficient attention mechanisms, such as linear and hybrid attention, have been developed. A fundamental gap remains: their expressive power relative to full attention lacks a rigorous theoretical characterization. In this work, we theoretically characterize the performance differences among these attention mechanisms. Our theory applies to all linear attention variants that can be formulated as a recurrence, including Mamba, DeltaNet, etc. Specifically, we establish an expressiveness hierarchy: for the sequential function composition-a multi-step reasoning task that must occur within a model's forward pass, an (L+1)-layer full attention network is sufficient, whereas any hybrid network interleaving L-1 layers of full attention with a substantially larger number (2^{3L^2}) of linear attention layers cannot solve it. This result demonstrates a clear separation in expressive power between the two types of attention. Our work provides the first provable separation between hybrid attention and standard full attention, offering a theoretical perspective for understanding the fundamental capabilities and limitations of different attention mechanisms.

Hybrid language models combine softmax attention with linear-time sequence mechanisms such as state-space or linear-attention layers, but the functional contribution of each component type remains insufficiently characterized. We study component-level ablation in two sub-1B hybrid language models, Qwen3.5-0.8B and Falcon-H1-0.5B, using likelihood-based evaluation, downstream benchmarks, layer-wise interventions, random controls, and representation-level diagnostics. Across the tested models, removing either attention or the alternative sequence-processing pathway substantially degrades performance, indicating that both component types contribute to model behavior. Likelihood metrics are especially sensitive to the linear-attention or state-space pathway, while downstream benchmark degradation depends on task and architecture. Layer-wise ablations show that component importance is position-dependent, with the strongest effects concentrated in early or mid-network components rather than uniformly across depth. Random-removal controls further show that hybrid architectures and same-family Transformer baselines degrade differently under structural perturbation. These results suggest that component ablation is a useful diagnostic for understanding hybrid language model architectures. The findings provide evidence relevant to efficient model design, compression, robustness analysis, and deployment decisions in architectures that combine attention with alternative sequence-processing mechanisms.

Transformers face quadratic complexity and memory issues with long sequences, prompting the adoption of linear attention mechanisms using fixed-size hidden states. However, linear models often suffer from limited recall performance, leading to hybrid architectures that combine linear and full attention layers. Despite extensive hybrid architecture research, the choice of linear attention component has not been deeply explored. We systematically evaluate various linear attention models across generations - vector recurrences to advanced gating mechanisms - both standalone and hybridized. To enable this comprehensive analysis, we trained and open-sourced 72 models: 36 at 340M parameters (20B tokens) and 36 at 1.3B parameters (100B tokens), covering six linear attention variants across five hybridization ratios. Benchmarking on standard language modeling and recall tasks reveals that superior standalone linear models do not necessarily excel in hybrids. While language modeling remains stable across linear-to-full attention ratios, recall significantly improves with increased full attention layers, particularly below a 3:1 ratio. Our study highlights selective gating, hierarchical recurrence, and controlled forgetting as critical for effective hybrid models. We recommend architectures such as HGRN-2 or GatedDeltaNet with a linear-to-full ratio between 3:1 and 6:1 to achieve Transformer-level recall efficiently. Our models are open-sourced at https://huggingface.co/collections/m-a-p/hybrid-linear-attention-research-686c488a63d609d2f20e2b1e.

Diffusion language models (dLLMs) offer a promising alternative to autoregressive models by accelerating inference through parallel decoding. Recent dLLMs commonly use blockwise semi-autoregressive decoding, generating blocks autoregressively while denoising tokens within each active block in parallel. However, despite KV caching, each denoising step still attends to all previous blocks, repeatedly incurring prefix-attention cost. Motivated by this bottleneck, we ask whether dLLM inference can be further accelerated by linearizing attention over previous blocks. We introduce block-hybrid attention, which retains exact softmax attention within the active denoising block while applying linear attention over previous blocks. We show that this hybrid attention can be retrofitted into a pretrained dLLM with minimal post-training: LLaDA-Hybrid replaces 6 of the 20 attention layers in LLaDA~2.1, a 16B open-source dLLM, largely following LoLCAT (Zhang et al, 2024). The conversion takes only approximately 60 hours while preserving benchmark performance: 72.0% vs. 75.6% on HumanEval, 63.0% vs. 57.7% on MBPP+, and 86.7% vs. 88.3% on CMATH. With a Triton implementation, LLaDA-Hybrid achieves up to 1.7times higher decoding throughput and supports more concurrent requests before exhausting memory, showing that pretrained dLLMs can be efficiently linearized for faster inference. Our code is available at: https://github.com/Diuven/LLaDA-Hybrid.

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.

Recurrent-attention hybrids aim to combine the efficiency of recurrence with the expressivity of attention, but existing approaches typically apply attention uniformly across all positions, even when the recurrent state alone is sufficient for accurate prediction. We introduce AMOR (Adaptive Metacognitive Output Router), a post-hoc hybrid architecture that selectively invokes attention based on predictive uncertainty. A recurrent backbone is augmented with entropy-gated attention blocks that activate only when the model's output entropy exceeds a dynamic threshold derived from a running batch median and scaled standard deviation. This yields a simple, gradient-free routing mechanism inspired by uncertainty-driven computation and the System 1 / System 2 distinction. Across Mamba2 and Gated DeltaNet backbones (180M-1.5B), AMOR consistently matches or outperforms both pure recurrent models and fixed-schedule hybrid baselines while invoking attention on only ~22% of tokens. It achieves strong performance on common-sense reasoning benchmarks and maintains stable long-context performance on LongBench, where prior hybrid models degrade under distribution shift. These results suggest that when attention is applied matters as much as how much: selectively allocating attention based on predictive uncertainty improves both efficiency and robustness, offering a simple alternative to uniform or fixed routing strategies and pointing toward adaptive hybrid architectures that dynamically match computation to input difficulty.

Gated attention is an effective approach to mitigate attention sinks and enhance the representational capacity of attention. To further extend its effectiveness-efficiency Pareto frontier, we propose a Hybrid Gated Attention (HyGA) framework that contains three types of gating strategies. Specifically, these gates leverage diverse information from multiple stages of attention, and collaboratively build element-wise/head-wise gating from multiple perspectives, capturing intra-head and cross-head information interactions. Through our hybrid gating components, HyGA could provide multi-source modulation signals, enabling more comprehensive control over information flow and improving the representational capacity of attention. We also introduce low-rank matrix decomposition and learnable attention sink to further enhance training efficiency and stability. In experiments, we evaluate HyGA on widely-used benchmarks based on different backbones. The experimental results show that our HyGA comprehensively improves both training loss and various downstream performances compared with Gated attention. HyGA has also been verified to achieve the best performance at different computation costs, with comprehensive model analyses for better understanding. The proposed HyGA sheds light on a more effective, efficient, and stable attention mechanism.

Softmax attention is the cornerstone of modern large language models, but its memory scales linearly and compute quadratically with sequence length. Linear recurrent models, such as linear attention and state space models, have become widely studied as alternatives to attention due to their linear compute and constant memory. While these sub-quadratic token mixing methods, or mixers, achieve promising efficiency gains and competitive results on a wide range of benchmarks, current linear recurrent models still lag behind on tasks that require long-context retrieval or in-context learning. A growing body of work studies hybrid architectures that attempt to mitigate these trade-offs by statically interleaving or merging attention and recurrent blocks. In this work, we explore a new axis of developing hybrid models: across the token sequence. We propose Oryx, a hybrid model that can, throughout a sequence, flexibly switch between different mixers, for example quadratic attention for rich context utilization and linear recurrences for efficient generation. Oryx ties at least 90% of its parameters across mixers, enabling attention and recurrent modes to operate over shared internal representations. We validate our design with Mamba-2 and Gated DeltaNet variants, up to 1.4B models. Under fixed token budgets and a mixed-training strategy, Oryx achieves comparable or better performance than its single-mixer baselines. At the 1.4B scale, all instances of Oryx outperform their respective baselines by at least 0.7 percentage points on averaged language modeling tasks. On retrieval tasks, Oryx achieves performance comparable to the Transformer baseline even when processing only a tiny fraction (<10%) of the tokens in attention mode. These results suggest that attention and linear recurrent models can share internal representations, and motivate sequence-axis hybridization as a promising direction.

In this work, we upgrade the multi-head attention mechanism, the core of the Transformer model, to improve efficiency while maintaining or surpassing the previous accuracy level. We show that multi-head attention can be expressed in the summation form. Drawing on the insight that not all attention heads hold equal significance, we propose Mixture-of-Head attention (MoH), a new architecture that treats attention heads as experts in the Mixture-of-Experts (MoE) mechanism. MoH has two significant advantages: First, MoH enables each token to select the appropriate attention heads, enhancing inference efficiency without compromising accuracy or increasing the number of parameters. Second, MoH replaces the standard summation in multi-head attention with a weighted summation, introducing flexibility to the attention mechanism and unlocking extra performance potential. Extensive experiments on ViT, DiT, and LLMs demonstrate that MoH outperforms multi-head attention by using only 50%-90% of the attention heads. Moreover, we demonstrate that pre-trained multi-head attention models, such as LLaMA3-8B, can be further continue-tuned into our MoH models. Notably, MoH-LLaMA3-8B achieves an average accuracy of 64.0% across 14 benchmarks, outperforming LLaMA3-8B by 2.4% by utilizing only 75% of the attention heads. We believe the proposed MoH is a promising alternative to multi-head attention and provides a strong foundation for developing advanced and efficient attention-based models.

