STILL: Selecting Tokens for Intra-Layer Hybrid Attention to Linearize LLMs

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.

More Than a Quick Glance: Overcoming the Greedy Bias in KV-Cache Compression

While Large Language Models (LLMs) can theoretically support extensive context windows, their actual deployment is constrained by the linear growth of Key-Value (KV) cache memory. Prevailing compression strategies mitigate this through various pruning mechanisms, yet trade-off semantic recall for memory efficiency. In this work, we present LASER-KV (Layer Accumulated Selection with Exact-LSH Recall), a framework designed to test the limits of KV compression under a strict accumulative budgeting policy. We deviate from the standard fixed summary size approach by implementing a block-wise accumulation strategy governed by a protection divisor (n). This allows us to isolate the effects of compression from sliding window artifacts. Our experiments on the Babilong benchmark reveal performance degradation in previous compression methods by 15-30% on various long context tasks. LASER-KV maintains stable performance, achieving superior accuracies by a margin of upto 10% at 128k. These findings challenge the prevailing assumption that attention scores alone are a sufficient proxy for token utility.

Power-based Partial Attention: Bridging Linear-Complexity and Full Attention

It is widely accepted from transformer research that "attention is all we need", but the amount of attention required has never been systematically quantified. Is quadratic $O(L^2)$ attention necessary, or is there a sub-quadratic attention mechanism that can achieve comparable performance? To answer this question, we introduce power-based partial attention (PPA), an attention mechanism of order $O(L^{1+p})$, where $0 \leq p \leq 1$, such that $p=0$ corresponds to sliding window attention with linear complexity, and $p=1$ corresponds to full attention. With this attention construction, we can explore how transformer architecture performance varies as a function of the attention scaling behavior controlled by $p$. The overall trend from our experiments shows an S-curve-like behavior where the performance transitions from sliding-window (linear-complexity) attention to full attention over a narrow window of $p$ values, and plateaus as $p$ approaches $1$. In our experiments, we show that there exists $0<p<1$ such that $O(L^{1+p})$ attention is sufficient to achieve similar results as $O(L^2)$ full attention.

Window-Diffusion: Accelerating Diffusion Language Model Inference with Windowed Token Pruning and Caching

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

EEG-Titans: Long-Horizon Seizure Forecasting via Dual-Branch Attention and Neural Memory

Accurate epileptic seizure prediction from electroencephalography (EEG) remains challenging because pre-ictal dynamics may span long time horizons while clinically relevant signatures can be subtle and transient. Many deep learning models face a persistent trade-off between capturing local spatiotemporal patterns and maintaining informative long-range context when operating on ultralong sequences. We propose EEG-Titans, a dualbranch architecture that incorporates a modern neural memory mechanism for long-context modeling. The model combines sliding-window attention to capture short-term anomalies with a recurrent memory pathway that summarizes slower, progressive trends over time. On the CHB-MIT scalp EEG dataset, evaluated under a chronological holdout protocol, EEG-Titans achieves 99.46% average segment-level sensitivity across 18 subjects. We further analyze safety-first operating points on artifact-prone recordings and show that a hierarchical context strategy extending the receptive field for high-noise subjects can markedly reduce false alarms (down to 0.00 FPR/h in an extreme outlier) without sacrificing sensitivity. These results indicate that memory-augmented long-context modeling can provide robust seizure forecasting under clinically constrained evaluation

Forecasting Continuum Intensity for Solar Active Region Emergence Prediction using Transformers

Early and accurate prediction of solar active region (AR) emergence is crucial for space weather forecasting. Building on established Long Short-Term Memory (LSTM) based approaches for forecasting the continuum intensity decrease associated with AR emergence, this work expands the modeling with new architectures and targets. We investigate a sliding-window Transformer architecture to forecast continuum intensity evolution up to 12 hours ahead using data from 46 ARs observed by SDO/HMI. We conduct a systematic ablation study to evaluate two key components: (1) the inclusion of a temporal 1D convolutional (Conv1D) front-end and (2) a novel `Early Detection' architecture featuring attention biases and a timing-aware loss function. Our best-performing model, combining the Early Detection architecture without the Conv1D layer, achieved a Root Mean Square Error (RMSE) of 0.1189 (representing a 10.6% improvement over the LSTM baseline) and an average advance warning time of 4.73 hours (timing difference of -4.73h), even under a stricter emergence criterion than previous studies. While the Transformer demonstrates superior aggregate timing and accuracy, we note that this high-sensitivity detection comes with increased variance compared to smoother baseline models. However, this volatility is a necessary trade-off for operational warning systems: the model's ability to detect micro-changes in precursor signals enables significantly earlier detection, outweighing the cost of increased noise. Our results demonstrate that Transformer architectures modified with early detection biases, when used without temporal smoothing layers, provide a high-sensitivity alternative for forecasting AR emergence that prioritizes advance warning over statistical smoothness.

