Hybrid Attention

Combining attention's global retrieval with the sequential importance signal of state space models (SSMs) is the open challenge of hybrid language modeling. Transformers see everywhere but cannot prioritize; SSMs know what matters but cannot revisit. Existing hybrids -- Jamba (block level) and Hymba (head level) -- place the two in separate compartments, so neither informs the other during the attention computation itself. We propose SISA (SSM-Informed Softmax Attention), which adds an SSM-derived importance term directly inside the attention score and realizes the full operation as a single SDPA call on augmented query/key vectors -- no recurrent state, no custom kernel. At 152M / 5B tokens, SISA reaches LAMBADA-greedy 17.3% (vs. Transformer 13.9 and Mamba-3 15.5) and attains NIAH 100% from step 1K, 7x faster than Transformer's retrieval convergence; at 369M, Mamba-3 leads LAMBADA while SISA preserves perfect NIAH and stock-SDPA execution. SISA thus defines a third design axis for SSM-attention hybrids -- score-level fusion -- beyond the block-level and head-level paradigms that have dominated the field.

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.

We investigate the expressive power of hybrid recurrent-attention decoders, a class of architectures used in recent open-source language models such as Qwen3-Next and its successors. These models combine Gated Attention heads with recurrent Gated DeltaNet heads. Is there a formal advantage, in terms of model expressivity or efficiency, to such a hybrid architecture? We show that there is. We define parity-conditioned retrieval task and show that under constant-precision assumption, a Qwen-style hybrid of Gated DeltaNet and Gated Attention solves this task with a constant scratchpad, or equivalently $O(1)$ chain-of-thought steps. In contrast, no similar solution exists for pure Gated DeltaNet models, while pure Gated Attention requires at least a polynomial scratchpad.

Standard Mixture-of-Experts (MoE) transformers route tokens to expert subnetworks within each layer, but the layer structure itself remains monolithic. We introduce Mixture of Layers (MoL), which replaces full-width transformer blocks (d_model) with K parallel thin blocks at reduced dimensionality (d_thin << d_model), connected via learned down/up projections and composed via top-k block routing. Scaling sparse block routing to many blocks creates an attention coverage problem, as each block sees fewer tokens. We address this by introducing hybrid attention, which pairs one shared softmax block for global context with Gated DeltaNet linear attention in routed blocks.

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.

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 8$\times$A800 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 $2.8\times$ and $2.0\times$ in the prefill and decode stages.

Post-training hybridization of large language models (LLMs) often replaces quadratic self-attention with sliding-window attention (SWA) to reduce KV cache usage and improve latency. Existing hybridization schemes are typically defined either at the layer level (e.g., interleaving) or at the head level via static rankings from local to global. Layer-level schemes ignore that local and global dependencies are routed through heads within the same layer, while static head-level rankings suffer from entanglement: a head's local/global behavior can change after hybridization. We propose BOSCH, Black-box Binary Optimization for Short-context Head Selection, a training-free method that formulates the problem as a Large Neighborhood Search and decomposes it into three subproblems: (i) layer-importance detection via small-budget black-box probes, (ii) adaptive per-layer SWA-ratio assignment based on these sensitivities, and (iii) grouped head-level optimization within ratio buckets. Extensive experiments on 4 LLMs ranging from 1.7B to 30B parameters, across 4 SWA ratios, show that BOSCH consistently outperforms layer-level heuristics and 6 strong static head-level methods, with larger gains at higher SWA ratios. Under continual pretraining, BOSCH recover original long-context performance faster and to a higher level. Analysis of the selected heads reveals substantial turnover for BOSCH across different SWA ratios, underscoring the importance of performing head-level selection for each target ratio rather than relying on fixed locality rankings.

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.

The attention mechanism has been the core component in modern transformer architectures. However, the computation of standard full attention scales quadratically with the sequence length, serving as a major bottleneck in long-context language modeling. Sliding window attention restricts the context length for better efficiency at the cost of narrower receptive fields. While existing efforts attempt to take the benefits from both sides by building hybrid models, they often resort to static, heuristically designed alternating patterns that limit efficient allocation of computation in various scenarios. In this paper, we propose Switch Attention (SwiAttn), a novel hybrid transformer that enables dynamic and fine-grained routing between full attention and sliding window attention. For each token at each transformer layer, SwiAttn dynamically routes the computation to either a full-attention branch for global information aggregation or a sliding-window branch for efficient local pattern matching. An adaptive regularization objective is designed to encourage the model towards efficiency. Moreover, we adopt continual pretraining to optimize the model, transferring the full attention architecture to the hybrid one. Extensive experiments are conducted on twenty-three benchmark datasets across both regular (4K) and long (32K) context lengths, demonstrating the effectiveness of the proposed method.

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.

Jet flavor tagging plays an important role in precise Standard Model measurement enabling the extraction of mass dependence in jet-quark interaction and quark-gluon plasma (QGP) interactions. They also enable inferring the nature of particles produced in high-energy particle collisions that contain heavy quarks. The classification of bottom jets is vital for exploring new Physics scenarios in proton-proton collisions. In this research, we present a hybrid deep learning architecture that integrates edge convolutions with transformer self-attention mechanisms, into one single architecture called the Edge Convolution Transformer (ECT) model for bottom-quark jet tagging. ECT processes track-level features (impact parameters, momentum, and their significances) alongside jet-level observables (vertex information and kinematics) to achieve state-of-the-art performance. The study utilizes the ATLAS simulation dataset. We demonstrate that ECT achieves 0.9333 AUC for b-jet versus combined charm and light jet discrimination, surpassing ParticleNet (0.8904 AUC) and the pure transformer baseline (0.9216 AUC). The model maintains inference latency below 0.060 ms per jet on modern GPUs, meeting the stringent requirements for real-time event selection at the LHC. Our results demonstrate that hybrid architectures combining local and global features offer superior performance for challenging jet classification tasks. The proposed architecture achieves good results in b-jet tagging, particularly excelling in charm jet rejection (the most challenging task), while maintaining competitive light-jet discrimination comparable to pure transformer models.