Streaming video generation (SVG) distills a pretrained bidirectional video diffusion model into an autoregressive model equipped with sliding window attention (SWA). However, SWA inevitably loses distant history during long video generation, and its computational overhead remains a critical challenge to real-time deployment. In this work, we propose Hybrid Forcing, which jointly optimizes temporal information retention and computational efficiency through a hybrid attention design. First, we introduce lightweight linear temporal attention to preserve long-range dependencies beyond the sliding window. In particular, we maintain a compact key-value state to incrementally absorb evicted tokens, retaining temporal context with negligible memory and computational overhead. Second, we incorporate block-sparse attention into the local sliding window to reduce redundant computation within short-range modeling, reallocating computational capacity toward more critical dependencies. Finally, we introduce a decoupled distillation strategy tailored to the hybrid attention design. A few-step initial distillation is performed under dense attention, then the distillation of our proposed linear temporal and block-sparse attention is activated for streaming modeling, ensuring stable optimization. Extensive experiments on both short- and long-form video generation benchmarks demonstrate that Hybrid Forcing consistently achieves state-of-the-art performance. Notably, our model achieves real-time, unbounded 832x480 video generation at 29.5 FPS on a single NVIDIA H100 GPU without quantization or model compression. The source code and trained models are available at https://github.com/leeruibin/hybrid-forcing.

Transformers have become the cornerstone of modern large-scale language models; however, their dependence on softmax attention poses a major computational bottleneck, particularly in long-context settings. In this work, rather than following prevalent approaches such as linear attention (or SSMs) and local attention, we introduce an intermediate design called \rat between recurrence and attention mechanisms. It partitions the input into chunks, applies a simple linear recurrence within each chunk to capture local dependencies, and then performs softmax attention across chunks to model long-range interactions. By adjusting the size of the chunk, \rat enables flexible trade-offs, combining the strengths of RNN and attention. Empirically, with a chunk size of 16, the \rat layer achieves a \(7\times\) improvement in training speed with 100K token sequences and \(9\times\) in generation at 4K sequence length, while maintaining similar or sometimes even better accuracy compared to standard attention. We demonstrate this by training 1.3B parameter models from scratch and performing large-scale evaluations, including short- and long-context benchmarks, as well as supervised fine-tuning~(SFT). We further propose a hybrid architecture that interleaves \rat with local attention. By combining efficient long-range modeling with strong local interactions, this hybrid design not only improves inference speed and reduces cache memory usage compared to attention, but also consistently enhances performance, for example, achieving an average 1 point gain in commonsense reasoning tasks, up to 4 points on code tasks, and a 1 point Rouge-L increase in a summarization SFT task. Code is available at https://github.com/CLAIRE-Labo/RAT

Modeling long sequences of user behaviors has emerged as a critical frontier in generative recommendation. However, existing solutions face a dilemma: linear attention mechanisms achieve efficiency at the cost of retrieval precision due to limited state capacity, while softmax attention suffers from prohibitive computational overhead. To address this challenge, we propose HyTRec, a model featuring a Hybrid Attention architecture that explicitly decouples long-term stable preferences from short-term intent spikes. By assigning massive historical sequences to a linear attention branch and reserving a specialized softmax attention branch for recent interactions, our approach restores precise retrieval capabilities within industrial-scale contexts involving ten thousand interactions. To mitigate the lag in capturing rapid interest drifts within the linear layers, we furthermore design Temporal-Aware Delta Network (TADN) to dynamically upweight fresh behavioral signals while effectively suppressing historical noise. Empirical results on industrial-scale datasets confirm the superiority that our model maintains linear inference speed and outperforms strong baselines, notably delivering over 8% improvement in Hit Rate for users with ultra-long sequences with great efficiency.

The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.

To mitigate the computational complexity in the self-attention mechanism on long sequences, linear attention utilizes computation tricks to achieve linear complexity, while state space models (SSMs) popularize a favorable practice of using non-data-dependent memory pattern, i.e., emphasize the near and neglect the distant, to processing sequences. Recent studies have shown the priorities by combining them as one. However, the efficiency of linear attention remains only at the theoretical level in a causal setting, and SSMs require various designed constraints to operate effectively on specific data. Therefore, in order to unveil the true power of the hybrid design, the following two issues need to be addressed: (1) hardware-efficient implementation for linear attention and (2) stabilization of SSMs. To achieve this, we leverage the thought of tiling and hierarchy to propose CHELA (short-long Convolutions with Hardware-Efficient Linear Attention), which replaces SSMs with short-long convolutions and implements linear attention in a divide-and-conquer manner. This approach enjoys global abstraction and data-dependent selection from stable SSM and linear attention while maintaining real linear complexity. Our comprehensive experiments on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method.

High-frequency stock price prediction is challenging due to non-stationarity, noise, and volatility. To tackle these issues, we propose the Hybrid Attentive Ensemble Learning Transformer (HAELT), a deep learning framework combining a ResNet-based noise-mitigation module, temporal self-attention for dynamic focus on relevant history, and a hybrid LSTM-Transformer core that captures both local and long-range dependencies. These components are adaptively ensembled based on recent performance. Evaluated on hourly Apple Inc. (AAPL) data from Jan 2024 to May 2025, HAELT achieves the highest F1-Score on the test set, effectively identifying both upward and downward price movements. This demonstrates HAELT's potential for robust, practical financial forecasting and algorithmic trading.

When reading books, humans focus primarily on the current page, flipping back to recap prior context only when necessary. Similarly, we demonstrate that Large Language Models (LLMs) can learn to dynamically determine when to attend to global context. We propose All-or-Here Attention (AHA), which utilizes a binary router per attention head to dynamically toggle between full attention and local sliding window attention for each token. Our results indicate that with a window size of 256 tokens, up to 93\% of the original full attention operations can be replaced by sliding window attention without performance loss. Furthermore, by evaluating AHA across various window sizes, we identify a long-tail distribution in context dependency, where the necessity for full attention decays rapidly as the local window expands. By decoupling local processing from global access, AHA reveals that full attention is largely redundant, and that efficient inference requires only on-demand access to the global context.

We introduce convolutional multi-hybrid architectures, with a design grounded on two simple observations. First, operators in hybrid models can be tailored to token manipulation tasks such as in-context recall, multi-token recall, and compression, with input-dependent convolutions and attention offering complementary performance. Second, co-designing convolution operators and hardware-aware algorithms enables efficiency gains in regimes where previous alternative architectures struggle to surpass Transformers. At the 40 billion parameter scale, we train end-to-end 1.2 to 2.9 times faster than optimized Transformers, and 1.1 to 1.4 times faster than previous generation hybrids. On H100 GPUs and model width 4096, individual operators in the proposed multi-hybrid StripedHyena 2 architecture achieve two-fold throughput improvement over linear attention and state-space models. Multi-hybrids excel at sequence modeling over byte-tokenized data, as demonstrated by the Evo 2 line of models. We discuss the foundations that enable these results, including architecture design, overlap-add blocked kernels for tensor cores, and dedicated all-to-all and point-to-point context parallelism strategies.

Over the past 7 years, attention has become one of the most important primitives in deep learning. The primary approach to optimize attention is FlashAttention, which fuses the operation together, drastically improving both the runtime and the memory consumption. However, the importance of FlashAttention combined with its monolithic nature poses a problem for researchers aiming to try new attention variants -- a "software lottery". This problem is exacerbated by the difficulty of writing efficient fused attention kernels, resisting traditional compiler-based approaches. We introduce FlexAttention, a novel compiler-driven programming model that allows implementing the majority of attention variants in a few lines of idiomatic PyTorch code. We demonstrate that many existing attention variants (e.g. Alibi, Document Masking, PagedAttention, etc.) can be implemented via FlexAttention, and that we achieve competitive performance compared to these handwritten kernels. Finally, we demonstrate how FlexAttention allows for easy composition of attention variants, solving the combinatorial explosion of attention variants.

Transformers have achieved remarkable success across multiple fields, yet their impact on 3D medical image segmentation remains limited with convolutional networks still dominating major benchmarks. In this work, we a) analyze current Transformer-based segmentation models and identify critical shortcomings, particularly their over-reliance on convolutional blocks. Further, we demonstrate that in some architectures, performance is unaffected by the absence of the Transformer, thereby demonstrating their limited effectiveness. To address these challenges, we move away from hybrid architectures and b) introduce a fully Transformer-based segmentation architecture, termed Primus. Primus leverages high-resolution tokens, combined with advances in positional embeddings and block design, to maximally leverage its Transformer blocks. Through these adaptations Primus surpasses current Transformer-based methods and competes with state-of-the-art convolutional models on multiple public datasets. By doing so, we create the first pure Transformer architecture and take a significant step towards making Transformers state-of-the-art for 3D medical image segmentation.

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.

Modern large language models based on softmax scaled-dot-product attention are constrained by their training sequence length: as the key-value sequence grows, softmax probability mass can dilute across a wider distribution, inducing activation shift and long-context performance collapse. Moreover, long-context language modeling faces a structural tension: a sliding-window attention core maintains a bounded local representation and low perplexity but is blind to long-range dependencies, while full-context attention preserves global recall but suffers from out-of-distribution perplexity explosion. To resolve these limitations, we introduce ATMA, a hybrid convolutional-attention architecture that integrates a novel three-channel attention mechanism. ATMA factorizes the attention mixing step into: (1) a count-blind, unit-vector direction channel, (2) a bounded magnitude channel driven by the participation ratio of effective matches over an extreme-value-corrected null sink, and (3) a long-term recurrent compression memory optimized via a gated-delta fast-weights rule. Neither the Polar Attention core nor the recurrent memory is sufficient alone; their combination enables monotonic perplexity reduction and high-fidelity long-range retrieval simultaneously. We evaluate ATMA using a 100-run factorial ablation sweep, demonstrating that the combined Polar + memory model maintains induction needle-in-a-haystack retrieval accuracy above 90% out to 64K tokens (32 times the training length of 2K) while its document perplexity improves monotonically, outperforming softmax-based memory baselines which collapse at extreme context lengths. Code: https://github.com/kreasof-ai/atma

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.

