{"name":"RWKV papers catalog","url":"https://www.rwkv.cn/eco/papers","count":232,"lastUpdated":"2026-06-26","items":[{"id":26062401,"created_at":"2026-06-26T12:00:11+00:00","title":"ASSCG: Just-Right Gating over Chattering for Fast-Slow LLM Planning in Autonomous Driving","date":"2026-06-24","content":"基于RWKV构建ASSCG慢系统调用门控器，将自动驾驶快慢双系统中的LLM查询建模为序列决策，在查询、缓存和丢弃间选择，从而降低延迟并提升闭环规划表现。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260624-1.png","author":"Sining Ang","link":"https://arxiv.org/abs/2606.25509","category":["序列/强化","规划决策"],"conference_name":null,"conference_url":null},{"id":26061801,"created_at":"2026-06-26T12:00:10+00:00","title":"Neural Events: Discrete Asynchronous Autoencoders for Event-Based Vision","date":"2026-06-18","content":"基于RWKV-7构建离散异步事件编码器，用RWKV递归更新局部记忆并为每个事件输出离散码 logits，再经量化生成神经事件，从而压缩事件流并保留精确时序信息。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260618-1.png","author":"Roberto Pellerito","link":"https://arxiv.org/abs/2606.19835","category":["图像","事件视觉"],"conference_name":null,"conference_url":null},{"id":26061502,"created_at":"2026-06-26T12:00:09+00:00","title":"Synergizing Global Pattern Learning and Time Order Characterization in Mobile Channel Prediction: An RWKV-Based Approach","date":"2026-06-15","content":"基于RWKV构建移动信道预测网络，利用其交错结构学习多尺度全局信道模式，并借助指数衰减机制刻画时间顺序和因果动态，在低数据、噪声和相关性变化场景下更稳健。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260615-2.png","author":"Zili Wang","link":"https://arxiv.org/abs/2606.16170","category":["序列/强化","信道预测"],"conference_name":null,"conference_url":null},{"id":26061501,"created_at":"2026-06-26T12:00:08+00:00","title":"RWKV-CVM: Gated Cross-Variate Mixing for Multivariate Power Load Forecasting","date":"2026-06-15","content":"基于RWKV-TS加入轻量门控跨变量混合模块CVM，用变量相关性选择性融合天气、区域负荷等多变量信息，在保持线性复杂度的同时提升多变量电力负荷预测表现。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260615-1.png","author":"Adil Rizki","link":"https://www.mdpi.com/2673-4826/7/2/58","category":["序列/强化","电力负荷预测"],"conference_name":null,"conference_url":null},{"id":26061004,"created_at":"2026-06-26T12:00:07+00:00","title":"Random shuffle high-order RWKV for pan-sharpening","date":"2026-06-10","content":"基于视觉RWKV提出高阶随机洗牌全色锐化框架，通过随机扫描与逆洗牌缓解二维依赖建模的方向偏置，并结合WKV共享、高阶通道混合和最大池化提升遥感图像融合质量。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260610-4.png","author":"Man Zhou","link":"https://www.sciencedirect.com/science/article/pii/S1566253526004239","category":["图像","遥感图像融合"],"conference_name":"Information Fusion | 中科院一区 TOP","conference_url":"https://www.sciencedirect.com/science/article/pii/S1566253526004239"},{"id":26061003,"created_at":"2026-06-26T12:00:06+00:00","title":"EGGROLL-IPO: Pluralistic Alignment via Decentralised Post-Training with Population Preferences","date":"2026-06-10","content":"基于RWKV-7G 10.1B作为基础策略，论文在EGGROLL去中心化后训练中引入Online IPO损失，把群体偏好转化为每个工作者可广播的标量适应度，用于更稳定地实现多元偏好对齐。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260610-3.png","author":"Alfie Lamerton","link":"https://openreview.net/forum?id=5P7ihyDNi8","category":["通用","偏好对齐"],"conference_name":null,"conference_url":null},{"id":26061002,"created_at":"2026-06-26T12:00:05+00:00","title":"Degradation-Aware Blind Light-Field Image Quality Assessment With Linear Attention","date":"2026-06-10","content":"基于共享RWKV模块建模光场图像空间与角度分支的长程依赖，并先估计微透镜可靠性图抑制局部退化，使盲光场图像质量评价在复杂退化下更稳健高效。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260610-2.png","author":"Youzhi Zhang","link":"https://ieeexplore.ieee.org/abstract/document/11557154","category":["图像","光场图像质量评价"],"conference_name":"IEEE SPL | JCR Q2","conference_url":"https://ieeexplore.ieee.org/abstract/document/11557154"},{"id":26061001,"created_at":"2026-06-26T12:00:04+00:00","title":"Benchmarking Neural Speech Compression from a Rate-Distortion Perspective","date":"2026-06-10","content":"基于Conv-RWKV Mixture模块构建ECC语音压缩框架，用CNN分支提取局部时频特征，RWKV分支建模长程语音依赖，并结合熵约束优化提升低码率下的率失真表现。