Hybrid Attention

The architectural design of Large Language Models (LLMs) is shifting from traditional full-attention-only models to hybrid models, which combine different attention modules to improve long-context efficiency and performance in length extrapolation and context extension. To explain why hybrid models work and how to design them better, we propose Mechanics of Long-Context Hybrid Models. As Part 1.1 of this series, we begin with hybrids of full attention and either sliding-window attention (SWA) or gated variants of linear attention (LA), represented by GLA and GDN. We first observe a Seesaw Effect in Context Extension: LA hybrids benefit more from long-context continual pretraining, whereas SWA hybrids perform better under length extrapolation. We attribute this behavior to differences in the positional inductive biases induced by these attention mechanisms. We find that SWA hybrids suffer from a Short-Context Learning Trap, Short-Window Weariness, and Long-Window Laziness, and require extended windows to enhance performance in continual long-context pretraining. For LA hybrids, we summarize the Matthew Effect of Hybrid Position Extrapolation and propose Sliding-Window Linear Attention, achieving 16$\times$ training-free length extrapolation while maintaining 100% accuracy on NIAH-SK1 in 64k context length.

Hybrid architectures combining linear sequence models with softmax attention provide an effective balance between efficient long-context modeling and precise token retrieval. Existing designs such as Native Hybrid Attention (NHA) combine compressed long-term states with sliding-window attention, but their exact attention is restricted to a fixed local window. In this work, we introduce Prefix-State Hybrid Block Attention (PHBA), which replaces local sliding-window attention with top-k block-sparse retrieval and couples each retrieved block with a compact prefix state summarizing its preceding context. The prefix states are constructed by a gated linear recurrence at block boundaries and retrieved together with the corresponding token blocks, allowing the model to combine precise long-range evidence with compressed historical context within a unified layer. We further develop a hardware-aware Triton implementation that streams routed token blocks and prefix states without materializing large intermediate tensors. Experiments show that PHBA improves long-context and retrieval performance over strong linear and hybrid baselines while retaining efficient training and inference.

Looped language models apply the same stack of layers T times to each token, which deepens the model without adding parameters but multiplies its key-value (KV) cache by T. The larger cache limits how many sequences a GPU can decode at once and slows each decoding step, which reads the whole cache. We propose Hybrid Latent Attention (HLA), which keeps exact keys and values within a sliding window of W recent tokens and stores each older token as a compact latent that the query of each loop reads directly, without reconstructing keys and values. We uptrain HLA on Ouro looped models (T=4) with 1.4B and 2.6B parameters, keeping the pretrained weights frozen and training only the added parameters to reproduce the original attention. The cache shrinks by 10.7x per token, fitting 4.0-8.8x as many concurrent sequences per GPU, and decoding throughput improves by 2.5x at 1K-token contexts and by up to 7.4x at 16K. HLA retains over 97% of the original accuracy on math, knowledge and reasoning benchmarks, and 96-100% on long-context retrieval up to 16K tokens. After supervised fine-tuning, it performs on par with the fine-tuned original model on competition-level math.

Recurrent-attention hybrid language models (LMs), which interleave attention and recurrent layers, are increasingly used to combine the efficiency of the recurrent layers with the strong performance of attention layers. Prior work suggests that attention and recurrent layers offer complementary pathways to use past information: attention supports precise memory recall from earlier tokens, while recurrent layers support consolidation of disparate information over long contexts. However, we observe that simply having access to both pathways does not mean that hybrid LMs are effectively using them. We find that they rely substantially more on attention than on the recurrent state. Standard supervised fine-tuning improves overall performance but does not improve how the two memory pathways are coordinated: the model becomes more reliant on information propagated by attention layers, while its use of information propagated by recurrent layers remains limited. To encourage better coordination between the two memory pathways, we add an auxiliary loss that limits attention's access to earlier context while the recurrent state propagates through the full sequence. This objective encourages the model to retain and use information through the recurrent pathway alongside attention. It improves overall performance, with particularly strong gains on tasks involving longer contexts or requiring information aggregation, consistent with the strengths of recurrent layers observed in analysis. Crucially, this imbalance and the benefit of our auxiliary loss generalize: they apply to multiple recurrent-attention LMs in question-answering and agentic tasks, as well as to attention-based LMs that combine different forms of memory. Together, our findings show that simply providing multiple memory pathways does not ensure their effective use, and that targeted supervision is needed to better coordinate them.