Looped Transformers (LT) have emerged as a powerful architecture by iterating their layers multiple times before decoding the final token. However, pairing them with full attention retains quadratic complexity, making them computationally expensive and slow. We introduce LT2 (Linear-Time Looped Transformers), a family of looped architectures that replace quadratic softmax attention with subquadratic, linear-time attention. We study two variants: LT2-linear with linear attention and LT2-sparse with sparse attention. We find that looping uniquely synergizes with these variants: it enables iterative memory refinement in linear attention and progressively expands the effective receptive field in sparse attention. We formalize these benefits theoretically and demonstrate consistent empirical gains across controlled recall, state-tracking, and language modeling tasks. We then explore LT2-hybrid, which combines different attention variants in a looped setting. Two variants are especially promising: LT2-hybrid (GDN+DSA), which interleaves linear and sparse attention to maximize efficiency and matches the standard looped transformer's quality at fully linear-time cost; and LT2-hybrid (Full+GDN), which interleaves GDN with a small fraction of full attention layers to maximize quality, surpassing the standard looped transformer in both performance and efficiency. We also show how to convert a pre-trained LT into an LT2-hybrid model. With about 1B tokens of training, our converted model, Ouro-hybrid-1.4B, outperforms industry-level 1B models and is competitive with industry-level 4B models while retaining the speed benefits of linear-time attention. Together, these results show a clear path toward making looped transformers more scalable and advancing efficient, capable small language models.

Autoregressive (AR) large language models (LLMs) have achieved broad practical success, but sequential decoding remains a key bottleneck for low-latency deployment. Recent efficient-inference work has progressed along two axes: reducing the cost of each model invocation through efficient architectures, and reducing serial decoding steps through parallel generation. Hybrid attention backbones address the former, while diffusion language models (dLLMs) pursue the latter via iterative parallel denoising. Combining these advantages remains challenging: AR-to-dLLM conversion often fails to preserve seed-checkpoint capability, and hybrid-attention recurrent states and masking constraints make diffusion training and serving nontrivial. We present FLARE, a systematic conversion framework for hybrid-attention LLMs. Our analysis identifies transfer data quality as the primary determinant of capability preservation, outweighing loss formulation and attention-mask design. The resulting framework combines a token-equal AR-and-diffusion objective, hardware-aware kernels, and unified inference, enabling one checkpoint to support both AR-style verified decoding and diffusion-style parallel denoising. Starting from strong AR checkpoints with limited post-training data, FLARE is competitive with leading open-source dLLMs across model scales and delivers consistent throughput gains over open-source dLLM baselines in single-GPU concurrent serving. Our results further suggest that practical dLLMs are limited not only by decoding algorithms, but also by transfer data quality and the training inefficiency of current block-diffusion objectives, motivating joint design of data, objectives, architectures, and inference systems.

Scaling pre-trained language models has resulted in large performance gains in various natural language processing tasks but comes with a large cost in memory requirements. Inspired by the position embeddings in transformers, we aim to simplify and reduce the memory footprint of the multi-head attention (MHA) mechanism. We propose an alternative module that uses only a single shared projection matrix and multiple head embeddings (MHE), i.e. one per head. We empirically demonstrate that our MHE attention is substantially more memory efficient compared to alternative attention mechanisms while achieving high predictive performance retention ratio to vanilla MHA on several downstream tasks. MHE attention only requires a negligible fraction of additional parameters (3nd, where n is the number of attention heads and d the size of the head embeddings) compared to a single-head attention, while MHA requires (3n^2-3n)d^2-3nd additional parameters.

Modeling long-range dependencies remains a central challenge in natural language processing. Transformer architectures achieve strong performance via self-attention but scale quadratically (O(N^2)) with sequence length, while State Space Models (SSMs) scale linearly (O(N)) but suffer from a selective recall bottleneck, struggling to retrieve precise information from compressed states. This creates a fundamental tradeoff between efficiency and perplexity. To tackle these challenges, we propose the Parallel Hybrid Architecture (PHA), which runs Gated State Spaces (GSS), Grouped Query Attention (GQA), and Feed-Forward Networks (FFNs) as independent parallel branches fused by a learnable mixing mechanism. Instead of forcing SSMs to approximate attention or serializing the two paradigms, PHA allows each branch to specialize: GSS captures global context, while attention performs selective retrieval, with FFN providing complementary processing. On WikiText-103, PHA achieves 16.51 PPL at 125M parameters, outperforming Hedgehog (16.70) and H3-125M (23.70). Scaling to 180M parameters yields 16.42 PPL, which gives comparable results with the pure attention baseline while delivering 24\% higher throughput and up to 40\% lower memory usage at long contexts. On OpenWebText, our 125M model achieves 19.72 PPL, outperforming standard Transformers (20.60) and GSS hybrid baselines (19.80). These results demonstrate that separating sequence modeling paradigms into parallel specialists enables Transformer-level perplexity with substantially improved efficiency for long-context language modeling.

We propose Hymba, a family of small language models featuring a hybrid-head parallel architecture that integrates transformer attention mechanisms with state space models (SSMs) for enhanced efficiency. Attention heads provide high-resolution recall, while SSM heads enable efficient context summarization. Additionally, we introduce learnable meta tokens that are prepended to prompts, storing critical information and alleviating the "forced-to-attend" burden associated with attention mechanisms. This model is further optimized by incorporating cross-layer key-value (KV) sharing and partial sliding window attention, resulting in a compact cache size. During development, we conducted a controlled study comparing various architectures under identical settings and observed significant advantages of our proposed architecture. Notably, Hymba achieves state-of-the-art results for small LMs: Our Hymba-1.5B-Base model surpasses all sub-2B public models in performance and even outperforms Llama-3.2-3B with 1.32% higher average accuracy, an 11.67x cache size reduction, and 3.49x throughput.

Hybrid architectures combining full attention (FA) and sliding-window attention (SWA) are a promising paradigm for efficient LLM inference. However, existing methods typically rely on hand-crafted rules or simple post-hoc heuristics for FA/SWA allocation and offer limited analysis of the attention behaviors underlying these designs. We propose Controllable Sparsity in Hybrid Attention (ConSA), a framework that learns optimal FA/SWA assignment under a user-specified sparsity target. ConSA employs L0 regularization to learn binary masks selecting between FA and SWA for each attention unit, while an augmented Lagrangian constraint enforces the target sparsity at either layer or KV-head granularity. We evaluate ConSA on two LLMs at the 0.6B and 1.7B scales. Learned allocations consistently outperform rule-based baselines, with KV-head-wise allocation yielding clear gains over layer-wise allocation. The learned patterns place SWA in the bottom layers and concentrate FA into contiguous middle-layer blocks, diverging from evenly interleaved patterns in rule-based methods. This structure persists across model scales, sparsity levels, and allocation granularities, revealing a fine-grained spectrum of intrinsic attention behaviors that underlies the learned allocation.

Hybrid Transformer architectures, which combine softmax attention blocks and recurrent neural networks (RNNs), have shown a desirable performance-throughput tradeoff for long-context modeling, but their adoption and studies are hindered by the prohibitive cost of large-scale pre-training from scratch. Some recent studies have shown that pre-trained softmax attention blocks can be converted into RNN blocks through parameter transfer and knowledge distillation. However, these transfer methods require substantial amounts of training data (more than 10B tokens), and the resulting hybrid models also exhibit poor long-context performance, which is the scenario where hybrid models enjoy significant inference speedups over Transformer-based models. In this paper, we present HALO (Hybrid Attention via Layer Optimization), a pipeline for distilling Transformer models into RNN-attention hybrid models. We then present HypeNet, a hybrid architecture with superior length generalization enabled by a novel position encoding scheme (named HyPE) and various architectural modifications. We convert the Qwen3 series into HypeNet using HALO, achieving performance comparable to the original Transformer models while enjoying superior long-context performance and efficiency. The conversion requires just 2.3B tokens, less than 0.01% of their pre-training data

Large Language Models (LLMs) encounter significant performance bottlenecks in long-sequence tasks due to the computational complexity and memory overhead inherent in the self-attention mechanism. To address these challenges, we introduce AllMem, a novel and efficient hybrid architecture that integrates Sliding Window Attention (SWA) with non-linear Test-Time Training (TTT) memory networks. AllMem enables models to effectively scale to ultra-long contexts while mitigating catastrophic forgetting. This approach not only overcomes the representation constraints typical of linear memory models but also significantly reduces the computational and memory footprint during long-sequence inference. Furthermore, we implement a Memory-Efficient Fine-Tuning strategy to replace standard attention layers in pre-trained models with memory-augmented sliding window layers. This framework facilitates the efficient transformation of any off-the-shelf pre-trained LLM into an AllMem-based architecture. Empirical evaluations confirm that our 4k window model achieves near-lossless performance on 37k LongBench with a marginal 0.83 drop compared to full attention. Furthermore, on InfiniteBench at a 128k context, our 8k window variant outperforms full attention, which validates the effectiveness of our parameterized memory in mitigating noise and maintaining robust long-range modeling without the prohibitive costs of global attention.

We design a family of image classification architectures that optimize the trade-off between accuracy and efficiency in a high-speed regime. Our work exploits recent findings in attention-based architectures, which are competitive on highly parallel processing hardware. We revisit principles from the extensive literature on convolutional neural networks to apply them to transformers, in particular activation maps with decreasing resolutions. We also introduce the attention bias, a new way to integrate positional information in vision transformers. As a result, we propose LeVIT: a hybrid neural network for fast inference image classification. We consider different measures of efficiency on different hardware platforms, so as to best reflect a wide range of application scenarios. Our extensive experiments empirically validate our technical choices and show they are suitable to most architectures. Overall, LeViT significantly outperforms existing convnets and vision transformers with respect to the speed/accuracy tradeoff. For example, at 80% ImageNet top-1 accuracy, LeViT is 5 times faster than EfficientNet on CPU. We release the code at https://github.com/facebookresearch/LeViT

Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by 3.9times with the same average attention span, boosting retrieval accuracy by 1.5-7.1times over the uniform-attention baseline across Vicuna-7B, Vicuna-13B, and Llama3-8B models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from 9%-36% to within 5% across two long-context understanding benchmarks. MoA achieves a 1.2-1.4times GPU memory reduction and boosts decode throughput by 5.5-6.7 times for 7B and 13B dense models on a single GPU, with minimal impact on performance.