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260610-1.png","author":"Jun Xu","link":"https://arxiv.org/abs/2606.11631","category":["语音","语音压缩"],"conference_name":null,"conference_url":null},{"id":26060901,"created_at":"2026-06-26T12:00:03+00:00","title":"Efficient RWKV-based Representation Learning for 3D Point Clouds","date":"2026-06-09","content":"基于RWKV设计P-RWKV模块和PointER框架，将序列建模能力迁移到三维点云表征学习，通过局部感知扩展与空间上下文增强补足几何结构建模，并降低训练和推理开销。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260609-1.png","author":"Yun Liu","link":"https://arxiv.org/abs/2606.10395","category":["3D/视频","三维点云"],"conference_name":null,"conference_url":null},{"id":26060802,"created_at":"2026-06-26T12:00:02+00:00","title":"PathRWKV: Enhancing Whole Slide Image Inference with Asymmetric Recurrent Modeling","date":"2026-06-08","content":"基于RWKV构建PathRWKV病理全切片图像模型，用非对称循环结构实现并行训练与常数内存推理，并结合位置编码和多任务学习提升WSI长序列、多尺度组织区域建模能力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260608-2.png","author":"Tianyi Zhang","link":"https://ieeexplore.ieee.org/abstract/document/11554093/","category":["图像","病理图像分析"],"conference_name":"IEEE TMI | 中科院一区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11554093"},{"id":26060801,"created_at":"2026-06-26T12:00:01+00:00","title":"Fast and effective prediction of spatiotemporal nonlinear dynamics using the tailored RWKV architecture","date":"2026-06-08","content":"基于定制RWKV架构预测时空非线性动力学，将RWKV的线性复杂度序列建模用于多模光纤脉冲传播，捕获长程时空依赖，以较低计算成本实现快速有效的物理过程预测。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260608-1.png","author":"Jinho Lee","link":"https://iopscience.iop.org/article/10.1088/2515-7647/ae79be/meta","category":["序列/强化","非线性动力学预测"],"conference_name":"JPhys Photonics | JCR Q1","conference_url":"https://iopscience.iop.org/article/10.1088/2515-7647/ae79be/meta"},{"id":26052901,"created_at":"2026-05-30T12:00:02+00:00","title":"PLM-NIDS: A Protocol-Language Model for Network Intrusion Detection from Raw Packet Sequences Using RWKV State-Space Models","date":"2026-05-29","content":"提出PLM-NIDS协议语言模型用于无需深度包检测的网络入侵检测，将网络流量视为由L3/L4包元数据（长度、到达间隔、TTL、TCP标志等）构成的语言并使用RWKV-4状态空间模型学习良性流量语法结构，通过逐流困惑度分数在零攻击标签下区分正常与攻击流量达到PR-AUC=0.93，RWKV的O(T)推理复杂度支持逐包流式线速处理且天然兼容TLS1.3/QUIC等加密协议","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260529-1.png","author":"Vivek Kumar Sharma","link":"https://arxiv.org/abs/2606.00155","category":["序列/强化","网络入侵检测"],"conference_name":"TCSVT | 中科院一区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11536060"},{"id":26052701,"created_at":"2026-05-30T12:00:01+00:00","title":"SANet: Structure-Aware Deep Unfolding Network for Face Super-Resolution with Global-Local Modeling","date":"2026-05-27","content":"提出SANet结构感知深度展开网络用于人脸超分辨，将人脸重建任务建模为显式优化问题并迭代展开为可解释深度神经网络，在近端算子中嵌入结构感知RWKV模块利用线性复杂度架构实现高效全局上下文建模，设计结构感知可变形偏移机制根据面部结构动态调整空间聚合模式以保留细粒度面部几何细节，在基准数据集的定量指标和视觉质量上均超越现有SOTA方法","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260527-1.png","author":"Chenyang Wang","link":"https://ieeexplore.ieee.org/abstract/document/11536060","category":["图像","人脸超分辨率"],"conference_name":null,"conference_url":null},{"id":26051801,"created_at":"2026-05-28T12:00:04+00:00","title":"TemLo: Temporal-Local Synergy Enhanced RWKV for Audio-Visual