Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization

Transformer models face scalability challenges in causal language modeling (CLM) due to inefficient memory allocation for growing key-value (KV) caches, which strains compute and storage resources. Existing methods like Grouped Query Attention (GQA) and token-level KV optimization improve efficiency but rely on rigid resource allocation, often discarding "low-priority" tokens or statically grouping them, failing to address the dynamic spectrum of token importance. We propose mixSGA, a novel mixture-of-expert (MoE) approach that dynamically optimizes token-wise computation and memory allocation. Unlike prior approaches, mixSGA retains all tokens while adaptively routing them to specialized experts with varying KV group sizes, balancing granularity and efficiency. Our key novelties include: (1) a token-wise expert-choice routing mechanism guided by learned importance scores, enabling proportional resource allocation without token discard; (2) weight-sharing across grouped attention projections to minimize parameter overhead; and (3) an auxiliary loss to ensure one-hot routing decisions for training-inference consistency in CLMs. Extensive evaluations across Llama3, TinyLlama, OPT, and Gemma2 model families show mixSGA's superiority over static baselines. On instruction-following and continued pretraining tasks, mixSGA achieves higher ROUGE-L and lower perplexity under the same KV budgets.

Unleashing Diffusion and State Space Models for Medical Image Segmentation

Existing segmentation models trained on a single medical imaging dataset often lack robustness when encountering unseen organs or tumors. Developing a robust model capable of identifying rare or novel tumor categories not present during training is crucial for advancing medical imaging applications. We propose DSM, a novel framework that leverages diffusion and state space models to segment unseen tumor categories beyond the training data. DSM utilizes two sets of object queries trained within modified attention decoders to enhance classification accuracy. Initially, the model learns organ queries using an object-aware feature grouping strategy to capture organ-level visual features. It then refines tumor queries by focusing on diffusion-based visual prompts, enabling precise segmentation of previously unseen tumors. Furthermore, we incorporate diffusion-guided feature fusion to improve semantic segmentation performance. By integrating CLIP text embeddings, DSM captures category-sensitive classes to improve linguistic transfer knowledge, thereby enhancing the model's robustness across diverse scenarios and multi-label tasks. Extensive experiments demonstrate the superior performance of DSM in various tumor segmentation tasks. Code is available at https://github.com/Rows21/KMax-Mamba.

Hardware-Efficient Attention for Fast Decoding

LLM decoding is bottlenecked for large batches and long contexts by loading the key-value (KV) cache from high-bandwidth memory, which inflates per-token latency, while the sequential nature of decoding limits parallelism. We analyze the interplay among arithmetic intensity, parallelization, and model quality and question whether current architectures fully exploit modern hardware. This work redesigns attention to perform more computation per byte loaded from memory to maximize hardware efficiency without trading off parallel scalability. We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfers without compromising model quality. We then introduce Grouped Latent Attention (GLA), a parallel-friendly latent attention paired with low-level optimizations for fast decoding while maintaining high model quality. Experiments show that GTA matches Grouped-Query Attention (GQA) quality while using roughly half the KV cache and that GLA matches Multi-head Latent Attention (MLA) and is easier to shard. Our optimized GLA kernel is up to 2$\times$ faster than FlashMLA, for example, in a speculative decoding setting when the query length exceeds one. Furthermore, by fetching a smaller KV cache per device, GLA reduces end-to-end latency and increases throughput in online serving benchmarks by up to 2$\times$.

XDementNET: An Explainable Attention Based Deep Convolutional Network to Detect Alzheimer Progression from MRI data

A common neurodegenerative disease, Alzheimer's disease requires a precise diagnosis and efficient treatment, particularly in light of escalating healthcare expenses and the expanding use of artificial intelligence in medical diagnostics. Many recent studies shows that the combination of brain Magnetic Resonance Imaging (MRI) and deep neural networks have achieved promising results for diagnosing AD. Using deep convolutional neural networks, this paper introduces a novel deep learning architecture that incorporates multiresidual blocks, specialized spatial attention blocks, grouped query attention, and multi-head attention. The study assessed the model's performance on four publicly accessible datasets and concentrated on identifying binary and multiclass issues across various categories. This paper also takes into account of the explainability of AD's progression and compared with state-of-the-art methods namely Gradient Class Activation Mapping (GradCAM), Score-CAM, Faster Score-CAM, and XGRADCAM. Our methodology consistently outperforms current approaches, achieving 99.66\% accuracy in 4-class classification, 99.63\% in 3-class classification, and 100\% in binary classification using Kaggle datasets. For Open Access Series of Imaging Studies (OASIS) datasets the accuracies are 99.92\%, 99.90\%, and 99.95\% respectively. The Alzheimer's Disease Neuroimaging Initiative-1 (ADNI-1) dataset was used for experiments in three planes (axial, sagittal, and coronal) and a combination of all planes. The study achieved accuracies of 99.08\% for axis, 99.85\% for sagittal, 99.5\% for coronal, and 99.17\% for all axis, and 97.79\% and 8.60\% respectively for ADNI-2. The network's ability to retrieve important information from MRI images is demonstrated by its excellent accuracy in categorizing AD stages.

Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Grouping First, Attending Smartly: Training-Free Acceleration for Diffusion Transformers

Diffusion-based Transformers have demonstrated impressive generative capabilities, but their high computational costs hinder practical deployment, for example, generating an $8192\times 8192$ image can take over an hour on an A100 GPU. In this work, we propose GRAT (\textbf{GR}ouping first, \textbf{AT}tending smartly), a training-free attention acceleration strategy for fast image and video generation without compromising output quality. The key insight is to exploit the inherent sparsity in learned attention maps (which tend to be locally focused) in pretrained Diffusion Transformers and leverage better GPU parallelism. Specifically, GRAT first partitions contiguous tokens into non-overlapping groups, aligning both with GPU execution patterns and the local attention structures learned in pretrained generative Transformers. It then accelerates attention by having all query tokens within the same group share a common set of attendable key and value tokens. These key and value tokens are further restricted to structured regions, such as surrounding blocks or criss-cross regions, significantly reducing computational overhead (e.g., attaining a \textbf{35.8$\times$} speedup over full attention when generating $8192\times 8192$ images) while preserving essential attention patterns and long-range context. We validate GRAT on pretrained Flux and HunyuanVideo for image and video generation, respectively. In both cases, GRAT achieves substantially faster inference without any fine-tuning, while maintaining the performance of full attention. We hope GRAT will inspire future research on accelerating Diffusion Transformers for scalable visual generation.

Stable Reinforcement Learning for Efficient Reasoning

The success of Deepseek-R1 has drawn the LLM community's attention to reinforcement learning (RL) methods like GRPO. However, such rule-based 0/1 outcome reward methods lack the capability to regulate the intermediate reasoning processes during chain-of-thought (CoT) generation, leading to severe overthinking phenomena. In response, recent studies have designed reward functions to reinforce models' behaviors in producing shorter yet correct completions. Nevertheless, we observe that these length-penalty reward functions exacerbate RL training instability: as the completion length decreases, model accuracy abruptly collapses, often occurring early in training. To address this issue, we propose a simple yet effective solution GRPO-$\lambda$, an efficient and stabilized variant of GRPO, which dynamically adjusts the reward strategy by monitoring the correctness ratio among completions within each query-sampled group. A low correctness ratio indicates the need to avoid length penalty that compromises CoT quality, triggering a switch to length-agnostic 0/1 rewards that prioritize reasoning capability. A high ratio maintains length penalties to boost efficiency. Experimental results show that our approach avoids training instability caused by length penalty while maintaining the optimal accuracy-efficiency trade-off. On the GSM8K, GPQA, MATH-500, AMC 2023, and AIME 2024 benchmarks, it improves average accuracy by 1.48% while reducing CoT sequence length by 47.3%.

MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection

3D Anomaly Detection (AD) is a promising means of controlling the quality of manufactured products. However, existing methods typically require carefully training a task-specific model for each category independently, leading to high cost, low efficiency, and weak generalization. Therefore, this paper presents a novel unified model for Multi-Category 3D Anomaly Detection (MC3D-AD) that aims to utilize both local and global geometry-aware information to reconstruct normal representations of all categories. First, to learn robust and generalized features of different categories, we propose an adaptive geometry-aware masked attention module that extracts geometry variation information to guide mask attention. Then, we introduce a local geometry-aware encoder reinforced by the improved mask attention to encode group-level feature tokens. Finally, we design a global query decoder that utilizes point cloud position embeddings to improve the decoding process and reconstruction ability. This leads to local and global geometry-aware reconstructed feature tokens for the AD task. MC3D-AD is evaluated on two publicly available Real3D-AD and Anomaly-ShapeNet datasets, and exhibits significant superiority over current state-of-the-art single-category methods, achieving 3.1\% and 9.3\% improvement in object-level AUROC over Real3D-AD and Anomaly-ShapeNet, respectively. The source code will be released upon acceptance.

Geometry-Informed Neural Operator Transformer

Machine-learning-based surrogate models offer significant computational efficiency and faster simulations compared to traditional numerical methods, especially for problems requiring repeated evaluations of partial differential equations. This work introduces the Geometry-Informed Neural Operator Transformer (GINOT), which integrates the transformer architecture with the neural operator framework to enable forward predictions for arbitrary geometries. GINOT encodes the surface points cloud of a geometry using a sampling and grouping mechanism combined with an attention mechanism, ensuring invariance to point order and padding while maintaining robustness to variations in point density. The geometry information is seamlessly integrated with query points in the solution decoder through the attention mechanism. The performance of GINOT is validated on multiple challenging datasets, showcasing its high accuracy and strong generalization capabilities for complex and arbitrary 2D and 3D geometries.

Cost-Optimal Grouped-Query Attention for Long-Context LLMs

Building effective and efficient Transformer-based large language models (LLMs) has recently become a research focus, requiring maximizing model language capabilities and minimizing training and deployment costs. Existing efforts have primarily described complex relationships among model performance, parameter size, and data size, as well as searched for the optimal compute allocation to train LLMs. However, they overlook the impacts of context length and attention head configuration (the number of query and key-value heads in grouped-query attention) on training and inference. In this paper, we systematically compare models with different parameter sizes, context lengths, and attention head configurations in terms of model performance, computational cost, and memory cost. Then, we extend the existing scaling methods, which are based solely on parameter size and training compute, to guide the construction of cost-optimal LLMs during both training and inference. Our quantitative scaling studies show that, when processing sufficiently long sequences, a larger model with fewer attention heads can achieve a lower loss while incurring lower computational and memory costs. Our findings provide valuable insights for developing practical LLMs, especially in long-context processing scenarios. We will publicly release our code and data.