Audio representation learning typically evaluates design choices such as input frontend, sequence backbone, and sequence length in isolation. We show that these axes are coupled, and conclusions from one setting often do not transfer to others. We introduce HELIX, a controlled framework comparing pure Mamba, pure attention, and a minimal hybrid with a single attention bottleneck. All models are parameter-matched at about 8.3M parameters to isolate architectural effects. Across six datasets, we find that the preferred input representation depends on the backbone, and that attention hurts performance on short, stationary audio but becomes important at longer sequence lengths. On a 5-minute speaker identification task with 30,000 tokens, pure attention fails with out-of-memory errors, while HELIX closes an 11.5-point gap over pure Mamba.

We identify a routing paradox in hybrid sequence models: content-based routing - deciding which tokens deserve expensive attention - requires pairwise computation, and this requirement is inescapable. Through 20+ controlled experiments across three tasks, multiple scales (200K to 1.4B parameters), and 15+ routing mechanisms, we map the routing landscape exhaustively. Every system that achieves high routing precision does so through pairwise token comparison. Every mechanism that avoids pairwise computation fails: recurrent models (Mamba-1.4B: 29%), memory banks (12%), bandits (0.7-3.6%), contrastive pretraining (1.6%), and 12 other approaches all cluster at 1-29%. Routing needs two ingredients: (1) per-token representations with bidirectional context and (2) pairwise token comparison. Bidirectional Mamba (O(n)) + pairwise comparison achieves 99.5%; replacing the full pairwise router with rank-1 projection improves this to 99.7%. Adding one bidirectional layer to frozen Pythia-1B recovers 99.4% routing. Six different O(n) preprocessing mechanisms (bidirectional Mamba, Perceiver inducing points, causal attention with E2E training, sparse attention, bidirectional attention, rank-1 projection) all succeed; global mean pooling (1.9%) and Fourier mixing (0.9%) fail. The routing signal occupies a ~34-dimensional latent subspace, invisible to cosine similarity. Non-learned indices (Bloom filter: 90.9%; BM25: 82.7%) bypass the bottleneck for exact/keyword matching. Combining O(n) bidirectional Mamba with rank-1 pairwise projection yields 99.7% routing at linear inference cost.

Recurrent neural networks (RNNs) and self-attention are both widely used sequence-mixing layers that maintain an internal memory. However, this memory is constructed using two orthogonal mechanisms: RNNs compress the entire past into a fixed-size state, whereas self-attention's state stores every past time step growing its state (the KV cache) linearly with the sequence length. This results in orthogonal strengths and weaknesses. Self-attention layers excel at retrieving information in the context but have large memory and computational costs, while RNNs are more efficient but degrade over longer contexts and underperform for precise recall tasks. Prior work combining these mechanisms has focused primarily on naively interleaving them to reduce computational cost without regard to their complementary mechanisms. We propose the Hybrid Associative Memory (HAM) layer, which combines self-attention and RNNs while leveraging their individual strengths: the RNN compresses the entire sequence, while attention supplements it only with information that is difficult for the RNN to predict, which is hence the most valuable information to explicitly store. HAM layers enable data-dependent growth of the KV cache, which can be precisely controlled by the user with a single, continuous threshold. We find that this fine-grained control of the KV cache growth rate has a smooth trade-off with loss and performance. Empirically, we show that our hybrid architecture offers strong, competitive performance relative to RNNs and Transformers even at substantially lower KV-cache usage.

Hybrid CNN-Transformer architectures achieve strong results in image super-resolution, but scaling attention windows or convolution kernels significantly increases computational cost, limiting deployment on resource-constrained devices. We present UCAN, a lightweight network that unifies convolution and attention to expand the effective receptive field efficiently. UCAN combines window-based spatial attention with a Hedgehog Attention mechanism to model both local texture and long-range dependencies, and introduces a distillation-based large-kernel module to preserve high-frequency structure without heavy computation. In addition, we employ cross-layer parameter sharing to further reduce complexity. On Manga109 ($4\times$), UCAN-L achieves 31.63 dB PSNR with only 48.4G MACs, surpassing recent lightweight models. On BSDS100, UCAN attains 27.79 dB, outperforming methods with significantly larger models. Extensive experiments show that UCAN achieves a superior trade-off between accuracy, efficiency, and scalability, making it well-suited for practical high-resolution image restoration.

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.

Structured dilated attention has an appealing inference-time efficiency knob: it reduces the FLOPs of attention and the KV cache size by a factor of the dilation size D, while preserving long-range connectivity. While prior work studies it by training each configuration from scratch, directly sparsifying a pretrained attention model into a dilated pattern leads to severe accuracy degradation, preventing flexible reuse across inference scenarios. We introduce RAT+, a dense-pretraining architecture that augments attention with full-sequence recurrence and active recurrence learning. A single RAT+ model is pretrained densely once and can then be flexibly switched at inference time to dilated attention (optionally with local windows) or hybrid layer/head compositions, requiring only a short 1B-token resolution adaptation rather than retraining separate sparse models. At 1.5B parameters trained on 100B tokens, RAT+ closely matches dense accuracy at D = 16, and drops by about 2-3 points at D = 64 on commonsense reasoning and LongBench tasks. We further scale to 2.6B and 7.6B parameters and observe even more promising performance (e.g., a 1-point average accuracy loss with a 64x reduction in attention FLOPs and KV cache size). Code is available at https://github.com/wimh966/rat-plus.

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

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.