Segmentation","date":"2026-05-18","content":"提出TemLo时序局部协同增强RWKV框架用于音视频分割，引入模态特定RWKV编码器捕获各流内的显著时空动态以降低模态内噪声和歧义，设计声学调制金字塔促进多尺度时空结构化音视觉交互丰富音频感知语义表示，提出视觉引导模态注入器确保时间一致且语义对齐推理，在AVS基准测试中建立新SOTA且使用PVT-v2骨干时在MS3数据集上MJ和MF分别提升10.28%和3.66%","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260518-1.png","author":"Yunzhi Zhuge","link":"https://ieeexplore.ieee.org/abstract/document/11523565","category":["通用","音视频分割"],"conference_name":"IEEE TMM | JCR Q1","conference_url":"https://ieeexplore.ieee.org/abstract/document/11523565"},{"id":26051501,"created_at":"2026-05-26T12:00:04+00:00","title":"FRWKV+: Adaptive Periodic-Position Branch Interaction for Frequency-Space Linear Time Series Forecasting","date":"2026-05-15","content":"提出FRWKV+增强型频域时序预测模型，基于RWKV架构引入跨分支门控交换实部虚部频率流上下文，采用自适应相位门机制提供信任控制的有符号周期位置校正，在匹配种子评估中实现FRWKV家族最大MSE胜出覆盖率","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260515-1.png","author":"Qingyuan Yang","link":"https://arxiv.org/abs/2605.15690","category":["序列/强化","时序预测"],"conference_name":null,"conference_url":null},{"id":26051402,"created_at":"2026-05-28T12:00:03+00:00","title":"C2F-VRWKV: A Lightweight Clustering-Aware and Cross-Frequency-Enhanced Vision-RWKV Network for Ship Detection in Remote Sensing Scenarios","date":"2026-05-14","content":"提出C2F-VRWKV轻量级聚类感知跨频率增强Vision-RWKV网络用于遥感图像船舶检测，C2-VRWKV骨干集成上下文聚类机制在线性复杂度扫描范式中聚合局部细粒度形态特征，跨频率融合模块采用频率引导交叉注意力策略解耦并强调船舶轮廓纹理同时抑制环境噪声，双流多频率注意力模块利用先验引导机制分离前景船舶特征与背景杂波，仅0.53M参数在Airbus-Ship数据集达到78.23% mAP","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260514-2.png","author":"Shounan Wang","link":"https://ieeexplore.ieee.org/abstract/document/11520379","category":["图像","遥感船舶检测"],"conference_name":"IEEE TGRS | 中科院 1 区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11520379"},{"id":26051401,"created_at":"2026-05-26T12:00:03+00:00","title":"SCRWKV: Ultra-Compact Structure-Calibrated Vision-RWKV for Topological Crack Segmentation","date":"2026-05-14","content":"提出SCRWKV超紧凑结构校准Vision-RWKV裂缝分割网络，Structure-Field Encoder骨干集成自适应多尺度级联调制器增强纹理表示，结构校准洞察单元通过几何引导双向结构变换捕获拓扑相关性，仅1.22M参数在TUT数据集上达到F1=0.8428","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260514-1.png","author":"Hanxu Zhang","link":"https://arxiv.org/abs/2605.14926","category":["图像","裂缝分割"],"conference_name":"ICML 2026 | CCF A","conference_url":"https://arxiv.org/abs/2605.14926"},{"id":26050701,"created_at":"2026-05-26T12:00:02+00:00","title":"Wavelet-Decoupled Spatiotemporal Network for Stock Return Prediction","date":"2026-05-07","content":"提出WaveDSTN小波解耦时空网络用于股票收益预测，利用小波变换将股票收益分解为高频短期波动和低频长期趋势分量，设计双路径时空编码器捕获动态时序依赖和跨股票信息传播，在保持时间序列因果一致性的同时显著提升预测准确率","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260507-1.png","author":"Lei Liao","link":"https://www.mdpi.com/1099-4300/28/5/548","category":["序列/强化","金融预测"],"conference_name":null,"conference_url":null},{"id":26042901,"created_at":"2026-05-26T12:00:01+00:00","title":"Rapid and High-Accuracy Three-Dimensional Airborne Transient Electromagnetic Forward Modeling Based on Machine Learning","date":"2026-04-29","content":"提出基于RWKV的3D航空瞬变电磁正演建模深度学习方法，采用双向加权键值(Bi-WKV)机制以线性复杂度捕获三维地电结构几何关系与连续性，引入收发高度调制机制适应飞行高度变化，在包含断层和褶皱的大规模数据集上实现1秒内高精度正演计算","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260429-1.png","author":"Xuben Wang","link":"https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2025JH001181","category":["3D/视频","地球物理"],"conference_name":null,"conference_url":null},{"id":26042401,"created_at":"2026-04-30T12:00:07+00:00","title":"Spec-RWKV: A Spectrum-Guided Multi-Scale Recurrent Modeling Framework for Multi-Center Resting-State fMRI-Assisted Diagnosis","date":"2026-04-24","content":"提出Spec-RWKV频谱引导的多尺度循环建模框架用于多中心静息态fMRI辅助诊断，基于RWKV架构显式建模物理采样间隔并联合时频信息，在ABIDE-I和ADHD-200数据集上展现竞争力且对TR扰动具有鲁棒性","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260424-1.png","author":"Sihang