We study architectural and optimization techniques for sample-efficient language modeling under the constraints of the BabyLM 2025 shared task. Our model, BLaLM, replaces self-attention with a linear-time mLSTM token mixer and explores lightweight enhancements, including short convolutions, sliding window attention with dynamic modulation, and Hedgehog feature maps. To support training in low-resource settings, we curate a high-quality corpus emphasizing readability and pedagogical structure. Experiments across both STRICT and STRICT-SMALL tracks show that (1) linear attention combined with sliding window attention consistently improves zero-shot performance, and (2) the Muon optimizer stabilizes convergence and reduces perplexity over AdamW. These results highlight effective strategies for efficient language modeling without relying on scale.

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.

Modeling attention in neural multi-source sequence-to-sequence learning remains a relatively unexplored area, despite its usefulness in tasks that incorporate multiple source languages or modalities. We propose two novel approaches to combine the outputs of attention mechanisms over each source sequence, flat and hierarchical. We compare the proposed methods with existing techniques and present results of systematic evaluation of those methods on the WMT16 Multimodal Translation and Automatic Post-editing tasks. We show that the proposed methods achieve competitive results on both tasks.

Transformer-based methods have shown impressive performance in low-level vision tasks, such as image super-resolution. However, we find that these networks can only utilize a limited spatial range of input information through attribution analysis. This implies that the potential of Transformer is still not fully exploited in existing networks. In order to activate more input pixels for better reconstruction, we propose a novel Hybrid Attention Transformer (HAT). It combines both channel attention and window-based self-attention schemes, thus making use of their complementary advantages of being able to utilize global statistics and strong local fitting capability. Moreover, to better aggregate the cross-window information, we introduce an overlapping cross-attention module to enhance the interaction between neighboring window features. In the training stage, we additionally adopt a same-task pre-training strategy to exploit the potential of the model for further improvement. Extensive experiments show the effectiveness of the proposed modules, and we further scale up the model to demonstrate that the performance of this task can be greatly improved. Our overall method significantly outperforms the state-of-the-art methods by more than 1dB. Codes and models are available at https://github.com/XPixelGroup/HAT.

Key-Value (KV) cache memory and bandwidth increasingly dominate large language model inference cost in long-context and long-generation regimes. Architectures such as multi-head latent attention (MLA) and hybrid sliding-window attention (SWA) can alleviate this bound, but integrating them into existing models remains difficult. Prior methods impose fine-grained structural requirements on both source and target attention modules, which cannot meet the feasible requirement in practical deployment. We present Attention Editing, a practical framework for converting already-trained large language models (LLMs) with new attention architectures without re-pretraining from scratch. Attention editing replaces the original attention with a learnable target module and trains it using progressive distillation, consisting of (1) layer-wise teacher-forced optimization with intermediate activation supervision to prevent cold-start error accumulation, and (2) model-level distillation on next-token distributions, optionally regularized by weak feature matching. We instantiate the framework on two different target--MLA and GateSWA, a gated hybrid SWA design, and apply it to Qwen3-8B and Qwen3-30B-A3B. The resulting models maintain competitive performance while delivering substantial efficiency improvements, demonstrating that large-scale attention conversion is both feasible and robust. Notably, experiments are conducted on an Ascend 910B clusters, offering a practical training case study on domestic hardware.

Hybrid sequence models--combining Transformer and state-space model layers--seek to gain the expressive versatility of attention as well as the computational efficiency of state-space model layers. Despite burgeoning interest in hybrid models, we lack a basic understanding of the settings where--and underlying mechanisms through which--they offer benefits over their constituent models. In this paper, we study this question, focusing on a broad family of core synthetic tasks. For this family of tasks, we prove the existence of fundamental limitations for non-hybrid models. Specifically, any Transformer or state-space model that solves the underlying task requires either a large number of parameters or a large working memory. On the other hand, for two prototypical tasks within this family--namely selective copying and associative recall--we construct hybrid models of small size and working memory that provably solve these tasks, thus achieving the best of both worlds. Our experimental evaluation empirically validates our theoretical findings. Importantly, going beyond the settings in our theoretical analysis, we empirically show that learned--rather than constructed--hybrids outperform non-hybrid models with up to 6x as many parameters. We additionally demonstrate that hybrid models exhibit stronger length generalization and out-of-distribution robustness than non-hybrids.

As large language models (LLMs) continue to advance rapidly, they are becoming increasingly capable while simultaneously demanding ever-longer context lengths. To improve the inference efficiency of long-context processing, several novel low-complexity hybrid architectures have recently been proposed, effectively alleviating the computational burden of long-context inference. However, existing research on long-context prefill acceleration remains predominantly focused on sparse attention mechanisms, which achieve their maximum speedup only on full-attention models. When transferred to emerging architectures--such as linear/full attention hybrids or sliding window/full attention hybrids--these prefill acceleration approaches suffer significant performance degradation. Furthermore, such methods are generally incompatible with continuous batching, making them difficult to integrate into modern inference engines such as vLLM. To this end, we propose UniPrefill, a prefill acceleration framework applicable to virtually any model architecture, which directly accelerates the model's computation at the token level. We further implement UniPrefill as a continuous batching operator and extend vLLM's scheduling strategy to natively support prefill-decode co-processing and tensor parallel for UniPrefill, enabling its seamless integration into vLLM. UniPrefill achieves up to 2.1x speedup in Time-To-First-Token (TTFT), with the acceleration becoming increasingly pronounced as the number of concurrent requests grows.

Transformers have achieved significant success across various domains, relying on self-attention to capture dependencies. However, the standard first-order attention mechanism is often limited by a low-rank bottleneck, struggling to capture intricate, multi-hop relationships within a single layer. In this paper, we propose the Nexus, a novel architecture designed to enhance representational power through a recursive framework. Unlike standard approaches that use static linear projections for Queries and Keys, Nexus dynamically refines these representations via nested self-attention mechanisms. Specifically, the Query and Key vectors are themselves outputs of inner attention loops, allowing tokens to aggregate global context and model high-order correlations prior to the final attention computation. We enforce a parameter-efficient weight-sharing strategy across recursive steps, ensuring that this enhanced expressivity incurs O(1) additional parameters. We provide theoretical analysis demonstrating that our method breaks the linear bottleneck of standard attention. Empirically, Nexus outperforms standard Transformers on multiple benchmarks.

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.

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 success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.

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.

Hybrid models that combine the language modeling capabilities of Attention layers with the efficiency of Recurrent layers (e.g., State Space Models) have gained traction in practically supporting long contexts in Large Language Model serving. Yet, the unique properties of these models complicate the usage of complementary efficiency optimizations such as prefix caching that skip redundant computations across requests. Most notably, their use of in-place state updates for recurrent layers precludes rolling back cache entries for partial sequence overlaps, and instead mandates only exact-match cache hits; the effect is a deluge of (large) cache entries per sequence, most of which yield minimal reuse opportunities. We present Marconi, the first system that supports efficient prefix caching with Hybrid LLMs. Key to Marconi are its novel admission and eviction policies that more judiciously assess potential cache entries based not only on recency, but also on (1) forecasts of their reuse likelihood across a taxonomy of different hit scenarios, and (2) the compute savings that hits deliver relative to memory footprints. Across diverse workloads and Hybrid models, Marconi achieves up to 34.4times higher token hit rates (71.1% or 617 ms lower TTFT) compared to state-of-the-art prefix caching systems.

Recent work has shown that attention-based language models excel at recall, the ability to ground generations in tokens previously seen in context. However, the efficiency of attention-based models is bottle-necked during inference by the KV-cache's aggressive memory consumption. In this work, we explore whether we can improve language model efficiency (e.g. by reducing memory consumption) without compromising on recall. By applying experiments and theory to a broad set of architectures, we identify a key tradeoff between a model's state size and recall ability. We show that efficient alternatives to attention (e.g. H3, Mamba, RWKV) maintain a fixed-size recurrent state, but struggle at recall. We propose BASED a simple architecture combining linear and sliding window attention. By varying BASED window size and linear attention feature dimension, we can dial the state size and traverse the pareto frontier of the recall-memory tradeoff curve, recovering the full quality of attention on one end and the small state size of attention-alternatives on the other. We train language models up to 1.3b parameters and show that BASED matches the strongest sub-quadratic models (e.g. Mamba) in perplexity and outperforms them on real-world recall-intensive tasks by 6.22 accuracy points. Implementations of linear attention are often less efficient than optimized standard attention implementations. To make BASED competitive, we develop IO-aware algorithms that enable 24x higher throughput on language generation than FlashAttention-2, when generating 1024 tokens using 1.3b parameter models. Code for this work is provided at: https://github.com/HazyResearch/based.