Long-horizon video world models require persistent memory to preserve scene consistency over extended rollouts. Softmax attention retains the full generation history through a growing KV cache, whereas recurrent linear attention compresses history into fixed-size states with substantially lower memory cost. However, we identify severe long-range forgetting in Gated DeltaNet (GDN), where information from distant but relevant scenes is progressively attenuated by subsequent state updates. To address this limitation, we propose HLA-WM, a training-free hybrid linear-attention framework that combines coarse-grained geometry-guided retrieval with fine-grained recurrent linear-state computation. HLA-WM exploits the affine structure of GDN to cache compact chunk-wise transition summaries, retrieve scene-relevant historical chunks using camera geometry, and recompose them into query-specific recurrent states. On the $60$-second SANA-WM-Bench, HLA-WM improves all six aggregate revisit-consistency and camera-control metrics of the base autoregressive generator without additional training, including a $0.74$ dB PSNR gain and a $28.5\%$ reduction in rotation error. The improvements persist after downstream refinement and generalize to MBench-A, where HLA-WM consistently improves all three revisit-consistency metrics across all four subsets and all evaluated inference modes over $547$ samples. At a $60$-second context, HLA-WM reduces historical-state memory by $12\times$ relative to full KV caching while incurring at most a $1.6\%$ reduction in inference throughput. These results demonstrate that selectively addressable recurrent memory can improve long-range scene recall while preserving the efficiency advantages of GDN. Project page: https://caesarhhh.github.io/hla-wm/

Developmental psychology holds that certain priors are given to infants prior to experience rather than induced from data, and that the influence of such priors is suppressed under strong, well-constrained conditions but reasserts itself under weak ones. We ask whether an analogous principle holds for the Transformer: can the activation function given to attention heads serve as an intrinsic inductive bias? We propose Mixture of Function Attention (MoFA), a parameter-free modification to multi-head attention that fixes a ratio of softmax and sigmoid heads before training. Across five ratios, a 124M-parameter GPT-2 model, and five seeds, we find that this given ratio has little effect in-distribution -- differences between ratios are statistically negligible for moderate mixtures and remain small even at the extremes -- but its influence re-emerges sharply under zero-shot distribution shift across 15 out-of-distribution domains. Perplexity gaps between ratios widen by more than an order of magnitude on several domains, and the best-performing ratio tracks a single axis of domain structure, separating short, informal text (softmax-favoring) from technical, long-form text (sigmoid-favoring), that explains 78.3% of the variance in domain response. This reorganization is visible at the head level: sigmoid heads show an accelerating drop in attention entropy as their ratio increases, while softmax heads respond more modestly, yielding a consistent division of labor between the two head types. Our results suggest that activation choice functions as a given prior whose influence is masked in-distribution and re-emerges out-of-distribution.

Language models increasingly use architectures that vary attention span and positional encoding across layers, such as applying RoPE with sliding-window attention and NoPE with global attention (SWA NoPE). However, how these choices shape in-context retrieval remains unclear. To study this question, we take a mechanistic view, tracing how positional encoding (PE) choice shapes the internal mechanisms models use for in-context retrieval. Across 22 open-weight models spanning eight families, we find that standard RoPE models rely primarily on positional retrieval, while PE hybrids shift toward semantic retrieval. We further show on a controlled pre-training ablation that confining positional encoding to local layers produces this semantic shift, degrading representations of positional information. Finally, we show that the reported long-context gains of PE hybrids mask a retrieval trade-off: SWA NoPE improves over RoPE on multiple-target retrieval and QA, but degrades when distinguishing competing keys. We show that these behavioral differences better track the mechanism shift from positional toward semantic mechanisms than a uniform improvement in long-context retrieval.