Peng","link":"https://www.mdpi.com/2076-3425/16/5/455","category":["图像","医学影像分析"],"conference_name":null,"conference_url":null},{"id":26042102,"created_at":"2026-04-30T12:00:07+00:00","title":"HA-ViTNet: Dual-Domain Collaborative Learning for Semantic Segmentation of High-Resolution Remote Sensing Images","date":"2026-04-21","content":"提出HA-ViTNet双域协作框架用于高分辨率遥感图像分割，设计Spatial RWKV Attention Block模拟空间序列内的时序依赖关系实现伪时序建模，将全局上下文反馈到空间流以平衡局部细节保留与全局语义理解","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260421-2.png","author":"Yu Bai","link":"https://ieeexplore.ieee.org/abstract/document/11461799","category":["图像","遥感分割"],"conference_name":"ICASSP 2026 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/11461799"},{"id":26042101,"created_at":"2026-04-30T12:00:07+00:00","title":"Enhance Deformation-Tolerant Unsupervised Infrared and Visible Image Fusion Via Hybrid Feature Representation Learning","date":"2026-04-21","content":"提出变形容忍红外与可见光图像融合方法，设计动态混合编码器结合CNN自适应权重特征与Vision-RWKV变体模型实现从局部特征感知到全局上下文建模的特征抽象，通过隐式特征级对齐解码器实现高精度多尺度特征对齐","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260421-1.png","author":"Ruhao Yan","link":"https://ieeexplore.ieee.org/abstract/document/11463142","category":["图像","图像融合"],"conference_name":"ICASSP 2026 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/11463142"},{"id":26042002,"created_at":"2026-05-28T12:00:02+00:00","title":"DCAF: Dynamic Affective Consistency-Aware Fusion with Disentangled Modality Representations for Multimodal Sentiment Analysis","date":"2026-04-20","content":"提出DCAF动态情感一致性感知融合框架处理多模态情感分析中的模态差距和情感冲突问题，采用跨模态正交解耦学习(CODL)通过三模态交叉注意力机制和监督对比目标桥接模态差距，设计一致性引导单模态标签推导(CULD)在双平面几何约束下缓解样本内情感冲突，利用RWKV实现线性效率并在三个基准测试达到SOTA结果","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260420-2.png","author":"Weihao Lv","link":"https://www.sciencedirect.com/science/article/abs/pii/S0925231226010520","category":["语言","多模态情感分析"],"conference_name":"Neurocomputing | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0925231226010520"},{"id":26042001,"created_at":"2026-04-30T12:00:04+00:00","title":"Trifusion-RWKV for Complex Degradation Restoration in Library and Archive Environments","date":"2026-04-20","content":"提出TriFusion-RWKV图书馆档案图像恢复模型，核心Fusion-RWKV模块包含频率自适应、多尺度膨胀注意力和动态LUT三个并行分支，通过RWKV风格门控和前馈网络精炼特征，在LOL-v2-real数据集上达到23.99dB PSNR且仅9.17M参数","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260420-1.png","author":"Yong Cui","link":"https://fyust.edu.cn/gjhyqk/cscwd2026/papers/paper_589.pdf","category":["图像","图像恢复"],"conference_name":null,"conference_url":null},{"id":26041901,"created_at":"2026-04-30T12:00:04+00:00","title":"PestVL-Net: Enabling Multimodal Pest Learning via Fine-grained Vision-Language Interaction","date":"2026-04-19","content":"提出PestVL-Net视觉语言框架用于害虫识别，在视觉通路中采用RWKV架构并嵌入显著性引导的自适应窗口分区方案建模害虫细粒度视觉特征，结合MLLM先验和多模态思维链推理生成精确语义描述","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260419-1.png","author":"Xueheng Li","link":"https://arxiv.org/abs/2604.17278","category":["图像","视觉语言模型"],"conference_name":null,"conference_url":null},{"id":26041801,"created_at":"2026-04-30T12:00:04+00:00","title":"RWKV4Rec: RWKV-Based Personalized Sequential Recommendation Model","date":"2026-04-18","content":"提出RWKV4Rec序列推荐模型，将RWKV的高效长序列处理能力应用于推荐系统，设计item-RWKV块模块并基于LoRA技术提出Low-Rank TimeMix自适应分配历史物品权重，在四个基准数据集上NDCG@10提升1.80-3.76%","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260418-1.png","author":"Mengwei Yuan","link":"https://dl.acm.org/doi/abs/10.1145/3810245","category":["序列/强化","推荐系统"],"conference_name":"ACM