In this paper, we propose Superlinear attention, a fully trainable multi-step attention architecture that achieves subquadratic complexity for long sequences while preserving random context access (a.k.a.\ structural non-exclusion): no eligible token position is structurally excluded from being selected for attention. Superlinear attention reformulates standard causal self-attention as a multi-step search problem with N steps, yielding an overall complexity of O(L^{1+1{N}}). To illustrate the architecture, we present a baseline N=2 implementation, which is algorithmically analogous to standard jump search. In this O(L^{3/2}) instantiation, the first step performs O(L^{3/2}) span-search to select relevant spans of the sequence, and the second step applies O(L^{3/2}) span-attention (standard attention restricted to the selected spans). In an upscaled O(L^{1.54}) configuration for robustness, we achieve an average decoding throughput of 114 tokens/sec at 1M context length and 80 tokens/sec at 10M context in our implementation on a modified 30B hybrid MoE model on a single B200 GPU. With limited training, we also obtain strong performance on the NIAH (Needle In A Haystack) task up to 256K context length, demonstrating that the routed span selection is learnable end-to-end. This paper emphasizes architectural formulation, scaling analysis, and systems feasibility, and presents initial validation; comprehensive quality evaluations across diverse long-context tasks are left to future work.

Sparse attention is becoming increasingly important for serving large language models (LLMs) as generation lengths continue to grow. However, deploying and evaluating new sparse attention algorithms at scale remains highly engineering-intensive, slowing both human researchers and AI agents in exploring the sparse attention design. To address this challenge, we present Vortex, a system that combines a Python-embedded frontend language atop a page-centric tensor abstraction for expressing a broad range of sparse attention algorithms, with an efficient backend tightly integrated into modern LLM serving stacks. Vortex enables rapid prototyping, deployment, and evaluation of sparse attention algorithms, effectively translating their theoretical efficiency gains into real-world throughput improvements. As a result, Vortex substantially accelerates the design and iteration of sparse attention algorithms. First, AI agents use Vortex to automatically generate and refine diverse algorithms, the best reaching up to 3.46times higher throughput than full attention while preserving accuracy. Second, Vortex extends sparse attention to emerging architectures and very large models that are otherwise hard to experiment with, reaching up to 4.7times higher throughput on the MLA-based GLM-4.7-Flash and 1.37times on the 229B-parameter MiniMax-M2.7 on NVIDIA B200 GPUs.

Linear attention has recently gained significant attention for long-context inference due to its constant decoding cost with respect to context length. However, existing serving systems typically serve linear attention by recurrently computing and updating a large linear attention state in every decoding step. Since the state is much larger than the per-token key and value, recurrent decoding incurs substantial memory access and becomes inefficient for serving linear attention. In this paper, we propose KVBuffer, an IO-aware serving mechanism for linear attention. By buffering recent keys and values, KVBuffer enables serving systems to compute linear attention outputs in more flexible and memory-efficient ways. For decoding, KVBuffer enables chunkwise computation, which reduces average memory access and decoding latency by deferring state updates and applying them in batch. For speculative decoding, KVBuffer verifies draft tokens in parallel and avoids storing temporary states. For short contexts, KVBuffer computes attention outputs directly from buffered keys and values, without creating or updating the linear attention state. We implement KVBuffer in SGLang for Qwen3-Next. Our evaluations show that KVBuffer can reduce linear attention decoding latency by up to 45.17% and increase the maximum number of serving requests by 5x for speculative decoding when verifying four draft tokens.

The performance of Large Language Models (LLMs) is significantly sensitive to the contextual position of information in the input. To investigate the mechanism behind this positional bias, our extensive experiments reveal a consistent phenomenon we term the attention basin: when presented with a sequence of structured items (e.g., retrieved documents or few-shot examples), models systematically assign higher attention to the items at the beginning and end of the sequence, while neglecting those in the middle. Crucially, our analysis further reveals that allocating higher attention to critical information is key to enhancing model performance. Based on these insights, we introduce Attention-Driven Reranking (AttnRank), a two-stage framework that (i) estimates a model's intrinsic positional attention preferences using a small calibration set, and (ii) reorders retrieved documents or few-shot examples to align the most salient content with these high-attention positions. AttnRank is a model-agnostic, training-free, and plug-and-play method with minimal computational overhead. Experiments on multi-hop QA and few-shot in-context learning tasks demonstrate that AttnRank achieves substantial improvements across 10 large language models of varying architectures and scales, without modifying model parameters or training procedures.

In this paper, we demonstrate that an inherent waveform pattern in the attention allocation of large language models (LLMs) significantly affects their performance in tasks demanding a high degree of context awareness, such as utilizing LLMs for tool-use. Specifically, the crucial information in the context will be potentially overlooked by model when it is positioned in the trough zone of the attention waveform, leading to decreased performance. To address this issue, we propose a novel inference method named Attention Buckets. It allows LLMs to process their input through multiple parallel processes. Each process utilizes a distinct base angle for the rotary position embedding, thereby creating a unique attention waveform. By compensating an attention trough of a particular process with an attention peak of another process, our approach enhances LLM's awareness to various contextual positions, thus mitigating the risk of overlooking crucial information. In the largest tool-use benchmark, our method elevates a 7B model to achieve state-of-the-art performance, comparable to that of GPT-4. On other benchmarks and some RAG tasks, which also demand a thorough understanding of contextual content, Attention Buckets also exhibited notable enhancements in performance.

Computational workloads composing traditional Transformer models are starkly bifurcated. Multi-Head Attention (MHA) is memory-bound, with low arithmetic intensity, while feedforward layers are compute-bound. This dichotomy has long motivated research into specialized hardware to mitigate the MHA bottleneck. This paper argues that recent architectural shifts, namely Multi-head Latent Attention (MLA) and Mixture-of-Experts (MoE), challenge the premise of specialized attention hardware. We make two key observations. First, the arithmetic intensity of MLA is over two orders of magnitude greater than that of MHA, shifting it close to a compute-bound regime well-suited for modern accelerators like GPUs. Second, by distributing MoE experts across a pool of accelerators, their arithmetic intensity can be tuned through batching to match that of the dense layers, creating a more balanced computational profile. These findings reveal a diminishing need for specialized attention hardware. The central challenge for next-generation Transformers is no longer accelerating a single memory-bound layer. Instead, the focus must shift to designing balanced systems with sufficient compute, memory capacity, memory bandwidth, and high-bandwidth interconnects to manage the diverse demands of large-scale models.

In this technical report, we present the Ring-linear model series, specifically including Ring-mini-linear-2.0 and Ring-flash-linear-2.0. Ring-mini-linear-2.0 comprises 16B parameters and 957M activations, while Ring-flash-linear-2.0 contains 104B parameters and 6.1B activations. Both models adopt a hybrid architecture that effectively integrates linear attention and softmax attention, significantly reducing I/O and computational overhead in long-context inference scenarios. Compared to a 32 billion parameter dense model, this series reduces inference cost to 1/10, and compared to the original Ring series, the cost is also reduced by over 50%. Furthermore, through systematic exploration of the ratio between different attention mechanisms in the hybrid architecture, we have identified the currently optimal model structure. Additionally, by leveraging our self-developed high-performance FP8 operator library-linghe, overall training efficiency has been improved by 50%. Benefiting from the high alignment between the training and inference engine operators, the models can undergo long-term, stable, and highly efficient optimization during the reinforcement learning phase, consistently maintaining SOTA performance across multiple challenging complex reasoning benchmarks.

We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attention layers, giving rise to inter-spike plateaus (ISP). As full attention becomes denser, successive PAS become increasingly connected through ISP, ultimately recovering the stable MA morphology of full attention LLMs. We establish the recurrence of this organization across five linear attention architectures, six hybridization configurations, five data domains, and representative open-source hybrid models spanning 1.2B to 397B total parameters. Controlled pretraining of GDN-based hybrids at scales up to 1.3B shows that both morphologies emerge early and respond asymmetrically to output gating: full attention output gating strongly attenuates their absolute magnitudes without eliminating their layerwise organization, whereas removing GDN gates yields comparatively modest amplification. Mechanistically, our systematic-outlier analysis supports a shared lifecycle account governed by the timing of MA cancellation. PAS follows a localized write-sink-cancel process, while the extended persistence of ISP is consistent with delayed cancellation. At the full attention limit, this account recovers the stable MA morphology characteristic of full attention LLMs. Our code is available at https://github.com/StartluxLabs/Massive-Activations-HLA.

We present Lightning Attention, the first linear attention implementation that maintains a constant training speed for various sequence lengths under fixed memory consumption. Due to the issue with cumulative summation operations (cumsum), previous linear attention implementations cannot achieve their theoretical advantage in a casual setting. However, this issue can be effectively solved by utilizing different attention calculation strategies to compute the different parts of attention. Specifically, we split the attention calculation into intra-blocks and inter-blocks and use conventional attention computation for intra-blocks and linear attention kernel tricks for inter-blocks. This eliminates the need for cumsum in the linear attention calculation. Furthermore, a tiling technique is adopted through both forward and backward procedures to take full advantage of the GPU hardware. To enhance accuracy while preserving efficacy, we introduce TransNormerLLM (TNL), a new architecture that is tailored to our lightning attention. We conduct rigorous testing on standard and self-collected datasets with varying model sizes and sequence lengths. TNL is notably more efficient than other language models. In addition, benchmark results indicate that TNL performs on par with state-of-the-art LLMs utilizing conventional transformer structures. The source code is released at github.com/OpenNLPLab/TransnormerLLM.

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.

We propose a technique for learning representations of parser states in transition-based dependency parsers. Our primary innovation is a new control structure for sequence-to-sequence neural networks---the stack LSTM. Like the conventional stack data structures used in transition-based parsing, elements can be pushed to or popped from the top of the stack in constant time, but, in addition, an LSTM maintains a continuous space embedding of the stack contents. This lets us formulate an efficient parsing model that captures three facets of a parser's state: (i) unbounded look-ahead into the buffer of incoming words, (ii) the complete history of actions taken by the parser, and (iii) the complete contents of the stack of partially built tree fragments, including their internal structures. Standard backpropagation techniques are used for training and yield state-of-the-art parsing performance.