The landscape of satellite imagery time series datasets and boundary-pushing architectures for cropland segmentation has never been richer. However, in this gold rush, important truths are being missed on both fronts, as a drive for the most novel concepts or the largest datasets pushes finer details to the side. In this paper, we present our hybrid transformer-convolutional model, Cropland Parallel Attention and Refinement Network for Segmentation (PAtteRNS), the first model to use self-attention mechanisms separately for each of the temporal, spectral, and spatial aspects of Sentinel-2 multispectral SITS data. To achieve fully-factorised attention in our proposed model, we introduce a novel parallel transformer architecture which significantly reduces the computational complexity of triple-factorised self-attention. We validate our architecture with an in-depth ablation study, and analyse the performance of our model against state-of-the-art crop segmentation models on multiple tile-size variants of the popular PASTIS and MTLCC datasets. Our findings show our model to outperform all others in the task of crop class segmentation, verified across multiple important segmentation metrics, with especially strong performance against compared models seen in the often under-reported parcel delineation quality, for which we use the Boundary IoU metric. We also find that flawed class groupings within datasets can have a significant negative impact on model performance, and report that alternate tile-size variants of crop segmentation datasets produce results incomparable to one-another, invalidating fair comparison between model performance when trained on different tile-sizes. Based on these findings, we suggest further work is required to standardise best practices when constructing SITS crop segmentation datasets, and to enable future dynamic-tile-sizing for ideal model performance.

The attention operation is naively position invariant. However, positional information is fundamental to natural language, and therefore a variety of explicit position encodings have been developed in transformer-based models, such as rotary position encoding (RoPE). Although explicit position encodings have long been assumed to be required, recent methods that interleave local mixing layers, such as sliding window attention (SWA) and gated linear attention, while not encoding position (NoPE) in global attention layers has recently been shown to be successful at scale. How and why this approach works is not well-understood. In this paper, we develop an explanation of how hybrid models of this sort can implicitly encode position at global NoPE layers. Supported by both theoretical and empirical evidence, our central argument is that SWA and gated linear attention induce a recency bias in the residual stream that propagates to, and is selected by, the global attention logits. Moreover, in contrast to the implicit position encodings found in models with only global NoPE attention, in which positional information arises solely from the causal mask, the recency bias in hybrid models can be maintained across long sequences. In addition to deepening our understanding of how hybrid models encode position, these findings may provide insights for how to encode position in a way that can extrapolate to longer sequence lengths indefinitely.

In response to the growing demand for long sequences in agentic and reasoning use cases, many state-of-the-art LLMs combine multiple variants of attention to mitigate the quadratic complexity of traditional softmax attention. These hybrid attention LLMs aim to balance the strengths and limitations of full attention and alternatives based on recurrence. This work presents a first study of how hybrid attention impacts the multilinguality of LLMs. Beyond the impact on long sequences in poorly tokenized languages, our study is motivated by the possibility that the inductive biases of the recurrent state alter linguistic processing. Our interpretability analysis confirms this, showing that cross-lingual representations in hybrid models develop in patterns tied to the ordering of recurrent and full-attention layers. Across diverse models, we notably observe a pronounced spike in cross-lingual alignment around the first full-attention layer. These findings lead us to question the conventional ordering of attention layers. In distillation experiments on multilingual data, all alternative layer orderings outperform the standard throughout training, learning up to 2.5X faster. These stark, replicable results prompt our theory that multilingual models would benefit from starting with a full-attention layer rather than recurrent layers.