TKDD","conference_url":"https://dl.acm.org/doi/abs/10.1145/3810245"},{"id":26041601,"created_at":"2026-04-30T12:00:04+00:00","title":"Multigrain-aware Semantic Prototype Scanning and Tri-Token Prompt Learning Embraced High-Order RWKV for Pan-Sharpening","date":"2026-04-16","content":"提出多粒度语义原型扫描全色锐化方法，在高阶RWKV架构上嵌入语义驱动的扫描策略，使用局部敏感哈希构建多粒度语义原型实现上下文感知token重排序，结合三token提示学习机制提升全局交互一致性","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260416-1.png","author":"Junfeng Li","link":"https://arxiv.org/abs/2604.14622","category":["图像","全色锐化"],"conference_name":null,"conference_url":null},{"id":26041301,"created_at":"2026-05-28T12:00:01+00:00","title":"PHMRNet: Persistent Homology Based Mamba-RWKV Network for LiDAR Place Recognition","date":"2026-04-13","content":"提出PHMRNet持久同调Mamba-RWKV网络用于LiDAR场景识别，将持久同调提取的拓扑信息作为拓扑感知采样通道补偿2D范围视图投影丢失的几何细节，引入时空融合模块链接跨帧拓扑与范围视图特征提升时空一致性，在公开数据集上显著提升描述符鲁棒性","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260413-1.png","author":"Dejing Zhou","link":"https://ieeexplore.ieee.org/abstract/document/11480752","category":["3D/视频","LiDAR定位"],"conference_name":"IEEE RA-L | JCR Q1","conference_url":"https://ieeexplore.ieee.org/abstract/document/11480752"},{"id":26040601,"created_at":"2026-04-14T12:00:00+00:00","title":"RICEFuse: A RWKV-based information complementarity enhancement network for infrared-visible image fusion","date":"2026-04-06","content":"提出鲁棒红外与彩色图像融合框架RICEFuse，将RWKV作为序列特征提取模块嵌入图像特征解码分支，增强长距离特征依赖建模能力，提升图像融合质量和鲁棒性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260406-1.png","author":"Unknown","link":"https://www.sciencedirect.com/science/article/abs/pii/S1350449526001945","category":["图像","图像融合"],"conference_name":"Infrared Physics & Technology","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S1350449526001945"},{"id":26040301,"created_at":"2026-04-14T12:00:00+00:00","title":"Learning the Signature of Memorization in Autoregressive Language Models","date":"2026-04-03","content":"基于RWKV-4线性注意力模型验证可迁移的成员推理攻击方法LT-MIA，取得0.972 AUC性能，证明通过梯度下降优化交叉熵损失的语言模型存在通用的记忆特征签名。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260403-1.png","author":"David Ilić","link":"https://arxiv.org/abs/2604.03199","category":["语言","模型安全"],"conference_name":null,"conference_url":null},{"id":26032701,"created_at":"2026-03-30T12:00:04+00:00","title":"MD-RWKV-UNet: Scale-Aware Anatomical Encoding with Cross-Stage Fusion for Multi-Organ Segmentation","date":"2026-03-27","content":"提出MD-RWKV-UNet多器官分割模型，使用MD-RWKV双路径模块整合可变形空间偏移与RWKV机制，结合跨阶段双注意力融合策略，在Synapse和ACDC数据集上小器官分割精度提升9.2%","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260327-1.png","author":"Zhuoyi Fang","link":"https://arxiv.org/abs/2603.27261","category":["图像","医学图像分割"],"conference_name":null,"conference_url":null},{"id":26032603,"created_at":"2026-04-14T12:00:00+00:00","title":"A novel TV-FEM-RWKV-TS hybrid framework for aging prediction of PEMFCs under static and quasi-dynamic conditions","date":"2026-03-26","content":"提出TV-FEM-RWKV-TS时序预测模型，将RWKV模块嵌入时间序列预测主干网络，结合有限元方法分解时间特征，提升长期时序预测精度，降低计算复杂度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260326-3.png","author":"Unknown","link":"https://www.sciencedirect.com/science/article/abs/pii/S0360319926012255","category":["序列/强化","时序预测"],"conference_name":"IJHE | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0360319926012255"},{"id":26032602,"created_at":"2026-04-14T12:00:00+00:00","title":"MMCGR: A multimodal cascaded gated fusion RWKV-based model for emotion recognition in