As context lengths scale, attention increasingly becomes a primary computational bottleneck in large language models. Standard Transformers remain powerful but computationally inefficient, as they allocate the same attention budget to every token regardless of its contextual demand. Existing local-global hybrids provide a more efficient alternative by mixing restricted- and full-context attention, but they typically allocate span statically across layers or heads. To address these limitations, we propose LoGo, a token-level dynamic local-global attention mechanism that uses attention span as a direct proxy for attention budget allocation. Each LoGo layer contains coupled local and global branches: all tokens receive efficient local attention over a restricted context window, while a learned gate activates global attention with full-context access only for tokens requiring long-range information. A threshold-based budget controller maintains a target global ratio without auxiliary losses, and a progressive masking schedule stabilizes training before sparse routing takes effect. We further implement query-sparse Triton kernels that convert reduced global-attention computation into practical speedups. Extensive experiments validate LoGo's effectiveness, showing that it preserves the scaling behavior of full-attention Transformers across model sizes. In controlled comparisons, LoGo improves over the full-attention Transformer and matched-budget static local-global hybrids, with clear gains on long-range retrieval. Analysis further shows that LoGo learns interpretable span allocation patterns. These results suggest that learned token-level span allocation is an effective and scalable way to improve the long-context performance-compute trade-off.

Machine comprehension (MC), answering a query about a given context paragraph, requires modeling complex interactions between the context and the query. Recently, attention mechanisms have been successfully extended to MC. Typically these methods use attention to focus on a small portion of the context and summarize it with a fixed-size vector, couple attentions temporally, and/or often form a uni-directional attention. In this paper we introduce the Bi-Directional Attention Flow (BIDAF) network, a multi-stage hierarchical process that represents the context at different levels of granularity and uses bi-directional attention flow mechanism to obtain a query-aware context representation without early summarization. Our experimental evaluations show that our model achieves the state-of-the-art results in Stanford Question Answering Dataset (SQuAD) and CNN/DailyMail cloze test.

In multi-source sequence-to-sequence tasks, the attention mechanism can be modeled in several ways. This topic has been thoroughly studied on recurrent architectures. In this paper, we extend the previous work to the encoder-decoder attention in the Transformer architecture. We propose four different input combination strategies for the encoder-decoder attention: serial, parallel, flat, and hierarchical. We evaluate our methods on tasks of multimodal translation and translation with multiple source languages. The experiments show that the models are able to use multiple sources and improve over single source baselines.

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.

Hybrid State-Space models combine Attention with recurrent State-Space Model (SSM) layers, balancing eidetic memory from Attention with compressed fading memory from SSMs. This yields smaller Key-Value caches and faster decoding than Transformers, along with a richer architectural design space. Exploring that design space at scale has so far required training from scratch, a barrier that has kept most large-model Hybrid research within a narrow range of architectures. We introduce Priming, a method that turns Hybrid architecture design from a pre-training problem into a knowledge transfer one. Priming initializes a Hybrid model from a pre-trained Transformer and, through short alignment and post-training phases, recovers downstream quality using less than 0.5% of the source model's pre-training token budget. Priming is agnostic to the source Transformer family (e.g., Qwen, Llama, Mistral), model class (dense or Mixture-of-Experts), and model scale. Priming enables us to run the first controlled comparison of SSM layer types at scale under identical conditions. We evaluate, Gated KalmaNet (GKA), Gated DeltaNet (GDN), and Mamba-2, and show that their expressiveness hierarchy, GKA>GDN>Mamba-2, directly predicts downstream performance on long-context reasoning tasks. We scale Priming to 8B/32B reasoning models with native 128K contexts. Our Hybrid GKA 32B improves over its source Qwen3-32B by +3.8 average reasoning points, while staying within 1% of a Transformer post-trained on the same data and enabling up to 2.3x higher decode throughput. To foster research on Hybrid architectures, we release a model zoo of primed Hybrid models for long-context reasoning and instruction following, together with the Priming training and inference code (Sequence Parallelism algorithms for long-context training, optimized GKA kernels, and vLLM serving plugin), all under Apache~2.0 License.

We present a very simple algorithm for attention that requires O(1) memory with respect to sequence length and an extension to self-attention that requires O(log n) memory. This is in contrast with the frequently stated belief that self-attention requires O(n^2) memory. While the time complexity is still O(n^2), device memory rather than compute capability is often the limiting factor on modern accelerators. Thus, reducing the memory requirements of attention allows processing of longer sequences than might otherwise be feasible. We provide a practical implementation for accelerators that requires O(n) memory, is numerically stable, and is within a few percent of the runtime of the standard implementation of attention. We also demonstrate how to differentiate the function while remaining memory-efficient. For sequence length 16384, the memory overhead of self-attention is reduced by 59X for inference and by 32X for differentiation.

The design choices in the Transformer attention mechanism, including weak inductive bias and quadratic computational complexity, have limited its application for modeling long sequences. In this paper, we introduce Mega, a simple, theoretically grounded, single-head gated attention mechanism equipped with (exponential) moving average to incorporate inductive bias of position-aware local dependencies into the position-agnostic attention mechanism. We further propose a variant of Mega that offers linear time and space complexity yet yields only minimal quality loss, by efficiently splitting the whole sequence into multiple chunks with fixed length. Extensive experiments on a wide range of sequence modeling benchmarks, including the Long Range Arena, neural machine translation, auto-regressive language modeling, and image and speech classification, show that Mega achieves significant improvements over other sequence models, including variants of Transformers and recent state space models.

The Attention module finds common usage in language modeling, presenting distinct challenges within the broader scope of Natural Language Processing. Multi-Head Attention (MHA) employs an absolute positional encoding, which imposes limitations on token length and entails substantial memory consumption during the processing of embedded inputs. The current remedy proposed by researchers involves the utilization of relative positional encoding, similar to the approach adopted in Transformer-XL or Relative Multi-Head Attention (RMHA), albeit the employed architecture consumes considerable memory resources. To address these challenges, this study endeavors to refine MHA, leveraging relative positional encoding in conjunction with the Depth-Wise Convolutional Layer architecture, which promises heightened accuracy coupled with minimized memory usage. The proposed RCMHA framework entails the modification of two integral components: firstly, the application of the Depth-Wise Convolutional Layer to the input embedding, encompassing Query, Key, and Value parameters; secondly, the incorporation of Relative Positional Encoding into the attention scoring phase, harmoniously integrated with Scaled Dot-Product Attention. Empirical experiments underscore the advantages of RCMHA, wherein it exhibits superior accuracy, boasting a score of 0.572 in comparison to alternative attention modules such as MHA, Multi-DConv-Head Attention (MDHA), and RMHA. Concerning memory utilization, RMHA emerges as the most frugal, demonstrating an average consumption of 2.98 GB, surpassing RMHA which necessitates 3.5 GB.

Attention-free language models that combine gating and convolutions are growing in popularity due to their efficiency and increasingly competitive performance. To better understand these architectures, we pretrain a suite of 17 attention and "gated-convolution" language models, finding that SoTA gated-convolution architectures still underperform attention by up to 2.1 perplexity points on the Pile. In fine-grained analysis, we find 82% of the gap is explained by each model's ability to recall information that is previously mentioned in-context, e.g. "Hakuna Matata means no worries Hakuna Matata it means no" rightarrow "??". On this task, termed "associative recall", we find that attention outperforms gated-convolutions by a large margin: a 70M parameter attention model outperforms a 1.4 billion parameter gated-convolution model on associative recall. This is surprising because prior work shows gated convolutions can perfectly solve synthetic tests for AR capability. To close the gap between synthetics and real language, we develop a new formalization of the task called multi-query associative recall (MQAR) that better reflects actual language. We perform an empirical and theoretical study of MQAR that elucidates differences in the parameter-efficiency of attention and gated-convolution recall. Informed by our analysis, we evaluate simple convolution-attention hybrids and show that hybrids with input-dependent sparse attention patterns can close 97.4% of the gap to attention, while maintaining sub-quadratic scaling. Our code is accessible at: https://github.com/HazyResearch/zoology.

Attention is a fundamental component behind the remarkable achievements of large language models (LLMs). However, our current understanding of the attention mechanism, especially regarding how attention distributions are established, remains limited. Inspired by recent studies that explore the presence of attention sink in the initial token, which receives disproportionately large attention scores despite their lack of semantic importance, this work delves deeper into this phenomenon. We aim to provide a more profound understanding of the existence of attention sinks within LLMs and to uncover ways to enhance the achievable accuracy of LLMs by directly optimizing the attention distributions, without the need for weight finetuning. Specifically, this work begins with comprehensive visualizations of the attention distributions in LLMs during inference across various inputs and tasks. Based on these visualizations, to the best of our knowledge, we are the first to discover that (1) attention sinks occur not only at the start of sequences but also within later tokens of the input, and (2) not all attention sinks have a positive impact on the achievable accuracy of LLMs. Building upon our findings, we propose a training-free Attention Calibration Technique (ACT) that automatically optimizes the attention distributions on the fly during inference in an input-adaptive manner. Extensive experiments validate that ACT consistently enhances the accuracy of various LLMs across different applications. Specifically, ACT achieves an average improvement of up to 7.30% in accuracy across different datasets when applied to Llama-30B. Our code is available at https://github.com/GATECH-EIC/ACT.

We introduce Kimi Linear, a hybrid linear attention architecture that, for the first time, outperforms full attention under fair comparisons across various scenarios -- including short-context, long-context, and reinforcement learning (RL) scaling regimes. At its core lies Kimi Delta Attention (KDA), an expressive linear attention module that extends Gated DeltaNet with a finer-grained gating mechanism, enabling more effective use of limited finite-state RNN memory. Our bespoke chunkwise algorithm achieves high hardware efficiency through a specialized variant of the Diagonal-Plus-Low-Rank (DPLR) transition matrices, which substantially reduces computation compared to the general DPLR formulation while remaining more consistent with the classical delta rule. We pretrain a Kimi Linear model with 3B activated parameters and 48B total parameters, based on a layerwise hybrid of KDA and Multi-Head Latent Attention (MLA). Our experiments show that with an identical training recipe, Kimi Linear outperforms full MLA with a sizeable margin across all evaluated tasks, while reducing KV cache usage by up to 75% and achieving up to 6 times decoding throughput for a 1M context. These results demonstrate that Kimi Linear can be a drop-in replacement for full attention architectures with superior performance and efficiency, including tasks with longer input and output lengths. To support further research, we open-source the KDA kernel and vLLM implementations, and release the pre-trained and instruction-tuned model checkpoints.