Leveraging the paradigm of autoregressive diffusion, text-conditioned interactive video world models aim to simulate temporally coherent environments guided by textual instructions. While enabling low-latency, long-duration generation is pivotal for embodied AI and simulation-based planning, current frameworks primarily rely on sliding-window mechanisms to bound computational complexity. However, this approach inherently sacrifices historical context, undermining the long-range interactive capabilities. Conversely, maintaining a full-history cache remains computationally prohibitive and memory-intensive: the quadratic complexity of attention leads to excessive computational overhead, while the linear growth of the KV cache inevitably leads to GPU memory saturation. To overcome these limitations, we propose WorldAttention, a system-oriented attention architecture that achieves high efficiency through the co-design of specialized attention kernels and hierarchical KV cache management. First, we introduce Hybrid Sparse Attention (HSA), which integrates linear global attention supplemented with head-adaptive sparse attention. Additionally, we design a Hierarchical KV Cache (HKV) that organizes historical KV pairs into semantically indexed pages across multi-tier memory, enabling fine-grained retrieval and controlled GPU residency. These two designs are supported by tailored kernels to effectively translate their theoretical efficiency into real-world performance. Extensive experiments on VBench-Long and InterVBench demonstrate that WorldAttention consistently surpasses prior state-of-the-art methods, achieving subject consistency scores of 0.9472 on VBench-Long and 0.9668 on InterVBench, respectively.

At each decoding step a language model attends over the key-value (KV) cache of every earlier token, so at long context the attention call is bounded by memory bandwidth. Sparse attention reads only a subset of keys chosen by a cheap score estimate, and most methods give the unread tokens zero weight. The output then draws on only a small fraction of the KV cache, and accuracy drops at small budgets, most on tasks that aggregate information across the context. An inverted-file product-quantization (IVF-PQ) index over the cached keys computes an approximate score for every indexed token in order to rank them; after ranking, those scores approximate the attention logits of the tokens left out. PQ-HSA (hybrid sparse-approximate attention) attends the selected tokens with their original keys and values, and the unselected tokens, the background, enter the same softmax through those scores, summed per inverted list and multiplied by the list's mean value. At 128K and a 1-2% retrieval budget, PQ-HSA is more accurate than Quest and SnapKV on Llama-3.1-8B and Qwen3-30B-A3B and stays close to full attention in macro accuracy; with the same selector, the background term raises macro accuracy on the 8B model from 0.71 to 0.83. In the same 128K setting, inside vLLM on one NVIDIA H20, the decode attention call runs 1.6x faster than the FlashAttention-3 kernel; the speedup grows with context length, and a cost model fitted on 8B to 30B models gives the context length at which it begins. A vLLM plugin runs PQ-HSA on two engine versions without changes to the engine source; code is available at https://github.com/KunmingSHAO/pqhsa_release.

Hybrid LLMs interleave full-attention layers with linear-attention layers to reduce long-context inference cost, but this structure complicates prefix caching. Full-attention KV caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing systems materialize recurrent-state checkpoints, restricting prefix reuse to checkpoint-aligned positions. We present SuffixReplay, the first prefix caching system that lets hybrid LLMs reuse cached prefixes at every cache-supported page boundary without materializing recurrent-state checkpoints. Our key insight is to just let linear states forget the distant past. Modern linear-attention mechanisms use recurrent decay and gating to attenuate the influence of old inputs. Therefore, instead of checkpointing every prefix boundary, SuffixReplay approximates the state at a matched boundary by replaying only a recent suffix of the layer's input hidden states, which we retain as anchors. At the algorithmic level, SuffixReplay combines layer-wise and token-wise anchor sparsity with a bounded replay budget to control storage, computation, and quality. At the system level, it uses an independently managed anchor sidecar and a pipelined replay path to overlap anchor movement and state reconstruction with the native serving pipeline. We evaluate SuffixReplay on three hybrid LLMs: OLMo-Hybrid-7B, Qwen3.5-4B, and Qwen3.6-27B-FP8. Across these models, SuffixReplay retains 91.4-100% of full-prefill quality on average across LongBench and RULER, while using only 0.36-0.51x the amortized per-token storage of SGLang's default 8192-token checkpoint cache. Integrated into SGLang, SuffixReplay reduces median TTFT by 15-70% on branching workloads, sustains 2.3-4.3x SGLang's throughput when the working set exceeds HBM, and matches SGLang on high-hit continuation traffic.

Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches where recurrent-state access becomes a major bottleneck. ReplaySSM amortizes state updates by buffering keys and values, but each new query still requires a full-state read even though the state remains unchanged between state updates. We observe that low-rank state-weighted query approximation accurately preserves state-read outputs. Although future queries are unknown, the basis vectors used to approximate them can be fixed offline. Based on this observation, we introduce SketchSSM, which preserves full-state updates while approximating reads. At each state update, SketchSSM reads the full state once to precompute outputs for these basis vectors, storing them in a compact sketch. Each subsequent decode step combines the sketch vectors with query-dependent coefficients to reconstruct the output without a full-state read. Across four Mamba-2-, GDN-, and KDA-based models, SketchSSM at a mean sketch rank of 8 reduces state-access traffic by approximately 10$\times$ while matching the average accuracy of the FP32 full-state baseline across four decode benchmarks, and preserves recall on four RULER retrieval tasks. At this rank on one NVIDIA B300, linear-attention kernel speedups over the Standard vLLM baseline reach 7.30$\times$, 5.02$\times$, and 5.24$\times$ for Mamba-2, GDN, and KDA, respectively, with up to 2.77$\times$ higher decode throughput on Nemotron 3 Super.

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

Time series foundation models (TSFMs) have recently delivered impressive zero-shot performance across diverse forecasting tasks. However, real-world decision-making frequently relies on \emph{irregular multivariate time series} (IMTS), where inconsistent inter-observation intervals and asynchronous sampling across variables coexist with informative missingness. Existing TSFMs handle such inputs either through imputation that injects spurious values or through index-based positional encodings that ignore continuous time. There is still a gap in the foundation model that follows the original IMTS patterns. In this paper, we propose a hybrid attention model that learns a unified time-aware patch representation for IMTS forecasting. We first design a \emph{time-aware patch encoding} that maps a variable number of intra-patch timestamps into a fixed-size embedding, producing a uniform format for irregular patches without resorting to imputation. We then introduce a \emph{time bias attention} mechanism that calibrates inter-patch temporal misalignment and asynchronous cross-channel dependencies as auxiliary attention offset. Finally, on top of a decoder-only Transformer backbone, we adopt a \emph{hybrid causal mask} that preserves a bidirectional full view over the historical context while keeping the forecast horizon strictly autoregressive. To support large-scale pretraining under irregular settings, we also curate VersaTSA, an archive of $30$B observations that retains the native sampling sparsity of its sources. Experiments on three IMTS benchmarks and a standard regular-MTS benchmark show that our model achieves state-of-the-art zero-shot performance on IMTS and remains competitive when transferred to regular forecasting.

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. GitHub code available at: https://github.com/OpenVDN/vdn-minimax-h3. Weights available at: https://huggingface.co/OpenVDN/vdn-minimax-h3

Hybrid architectures combining Full Attention (FA) and Linear Attention (LA) are increasingly prominent, yet their allocation remains heuristic. We seek an evidence-grounded basis in head-level functional organization learned by RoPE-based Transformers. Behavioral probes do not yield a complete taxonomy, so we propose two intervention metrics: RoPE Frequency Importance Score (RFIS), measuring how each frequency affects a head's attention distribution, and RoPE Positional Dependence (RPD), isolating dependence on rotary positional modulation. On Qwen3-series models and Llama3.1, RFIS suggests and RPD verifies a complete taxonomy of retrieval and positional heads separated by a salient mid-low-frequency band. Controlled Transformers show that this boundary follows the training-length positional scale; we term it the Global Positional Band (GPBand). The analysis suggests a potential cause of zero-shot length-extrapolation failure and yields two principles: positional modeling should operate only locally, with global access through position-independent retrieval; and both functions should be assigned at head granularity with layer-specific allocation. We instantiate them in Head-wise Hybrid Architecture (HwH), using NoPE FA for global retrieval and LA for local positional modeling. With an FA-to-LA ratio below 1:3, HwH retains strong language modeling and commonsense reasoning while improving retrieval and substantially strengthening zero-shot long-context extrapolation over Transformer, LA, and a layer-wise hybrid baseline. Ablations validate both principles and component roles, highlighting principled hybrid architecture design as a promising route toward future foundation models.