conversations","date":"2026-03-26","content":"提出多模态对比图推理框架MMCGR，将RWKV作为序列编码器嵌入多模态特征处理分支，增强跨模态时序关联建模能力，提升多模态任务的推理精度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260326-2.png","author":"Unknown","link":"https://www.sciencedirect.com/science/article/abs/pii/S0925231226008210","category":["图像","多模态"],"conference_name":"Neurocomputing | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0925231226008210"},{"id":26032601,"created_at":"2026-03-30T12:00:04+00:00","title":"Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework","date":"2026-03-26","content":"提出CLE-RWKV低光照增强框架，将RWKV状态空间模型用于稠密预测任务，通过空间到深度策略折叠邻域信息到通道维度，在7个基准数据集上PSNR提升2.1dB，实现可控制的亮度调节","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260326-1.png","author":"Zhuohua Ye","link":"https://arxiv.org/abs/2603.25296","category":["图像","低光照增强"],"conference_name":null,"conference_url":null},{"id":26032501,"created_at":"2026-03-30T12:00:04+00:00","title":"SCEAF-UNet: Medical image segmentation based on spatial-channel feature enhancement and adaptive fusion","date":"2026-03-25","content":"在RWKV-UNet医学图像分割骨干网络解码器中嵌入SCEAF模块，并行使用多尺度空间注意力门控块与通道注意力调制块，结合边缘注意力融合模块，在Synapse和ACDC数据集上分割精度提升6.3%","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260325-1.png","author":"Lingyun Zhao","link":"https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0345538","category":["图像","医学图像分割"],"conference_name":"PLOS ONE | JCR Q2","conference_url":"https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0345538"},{"id":26032401,"created_at":"2026-03-30T12:00:04+00:00","title":"StateLinFormer: Stateful Training Enhancing Long-term Memory in Navigation","date":"2026-03-24","content":"将RWKV线性注意力机制嵌入StateLinFormer导航模型，替换传统Transformer的注意力模块，配合有状态训练范式保留跨批次记忆状态，在MAZE和ProcTHOR环境中导航性能提升28%，支持无限长序列推理","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260324-1.png","author":"Zhiyuan Chen","link":"https://arxiv.org/abs/2603.23571","category":["3D/视频","导航"],"conference_name":null,"conference_url":null},{"id":26031901,"created_at":"2026-03-30T12:00:04+00:00","title":"Real-Frequency Correlation Functions from Neural Quantum States via Operator Lanczos Approach","date":"2026-03-19","content":"基于RWKV架构的神经量子态模型求解实频关联函数，嵌入RWKV时序模块处理算子演化过程，结合Operator Lanczos方法大幅提升计算精度与效率，在多体物理系统模拟中降低70%内存开销","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260319-1.png","author":"Research Square Team","link":"https://www.researchsquare.com/article/rs-8976411/v1","category":["通用","量子计算"],"conference_name":null,"conference_url":null},{"id":26031401,"created_at":"2026-03-23T15:00:02+00:00","title":"Beyond Quadratic: Linear-Time Change Detection with RWKV","date":"2026-03-14","content":"基于RWKV架构构建ChangeRWKV变化检测模型，使用分层RWKV编码器提取多分辨率特征，结合时空融合模块解决尺度空间错位问题，在LEVIR-CD基准上达到85.46% IoU，相比现有方法大幅降低参数量和计算量。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260314-1.png","author":"Zhenyu Yang","link":"https://ojs.aaai.org/index.php/AAAI/article/view/38167","category":["图像","变化检测"],"conference_name":"AAAI 2026 | CCF A","conference_url":"https://ojs.aaai.org/index.php/AAAI/article/view/38167"},{"id":26030901,"created_at":"2026-03-20T14:48:49.810762+00:00","title":"PoseRWGCN: an attention-free dual-stream RWKV–GCN architecture for real-time 3D human pose estimation","date":"2026-03-09","content":"在PoseRWGCN的双流架构中，RWKV流作为全局时间特征建模分支，通过递归门控机制捕捉长程时间依赖，与GCN流的局部空间特征经自适应融合，实现实时3D人体姿态估计，FLOPs减少62.1%且精度提升0.6mm。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260309-1.png","author":"Lintao Song","link":"https://link.springer.com/article/10.1007/s40747-026-02239-x","category":["3D/视频","姿态估计"],"conference_name":"JCR Q2 | Complex & Intelligent Systems","conference_url":"https://link.springer.com/article/10.1007/s40747-026-02239-x"},{"id":26030402,"created_at":"2026-03-23T15:00:01+00:00","title":"Why Are Linear RNNs More Parallelizable?","date":"2026-03-04","content":"从计算复杂度理论角度揭示线性RNN（LRNN）可并行化的本质，证明LRNN等价于对数深度算术电路，而非线性RNN存在L完全问题的并行化瓶颈，明确不同LRNN变体的表达能力差异。