The Transformer architecture processes sequences by implementing a form of neural message-passing that consists of iterative information retrieval (attention), followed by local processing (position-wise MLP). Two types of information are essential under this general computational paradigm: "sensory" information about individual objects, and "relational" information describing the relationships between objects. Standard attention naturally encodes the former, but does not explicitly encode the latter. In this paper, we present an extension of Transformers where multi-head attention is augmented with two distinct types of attention heads, each routing information of a different type. The first type is the standard attention mechanism of Transformers, which captures object-level features, while the second type is a novel attention mechanism we propose to explicitly capture relational information. The two types of attention heads each possess different inductive biases, giving the resulting architecture greater efficiency and versatility. The promise of this approach is demonstrated empirically across a range of tasks.

The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sharing undermines model performance by neglecting the functional diversity of attention heads. To address this, we propose LycheeDecode, an efficient decoding method centered on a fine-grained hybrid-head attention mechanism that employs a hardware-efficient top-k selection strategy. Specifically, the novel HardKuma-based mechanism partitions attention heads into a small subset of retrieval heads that dynamically identify crucial tokens and a majority of sparse heads that reuse them for efficient computation. Through extensive experiments on leading models like Llama3 and Qwen3 across diverse benchmarks for long-context understanding (e.g., LongBench, RULER) and complex reasoning (e.g., AIME24, OlympiadBench), we demonstrate that LycheeDecode achieves generative quality comparable to, and at times surpassing even the full-attention baseline. Crucially, this is accomplished with up to a 2.7x speedup at a 128K context length. By preserving the functional diversity of attention heads, our fine-grained strategy overcomes the performance bottlenecks of existing methods, providing a powerful and validated pathway to both efficient and high-quality long-context LLM inference.

Language models typically construct attention values from contextual hidden states, even when some of their content may be reusable across contexts. We investigate whether token-indexed memory can replace the dedicated value projection when complemented by contextual information. We propose Memory Attention (MA), which forms values by combining layer-specific token memory with contextual keys. The memory supplies token-specific representations, while the keys preserve context dependence. At inference, normalization can be folded into the memory tables, reducing value construction to lookup and addition. Token-indexed retrieval also enables CPU offloading with prefetching, reducing GPU parameter storage. Under matched training token budgets and with additional memory parameters, experiments across attention configurations show improved language modeling and average downstream performance.

Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .

Recent advances in video diffusion models have shifted towards transformer-based architectures, achieving state-of-the-art video generation but at the cost of quadratic attention complexity, which severely limits scalability for longer sequences. We introduce ReHyAt, a Recurrent Hybrid Attention mechanism that combines the fidelity of softmax attention with the efficiency of linear attention, enabling chunk-wise recurrent reformulation and constant memory usage. Unlike the concurrent linear-only SANA Video, ReHyAt's hybrid design allows efficient distillation from existing softmax-based models, reducing the training cost by two orders of magnitude to ~160 GPU hours, while being competitive in the quality. Our light-weight distillation and finetuning pipeline provides a recipe that can be applied to future state-of-the-art bidirectional softmax-based models. Experiments on VBench and VBench-2.0, as well as a human preference study, demonstrate that ReHyAt achieves state-of-the-art video quality while reducing attention cost from quadratic to linear, unlocking practical scalability for long-duration and on-device video generation. Project page is available at https://qualcomm-ai-research.github.io/rehyat.

In large language models built upon the Transformer architecture, recent studies have shown that inter-head interaction can enhance attention performance. Motivated by this, we propose Multi-head Explicit Attention (MEA), a simple yet effective attention variant that explicitly models cross-head interaction. MEA consists of two key components: a Head-level Linear Composition (HLC) module that separately applies learnable linear combinations to the key and value vectors across heads, thereby enabling rich inter-head communication; and a head-level Group Normalization layer that aligns the statistical properties of the recombined heads. MEA shows strong robustness in pretraining, which allows the use of larger learning rates that lead to faster convergence, ultimately resulting in lower validation loss and improved performance across a range of tasks. Furthermore, we explore the parameter efficiency of MEA by reducing the number of attention heads and leveraging HLC to reconstruct them using low-rank "virtual heads". This enables a practical key-value cache compression strategy that reduces KV-cache memory usage by 50% with negligible performance loss on knowledge-intensive and scientific reasoning tasks, and only a 3.59% accuracy drop for Olympiad-level mathematical benchmarks.

We present Neural Attention Search (NAtS), a framework that automatically evaluates the importance of each token within a sequence and determines if the corresponding token can be dropped after several steps. This approach can efficiently reduce the KV cache sizes required by transformer-based models during inference and thus reduce inference costs. In this paper, we design a search space that contains three token types: (i) Global Tokens will be preserved and queried by all the following tokens. (ii) Local Tokens survive until the next global token appears. (iii) Sliding Window Tokens have an impact on the inference of a fixed size of the next following tokens. Similar to the One-Shot Neural Architecture Search approach, this token-type information can be learned jointly with the architecture weights via a learnable attention mask. Experiments on both training a new transformer from scratch and fine-tuning existing large language models show that NAtS can efficiently reduce the KV cache size required for the models while maintaining the models' performance.

This paper addresses the limitations of traditional keyword-based search in understanding user intent and introduces a novel hybrid search approach that leverages the strengths of non-semantic search engines, Large Language Models (LLMs), and embedding models. The proposed system integrates keyword matching, semantic vector embeddings, and LLM-generated structured queries to deliver highly relevant and contextually appropriate search results. By combining these complementary methods, the hybrid approach effectively captures both explicit and implicit user intent.The paper further explores techniques to optimize query execution for faster response times and demonstrates the effectiveness of this hybrid search model in producing comprehensive and accurate search outcomes.

Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup.

Recent work has demonstrated the potential of non-transformer language models, especially linear recurrent neural networks (RNNs) and hybrid models that mix recurrence and attention. Yet there is no consensus on whether the potential benefits of these new architectures justify the risk and effort of scaling them up. To address this, we provide evidence for the advantages of hybrid models over pure transformers on several fronts. First, theoretically, we show that hybrid models do not merely inherit the expressivity of transformers and linear RNNs, but can express tasks beyond both, such as code execution. Putting this theory to practice, we train Olmo Hybrid, a 7B-parameter model largely comparable to Olmo 3 7B but with the sliding window layers replaced by Gated DeltaNet layers. We show that Olmo Hybrid outperforms Olmo 3 across standard pretraining and mid-training evaluations, demonstrating the benefit of hybrid models in a controlled, large-scale setting. We find that the hybrid model scales significantly more efficiently than the transformer, explaining its higher performance. However, its unclear why greater expressivity on specific formal problems should result in better scaling or superior performance on downstream tasks unrelated to those problems. To explain this apparent gap, we return to theory and argue why increased expressivity should translate to better scaling efficiency, completing the loop. Overall, our results suggest that hybrid models mixing attention and recurrent layers are a powerful extension to the language modeling paradigm: not merely to reduce memory during inference, but as a fundamental way to obtain more expressive models that scale better during pretraining.

Dot-product attention has wide applications in computer vision and natural language processing. However, its memory and computational costs grow quadratically with the input size. Such growth prohibits its application on high-resolution inputs. To remedy this drawback, this paper proposes a novel efficient attention mechanism equivalent to dot-product attention but with substantially less memory and computational costs. Its resource efficiency allows more widespread and flexible integration of attention modules into a network, which leads to better accuracies. Empirical evaluations demonstrated the effectiveness of its advantages. Efficient attention modules brought significant performance boosts to object detectors and instance segmenters on MS-COCO 2017. Further, the resource efficiency democratizes attention to complex models, where high costs prohibit the use of dot-product attention. As an exemplar, a model with efficient attention achieved state-of-the-art accuracies for stereo depth estimation on the Scene Flow dataset. Code is available at https://github.com/cmsflash/efficient-attention.

State-space models (SSMs) offer efficient sequence modeling but lag behind Transformers on benchmarks that require in-context retrieval. Prior work links this gap to a small set of attention heads, termed Gather-and-Aggregate (G&A), which SSMs struggle to reproduce. We propose *retrieval-aware distillation*, which converts a pretrained Transformer into a hybrid student by preserving only these retrieval-critical heads and distilling the rest into recurrent heads. We identify the essential heads via ablation on a synthetic retrieval task, producing a hybrid with sparse, non-uniform attention placement. We show that preserving **just 2% of attention heads recovers over 95% of teacher performance on retrieval-heavy tasks** (10 heads in a 1B model), requiring far fewer heads than hybrids that retain at least 25%. We further find that large recurrent states often compensate for missing retrieval: once retrieval is handled by these heads, the SSM backbone can be simplified with limited loss, even with an 8times reduction in state dimension. By reducing both the attention cache and the SSM state, the resulting hybrid is 5--6times more memory-efficient than comparable hybrids, closing the Transformer--SSM gap at a fraction of the memory cost.

Neural sequence-to-sequence models have provided a viable new approach for abstractive text summarization (meaning they are not restricted to simply selecting and rearranging passages from the original text). However, these models have two shortcomings: they are liable to reproduce factual details inaccurately, and they tend to repeat themselves. In this work we propose a novel architecture that augments the standard sequence-to-sequence attentional model in two orthogonal ways. First, we use a hybrid pointer-generator network that can copy words from the source text via pointing, which aids accurate reproduction of information, while retaining the ability to produce novel words through the generator. Second, we use coverage to keep track of what has been summarized, which discourages repetition. We apply our model to the CNN / Daily Mail summarization task, outperforming the current abstractive state-of-the-art by at least 2 ROUGE points.