Hybrid linear-attention architectures have recently scaled to large open-weight models, offering quality competitive with full attention while substantially reducing key/value (KV) cache growth. However, their in-place recurrent-state updates complicate cache management: prefix reuse requires state checkpoints alongside full-attention KV, while storing state checkpoints in full increases memory pressure, leading to more evictions and repeated prefill. By analyzing the decay structure of Gated DeltaNet (GDN) and Kimi Delta Attention (KDA), we find that different heads and channels retain prefix information over markedly different timescales, which we term \emph{retention horizons}. This variation suggests substantial compression potential in persistent state checkpoints. Building on this observation, we introduce \emph{Decay-Aware State Compression} (DASC), which derives retention horizons from model weights, selects long-horizon state units, and packs them into a ragged state checkpoint layout. To integrate efficiently with tensor-parallel inference engines, DASC furtherly balances compressed state checkpoints across TP ranks. On reuse, DASC either zero-fills omitted units or refreshes them from a bounded suffix with additional compute cost. Across retrieval and end-to-end reasoning benchmarks on Kimi-Linear, conservative DASC configurations remain close to full caching while compressing KDA recurrent state checkpoints by $2.63\times$. Under fixed state checkpoint memory budgets, the resulting capacity gains reduce mean Time to First Token (TTFT) by 42.6% and improve input throughput by 68.4%. At larger compression ratio, suffix refresh recovers much of the accuracy lost to more aggressive omission, at the cost of additional replay computation. Qwen with GDN exhibits a similar quality--efficiency trend, showing that DASC extends from channel-wise KDA to head-wise GDN.

Hybrid large language models interleave full-attention layers with linear-attention layers to reduce the cost of long-context inference. This structure complicates prefix caching: full-attention key-value caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing hybrid prefix caching methods address this mismatch by storing recurrent-state checkpoints. As a result, token-level matches are directly usable only at positions aligned with stored checkpoints, constraining prefix reuse to a discrete set of boundaries. We present Tail-Replay, a prefix caching mechanism that enables unconstrained token-level prefix reuse in hybrid large language models. The key insight is that linear-attention mechanisms such as Gated DeltaNet can be viewed as a structured, lossy compression of the input prefix: gated recurrent updates progressively attenuate the contributions of earlier inputs. Consequently, the recurrent state of a matched prefix can be well approximated by replaying only a short, recent suffix of that prefix. Tail-Replay exploits this property by caching the exact full-attention key-value cache while omitting recurrent-state checkpoints. On a cache hit, it reconstructs the linear-attention states by replaying a short, recent suffix of the matched prefix. As a result, the reuse boundary is determined by the shared tokens rather than by recurrent-state checkpoints. We evaluate Tail-Replay on three Gated DeltaNet-based hybrid models using the LongBench and RULER benchmarks. With only a 5--10% replay budget, it retains 92.8--99.9% of full-prefill quality on LongBench and RULER. For serving efficiency, we evaluate time-to-first-token speedups across multiple matched-prefix lengths---8K, 16K, and 32K. The speedup grows with prefix length, reaching $9.1$--$14.3\times$ over full prefill at 32K.