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260304-2.png","author":"William Merrill","link":"https://arxiv.org/abs/2603.03612","category":["通用","循环神经网络底层原理"],"conference_name":null,"conference_url":null},{"id":26030401,"created_at":"2026-03-20T14:48:06.5792+00:00","title":"A multi-prior fusion phase unwrapping method based on RwkvU-Net in digital holography","date":"2026-03-04","content":"在RwkvU-Net的Backbone网络中堆叠Vision-RWKV模块作为编码器，结合CNN的局部特征提取能力，建模全局上下文信息，提升复杂噪声下的相位展开精度和鲁棒性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260304-1.png","author":"Haoming Huang","link":"https://www.sciencedirect.com/science/article/abs/pii/S0143816626001119","category":["图像","相位展开"],"conference_name":"中科院 2 区 | Optics and Lasers in Engineering ","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0143816626001119"},{"id":26022701,"created_at":"2026-03-09T00:00:00+00:00","title":"RWKV-Inspired Multi-modal Relation Modeling for Vision-Language Tracking","date":"2026-02-27","content":"针对现有多模态交互方法缺乏对语言序列上下文与视觉特征关系的有效建模且计算时间成本高的问题，在RrmTrack框架中引入基于RWKV的多模态交互方法：借鉴RWKV架构设计时间混合模块来建模语言序列信息与图像特征之间的关系，借鉴RWKV架构设计通道混合模块来促进图像间的信息交互，利用RWKV的并行化训练、线性注意力机制和高效RNN推理的特性，在TNL2k、LaSOT等多个视觉语言跟踪基准上取得SOTA结果和速度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260227-1.png","author":"Guangtong Zhang","link":"https://ieeexplore.ieee.org/abstract/document/11417442","category":["图像","目标跟踪"],"conference_name":"IEEE TMM | 中科院 1 区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11417442"},{"id":26022501,"created_at":"2026-03-23T15:00:00+00:00","title":"OmniZip: Learning a Unified and Lightweight Lossless Compressor for Multi-Modal Data","date":"2026-02-25","content":"提出统一多模态无损压缩器OmniZip，包含模态统一分词器、模态路由上下文学习机制和模态路由前馈设计，在图像、文本、语音等多模态数据集上压缩效率相比gzip最高提升62%，支持边缘设备近实时推理。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260225-1.png","author":"Yan Zhao","link":"https://arxiv.org/abs/2602.22286","category":["通用","压缩算法"],"conference_name":null,"conference_url":null},{"id":26022401,"created_at":"2026-02-28T00:00:00+00:00","title":"A high-performance defect detection for titanium strip via receptance weighted key value architecture-inspired context modeling and hierarchical differential fusion","date":"2026-02-24","content":"在C2-RWKV模块中重构RWKV的时间和通道混合单元为二维空间通道混合单元，通过深度卷积将一维递归WKV算子映射到全向空间邻域混合，以线性复杂度实现图像特征的选择性动态累积，增强模型对多尺度和高相似度缺陷的判别能力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260224-1.png","author":"He Zeng","link":"https://www.sciencedirect.com/science/article/abs/pii/S0952197626003799","category":["图像","缺陷检测"],"conference_name":"EAAI | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0952197626003799"},{"id":26022301,"created_at":"2026-02-25T00:00:00+00:00","title":"Pre-trained multi-scale RWKV-GCN for multivariate time series forecasting","date":"2026-02-23","content":"构建两阶段框架PMSRWKV-GCN，第一阶段用FFT预处理后通过RWKV在自监督预训练中学干净的时序表示，第二阶段用GCN利用空间结构，提升多元时间序列预测性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260223-1.png","author":"Jianhua Hao","link":"https://link.springer.com/article/10.1038/s41598-026-41091-4","category":["序列/强化","时间序列预测"],"conference_name":"Scientific Reports","conference_url":"https://link.springer.com/article/10.1038/s41598-026-41091-4"},{"id":26021801,"created_at":"2026-02-25T00:00:00+00:00","title":"Serialized PointRWKV: A Serialized RWKV-like Model Employing Feature-Based Masked Autoencoders for Point Cloud