LiQSS: Post-Transformer Linear Quantum-Inspired State-Space Tensor Networks for Real-Time 6G

Proactive and agentic control in Sixth-Generation (6G) Open Radio Access Networks (O-RAN) requires control-grade prediction under stringent Near-Real-Time (Near-RT) latency and computational constraints. While Transformer-based models are effective for sequence modeling, their quadratic complexity limits scalability in Near-RT RAN Intelligent Controller (RIC) analytics. This paper investigates a post-Transformer design paradigm for efficient radio telemetry forecasting. We propose a quantum-inspired many-body state-space tensor network that replaces self-attention with stable structured state-space dynamics kernels, enabling linear-time sequence modeling. Tensor-network factorizations in the form of Tensor Train (TT) / Matrix Product State (MPS) representations are employed to reduce parameterization and data movement in both input projections and prediction heads, while lightweight channel gating and mixing layers capture non-stationary cross-Key Performance Indicator (KPI) dependencies. The proposed model is instantiated as an agentic perceive-predict xApp and evaluated on a bespoke O-RAN KPI time-series dataset comprising 59,441 sliding windows across 13 KPIs, using Reference Signal Received Power (RSRP) forecasting as a representative use case. Our proposed Linear Quantum-Inspired State-Space (LiQSS) model is 10.8x-15.8x smaller and approximately 1.4x faster than prior structured state-space baselines. Relative to Transformer-based models, LiQSS achieves up to a 155x reduction in parameter count and up to 2.74x faster inference, without sacrificing forecasting accuracy.

MiMo-V2-Flash Technical Report

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

UDPNet: Unleashing Depth-based Priors for Robust Image Dehazing

Image dehazing has witnessed significant advancements with the development of deep learning models. However, a few methods predominantly focus on single-modal RGB features, neglecting the inherent correlation between scene depth and haze distribution. Even those that jointly optimize depth estimation and image dehazing often suffer from suboptimal performance due to inadequate utilization of accurate depth information. In this paper, we present UDPNet, a general framework that leverages depth-based priors from large-scale pretrained depth estimation model DepthAnything V2 to boost existing image dehazing models. Specifically, our architecture comprises two typical components: the Depth-Guided Attention Module (DGAM) adaptively modulates features via lightweight depth-guided channel attention, and the Depth Prior Fusion Module (DPFM) enables hierarchical fusion of multi-scale depth map features by dual sliding-window multi-head cross-attention mechanism. These modules ensure both computational efficiency and effective integration of depth priors. Moreover, the intrinsic robustness of depth priors empowers the network to dynamically adapt to varying haze densities, illumination conditions, and domain gaps across synthetic and real-world data. Extensive experimental results demonstrate the effectiveness of our UDPNet, outperforming the state-of-the-art methods on popular dehazing datasets, such as 0.85 dB PSNR improvement on the SOTS dataset, 1.19 dB on the Haze4K dataset and 1.79 dB PSNR on the NHR dataset. Our proposed solution establishes a new benchmark for depth-aware dehazing across various scenarios. Pretrained models and codes will be released at our project https://github.com/Harbinzzy/UDPNet.

Spectral-Window Hybrid (SWH)

Scaling sequence modeling to extreme contexts requires balancing computational efficiency with representational expressivity. While Transformers provide precise retrieval via the attention mechanism, their quadratic $\mathcal{O}(T^2)$ complexity limits their application to long-horizon tasks. In this work, we propose the \textbf{Spectral-Window Hybrid (SWH)}, an architecture that decouples sequence modeling into two \textit{parallel} streams: a global branch utilizing the Convolution Theorem to model long-range decay dynamics in $\mathcal{O}(T \log T)$ time, and a local branch employing sliding-window attention for token interactions within a bounded context. By aggregating these representations, SWH avoids the computational bottleneck of global attention while retaining local precision. We demonstrate that SWH matches the perplexity of standard Transformers on short contexts while enabling efficient linear scaling to extended sequences. The code is available at https://github.com/VladimerKhasia/SWH