Attention dominates token mixing, but it collapses relation formation and flow allocation into a single score-to-flow step. We introduce Relation, which separates them by first organizing pairwise evidence into explicit Self and Exchange relations and deriving information flow afterward. Relation first decides whether a token should rely on itself or draw from its history, and if it draws from history, where to look. This relational organization gives rise to Full Relation, FlashRelation, Linear Relation, and Hybrid Relation. Across matched decoder-only models, Full Relation achieves lower mean final-validation NLL than MHA and reaches the paired MHA final training loss with 4.5-7.3% fewer tokens. Structural diagnostics further show that Relation learns a distinct depth organization: the first layer acts as a current-token anchor and a high-rank router, while later layers shift strongly toward history. In a fixed-context reference benchmark, FlashRelation is 4.17-5.28x faster than the materialized Full Relation implementation. Across scale-matched production workloads, it reaches 89.7-92.9% of PyTorch FlashAttention throughput while executing the exact Full Relation operator. Hybrid Relation demonstrates that Full and Linear Relation layers can be composed within a single decoder. These results support a relation-first view of token mixing: ask Self, ask Others, then let Flow follow Relation.

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.

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.

We present a linearized form of 2-simplicial attention by rewriting the trilinear score as an inner product between a composite query and a key, so that the sum over one token axis takes the same form as ordinary softmax attention. We then approximate this sum with positive random features and store the entire past in a fixed-size state, while the second axis stays explicit over a short window of recent tokens. This enables us to achieve linear cost in sequence length combined with a global reach that windowed 2-simplicial attention lacks. We implement it with custom Triton kernels and combine it with Kimi Delta Attention to build a model with no softmax attention at all. Under matched compute, this model achieves the highest mean downstream accuracy among the compared architectures, and at 16k context it improves mean accuracy over a KDA hybrid while lowering LAMBADA perplexity from 715.6 to 602.6.

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 $1.7\times$ 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.

Hybrid-attention large language models combine full attention with recurrent linear attention to reduce long-context inference costs, yet their autoregressive decoding remains memory-bound. Tree speculative decoding offers an attractive acceleration path, but existing tree-speculation systems are designed around the key--value caches of full-attention models. On hybrid models, they traverse recurrent layers branch by branch and materialize a full state for every proposal node, causing verification latency and transient memory to scale poorly with tree and batch sizes. We present Bole, a kernel--runtime co-design that enables efficient tree speculation for hybrid-attention LLMs. Bole transforms the linear-attention recurrence into a tree-structured closed form and realizes it with a resource-efficient GPU kernel, verifying all proposal nodes in parallel and accelerating linear-attention tree verification by 3.4--7.7$\times$. It losslessly encodes speculative state updates as token-level factors and reconstructs only the state selected after sampling, reducing transient state memory by 82--99$\times$ and freeing GPU capacity for KV caches. Its integration into SGLang, a widely deployed production LLM serving engine, couples efficient state management with a batch-wide verification budget calibrated to the complete hybrid forward. Across four models, two GPU platforms, and diverse datasets, Bole delivers up to $4.72\times$ the offline decode throughput of autoregressive decoding and up to $2.03\times$ that of the strongest tree-speculative baseline. Under online agent workloads, it reduces TTFT and TPOT by up to $67.6$ and $49.9$, respectively, over the strongest tree-speculative baseline.

This paper presents an applied case study on hybrid visual attention estimation for human-robot interaction using an expressive robotic head based on the InMoov ecosystem. The proposed pipeline combines a fast geometric perception layer with an independent semantic perception layer based on a vision-language model. The geometric layer provides high-frequency face and head-pose information for temporal regulation, while the semantic layer receives only raw egocentric camera frames and produces contextual attention labels related to attention toward the robot, phone use, or attention elsewhere. These signals are integrated through a finite state machine that regulates adaptive interaction behavior, including activation, waiting, interaction resumption, and return to rest. The system was evaluated with 10 participants across 40 trials covering baseline and adaptive interaction conditions. Results show reliable interaction start across all trials, consistent pause behavior in the adaptive distraction condition, and non-redundant semantic information between the geometric and semantic outputs.