Hybrid LLM architectures that combine Attention and State Space Models (SSMs) achieve state-of-the-art accuracy and runtime performance. Recent work has demonstrated that applying compression and distillation to Attention-only models yields smaller, more accurate models at a fraction of the training cost. In this work, we explore the effectiveness of compressing Hybrid architectures. We introduce a novel group-aware pruning strategy that preserves the structural integrity of SSM blocks and their sequence modeling capabilities. Furthermore, we demonstrate the necessity of such SSM pruning to achieve improved accuracy and inference speed compared to traditional approaches. Our compression recipe combines SSM, FFN, embedding dimension, and layer pruning, followed by knowledge distillation-based retraining, similar to the MINITRON technique. Using this approach, we compress the Nemotron-H 8B Hybrid model down to 4B parameters with up to 40x fewer training tokens. The resulting model surpasses the accuracy of similarly-sized models while achieving 2x faster inference, significantly advancing the Pareto frontier.

Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.

Linear attention is an efficient attention mechanism that has recently emerged as a promising alternative to conventional softmax attention. With its ability to process tokens in linear computational complexities, linear attention, in theory, can handle sequences of unlimited length without sacrificing speed, i.e., maintaining a constant training speed for various sequence lengths with a fixed memory consumption. However, due to the issue with cumulative summation (cumsum), current linear attention algorithms cannot demonstrate their theoretical advantage in a causal setting. In this paper, we present Lightning Attention-2, the first linear attention implementation that enables linear attention to realize its theoretical computational benefits. To achieve this, we leverage the thought of tiling, separately handling the intra-block and inter-block components in linear attention calculation. Specifically, we utilize the conventional attention computation mechanism for the intra-blocks and apply linear attention kernel tricks for the inter-blocks. A tiling technique is adopted through both forward and backward procedures to take full advantage of the GPU hardware. We implement our algorithm in Triton to make it IO-aware and hardware-friendly. Various experiments are conducted on different model sizes and sequence lengths. Lightning Attention-2 retains consistent training and inference speed regardless of input sequence length and is significantly faster than other attention mechanisms. The source code is available at https://github.com/OpenNLPLab/lightning-attention.

Multi-Head Latent Attention (MLA), introduced in DeepSeek-V2, improves the efficiency of large language models by projecting query, key, and value tensors into a compact latent space. This architectural change reduces the KV-cache size and significantly lowers memory bandwidth demands, particularly in the autoregressive decode phase. This letter presents the first hardware-centric analysis of MLA, comparing it to conventional Multi-Head Attention (MHA) and evaluating its implications for accelerator performance. We identify two alternative execution schemes of MLA--reusing, resp. recomputing latent projection matrices--which offer distinct trade-offs between compute and memory access. Using the Stream design space exploration framework, we model their throughput and energy cost across a range of hardware platforms and find that MLA can shift attention workloads toward the compute-bound regime. Our results show that MLA not only reduces bandwidth usage but also enables adaptable execution strategies aligned with hardware constraints. Compared to MHA, it provides more stable and efficient performance, particularly on bandwidth-limited hardware platforms. These findings emphasize MLA's relevance as a co-design opportunity for future AI accelerators.

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 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.

While transformers have shown remarkable success in natural language processing, their attention mechanism's large memory requirements have limited their ability to handle longer contexts. Prior approaches, such as recurrent memory or retrieval-based augmentation, have either compromised the random-access flexibility of attention (i.e., the capability to select any token in the entire context) or relied on separate mechanisms for relevant context retrieval, which may not be compatible with the model's attention. In this paper, we present a novel approach that allows access to the complete context while retaining random-access flexibility, closely resembling running attention on the entire context. Our method uses a landmark token to represent each block of the input and trains the attention to use it for selecting relevant blocks, enabling retrieval of blocks directly through the attention mechanism instead of by relying on a separate mechanism. Our approach seamlessly integrates with specialized data structures and the system's memory hierarchy, enabling processing of arbitrarily long context lengths. We demonstrate that our method can obtain comparable performance with Transformer-XL while significantly reducing the number of retrieved tokens in each step. Finally, we show that fine-tuning LLaMA 7B with our method successfully extends its context length capacity up to 32k tokens, allowing for inference at the context lengths of GPT-4.

Mixture-of-Experts (MoE) networks have been proposed as an efficient way to scale up model capacity and implement conditional computing. However, the study of MoE components mostly focused on the feedforward layer in Transformer architecture. This paper proposes the Mixture of Attention Heads (MoA), a new architecture that combines multi-head attention with the MoE mechanism. MoA includes a set of attention heads that each has its own set of parameters. Given an input, a router dynamically selects a subset of k attention heads per token. This conditional computation schema allows MoA to achieve stronger performance than the standard multi-head attention layer. Furthermore, the sparsely gated MoA can easily scale up the number of attention heads and the number of parameters while preserving computational efficiency. In addition to the performance improvements, MoA also automatically differentiates heads' utilities, providing a new perspective to discuss the model's interpretability. We conducted experiments on several important tasks, including Machine Translation and Masked Language Modeling. Experiments have shown promising results on several tasks against strong baselines that involve large and very deep models.

Effective attention modules have played a crucial role in the success of Transformer-based large language models (LLMs), but the quadratic time and memory complexities of these attention modules also pose a challenge when processing long sequences. One potential solution for the long sequence problem is to utilize distributed clusters to parallelize the computation of attention modules across multiple devices (e.g., GPUs). However, adopting a distributed approach inevitably introduces extra memory overheads to store local attention results and incurs additional communication costs to aggregate local results into global ones. In this paper, we propose a distributed attention framework named ``BurstAttention'' to optimize memory access and communication operations at both the global cluster and local device levels. In our experiments, we compare BurstAttention with other competitive distributed attention solutions for long sequence processing. The experimental results under different length settings demonstrate that BurstAttention offers significant advantages for processing long sequences compared with these competitive baselines, reducing 40% communication overheads and achieving 2 X speedup during training 32K sequence length on 8 X A100.

Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architectural based on a hybrid attention mechanism that not only surpasses conventional RoPE-based transformer models in long context tasks but also achieves competitive performance on benchmarks requiring shorter context lengths.

Transformer-based large language models (LLMs) have achieved remarkable success, yet their standard attention mechanism incurs quadratic computation and memory costs with respect to sequence length, posing a major bottleneck for long-context training. Prior work tackles this challenge along two directions: (1) kernel-level optimizations, which accelerate dense and sparse attention operators; and (2) module-level strategies, often referred to as distributed attention or context parallel training, which scale attention across multiple devices. However, systematic evaluation still remains limited: operator-level comparisons are often incomplete, while context parallel strategies are typically framework-specific, with unclear performance analysis across contexts. To address these gaps, we propose a unified benchmark that integrates representative attention kernels and context parallel mechanisms with a modular and extensible interface for evaluation. The benchmark evaluates methods along two critical dimensions: (1) attention mask patterns, which strongly affect efficiency, scalability, and usability, and (2) sequence length and distributed scale, which determine performance under extreme long-context training. Through comprehensive experiments on the cluster of up to 96 GPUs, our benchmark enables reproducible comparisons, highlights method-specific trade-offs, and provides practical guidance for designing and deploying attention mechanisms in long-context LLM training.

Large language models are increasingly applied to multi-document and long-form inputs, yet long-context inference remains memory- and noise-inefficient. Key-value (KV) caching scales linearly with context length, while external retrieval methods often return lexically similar but causally irrelevant passages. We present S3-Attention, a memory-first inference-time framework that treats long-context processing as attention-aligned endogenous retrieval. S3-Attention decodes transient key and query projections into top-k sparse feature identifiers using lightweight sparse autoencoders, and constructs a CPU-based inverted index mapping features to token positions or spans during a single streaming scan. This design allows the KV cache to be discarded entirely and bounds GPU memory usage by the scan chunk size. At generation time, feature co-activation is used to retrieve compact evidence spans, optionally fused with BM25 for exact lexical matching. Under a unified LongBench evaluation protocol with fixed prompting, decoding, and matched token budgets, S3-Hybrid closely matches full-context inference across multiple model families and improves robustness in several information-dense settings. We also report an engineering limitation of the current prototype, which incurs higher wall-clock latency than optimized full-KV baselines, motivating future kernel-level optimization.

We compare the fast training and decoding speed of RETURNN of attention models for translation, due to fast CUDA LSTM kernels, and a fast pure TensorFlow beam search decoder. We show that a layer-wise pretraining scheme for recurrent attention models gives over 1% BLEU improvement absolute and it allows to train deeper recurrent encoder networks. Promising preliminary results on max. expected BLEU training are presented. We are able to train state-of-the-art models for translation and end-to-end models for speech recognition and show results on WMT 2017 and Switchboard. The flexibility of RETURNN allows a fast research feedback loop to experiment with alternative architectures, and its generality allows to use it on a wide range of applications.

Multi-head attention layers, as used in the Transformer neural sequence model, are a powerful alternative to RNNs for moving information across and between sequences. While training these layers is generally fast and simple, due to parallelizability across the length of the sequence, incremental inference (where such paralleization is impossible) is often slow, due to the memory-bandwidth cost of repeatedly loading the large "keys" and "values" tensors. We propose a variant called multi-query attention, where the keys and values are shared across all of the different attention "heads", greatly reducing the size of these tensors and hence the memory bandwidth requirements of incremental decoding. We verify experimentally that the resulting models can indeed be much faster to decode, and incur only minor quality degradation from the baseline.