LLM serving is increasingly accelerated by position-independent caching (PIC). Existing PIC methods, however, are built for full-attention models, where a token-indexed KV cache underlies its core operations: matching reusable token chunks, concatenating their KV entries, and selectively recomputing a few tokens to restore cross-chunk context. Hybrid LLMs break these primitives---they replace most attention layers with linear recurrences that expose only a fixed-size state, leaving no token-indexed KV to concatenate or to locally repair. This raises a natural question: can PIC benefit hybrid models, and what would it take? We present LinearKV, a training-free hybrid-PIC framework. Its key insight is a \emph{decoupled initialization}: each linear layer maps its $K$ matched local states to a single initial state, while full-attention layers concatenate their KV as before. LinearKV is therefore compatible with existing PIC methods, reusing their token selection and recomputation as-is. Under this framework, we find that a \emph{single cached state} suffices as the linear layer's initializer. The algebraically principled alternative---composing all $K$ cached states into the exact full-prefix state, as concurrent work HYPIC does---is unnecessary and, on some architectures, even harmful. We compare the two across three hybrid models and three PIC selectors. On the two GDN models the two tie, both recovering most of full quality (up to $92\%$); on the Mamba-2 model, exact composition instead collapses under every selector---under EPIC, for instance, it recovers only $46.6\%$ of full quality, versus $86.8\%$ for a single cached block initializer. A single state initializer is also cheaper, cutting time-to-first-token to $0.46\times$ full prefill versus a further $5$--$17\%$ overhead for exact composition; results hold across LongBench QA and RULER at 8K--32K.

Visual generation increasingly requires high-resolution images, long videos, and multimodal context, making the quadratic cost of full attention prohibitive. We introduce Chimera, a hybrid visual diffusion backbone with a principled scaling recipe. Chimera processes text, image, and video tokens in one raster-ordered stream without positional embeddings. It combines Kimi Delta Attention (KDA) for long-context state tracking with O(N) complexity, interleaved Multi-head Latent Attention (MLA) for direct global interaction, and modality-aware short convolutions for local spatiotemporal context. Sparse Mixture-of-Experts (MoE) layers expand capacity while controlling activated compute. To scale this heterogeneous architecture, we introduce HeteroP, a module-wise scheme that transfers hyperparameters across width and depth according to each tensor's functional fan-in and model depth. HeteroP yields a consistently tuned family used to fit Chinchilla-style compute-optimal laws for activated model size, training-token count, and image-video data ratio. Guided by these laws, we train an 11B-parameter Chimera with 2B activated parameters. Experiments show three results. First, measured by pretraining diffusion loss, the dense backbone is 1.7x as compute-efficient as a matched full-attention Wan-2.1 2B baseline, while the complete system reaches 7.3x. Second, without length-specific fine-tuning, Chimera extrapolates zero-shot from 5-second training clips to 30-second videos, with only 6.5% FID degradation in the last five seconds. Third, the fitted laws show that compute-optimal image pretraining divides compute nearly evenly between activated model size and training-token count, whereas video pretraining modestly favors model size at higher budgets. These results establish a foundation for designing and scaling efficient long-context diffusion architectures.

Linear-attention and state-space language models compress the prefix into a fixed-size recurrent state, yielding O(1) memory at the cost of a lossy exact memory: when many key--value associations compete, earlier facts are overwritten and needle recall degrades. Inspired by Complementary Learning Systems, we give linear attention a hippocampal complement. HOLA (Hippocampal Linear Attention) keeps the usual delta-rule state as a compressive memory and adds a bounded exact KV cache, forming a semiparametric test-time memory: the state models linearly compressible structure, while the cache stores associations that should not be forced through that state. The cache writes without a learned eviction module, keeping tokens with large beta * ||e||, the prediction residual actually committed to the state; a decoupled RMSNorm-gamma cache read then turns these exact KV pairs into sharp retrieval rather than soft averaging. At 340M parameters trained on 15B SlimPajama tokens, HOLA lowers Wikitext perplexity from 27.32 to 22.92 (-16.1%), below a full-attention Transformer++ (26.88), and improves LAMBADA perplexity from 30.95 to 30.26. It also achieves the best linear in-context retrieval and remains much more robust than GDN or a matched HOLA+recency cache on RULER needle-in-a-haystack recall out to 32k tokens (16x its training length).