Analysis","date":"2026-02-18","content":"设计旋转不变重排序算法将无序3D点云转换为结构化序列，修改RWKV架构实现邻域插值和全局注意力，引入基于特征空间的MAE预训练，在ScanObjectNN和ModelNet40上达95.13%和94.6%准确率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260218-1.png","author":"Zhifeng Rao","link":"https://ieeexplore.ieee.org/document/11399627","category":["3D/视频","点云分析"],"conference_name":"IEEE Sensors Journal | JCR Q1","conference_url":"https://ieeexplore.ieee.org/document/11399627"},{"id":26021501,"created_at":"2026-02-25T00:00:00+00:00","title":"scMix: Learning Temporal Dynamics of Gene Expression under Irregular Time Intervals","date":"2026-02-15","content":"扩展RWKV架构提出Delta-RWKV块，通过Delta-Time Mixing将时间间隔Δt融入衰减项，处理不规则时间间隔的单细胞测序数据，建模基因表达的时间动态变化。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260215-1.png","author":"Shangjin Han","link":"https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btag080/8487133","category":["序列/强化","单细胞分析"],"conference_name":"Bioinformatics | JCR Q1","conference_url":"https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btag080/8487133"},{"id":26020601,"created_at":"2026-02-25T00:00:00+00:00","title":"MMSegRWKV: Enhancing Multimodal MRI Segmentation for Internet of Medical Things-Enabled Healthcare with RWKV-Inspired Architectures","date":"2026-02-06","content":"结合RWKV与U形架构构建MMSegRWKV，用DV-WKV在动态短时间窗口内建模双向时空依赖，通过ResFM显式建模二阶跨模态交互，提升多模态MRI分割精度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260206-1.png","author":"Yitong Cao","link":"https://ieeexplore.ieee.org/abstract/document/11373528","category":["3D/视频","医学图像分割"],"conference_name":"IEEE IoTJ | 中科院 1 区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11373528"},{"id":26020401,"created_at":"2026-02-25T00:00:00+00:00","title":"Surv-RWKV: Cross-modal receptance weighted key-value interaction with optimal transport feature alignment for survival analysis","date":"2026-02-04","content":"用RWKV编码器从WSI和基因通路序列提取特征，引入最优传输特征对齐模块映射到共享空间，通过RSF和CRDI模块动态建模跨模态交互，提升生存预测准确性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260204-1.png","author":"Xiyang Kuang","link":"https://www.sciencedirect.com/science/article/abs/pii/S0957417426004197","category":["序列/强化","生存分析"],"conference_name":"ESWA | 中科院 1 区 TOP","conference_url":"https://www.sciencedirect.com/journal/expert-systems-with-applications"},{"id":26020102,"created_at":"2026-04-14T12:00:00+00:00","title":"Asymptotic Semantic Collapse in Hierarchical Optimization","date":"2026-02-01","content":"基于RWKV-7 13B GGUF checkpoint进行无数据集基准测试，报告零哈希冲突，贪心解码下平均合规度0.50，随机解码下0.531，最终与锚点的Jaccard相似度分别为0.295和0.224。研究多代理语言系统中的渐近语义坍缩现象，建模语义状态为黎曼流形上的点，分析投影动力学，证明路径独立性和熵消失特性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260201-2.png","author":null,"link":"https://arxiv.org/abs/2602.18450","category":["语言","语义分析"],"conference_name":null,"conference_url":null},{"id":26020101,"created_at":"2026-02-05T18:16:15.80033+00:00","title":"KAN We Flow? Advancing Robotic Manipulation with 3D Flow Matching via KAN & RWKV","date":"2026-02-01","content":"替换流匹配的UNet风格骨干为RWKV-KAN UNet，通过RWKV时间/通道混合传播任务上下文及GroupKAN样条功能校准，实现参数减少86.8%、保持快速推理并达到最优成功率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260201-1.png","author":"Zhihao Chen","link":"https://arxiv.org/abs/2602.01115","category":["3D/视频","流匹配政策"],"conference_name":"ICRA 2026 | CCF B","conference_url":"https://arxiv.org/abs/2602.01115"},{"id":26012901,"created_at":"2026-02-25T00:00:00+00:00","title":"SeisRWKV: Multi-scale Feature Interaction with Linear Complexity for Seismic Neighboring-shot Interference Mitigation","date":"2026-01-29","content":"用RWKV层结合co-wkv双向注意力机制以线性复杂度全局建模，集成MCF模块增强跨通道信息融合与多尺度特征交互，精准抑制地震邻炮干扰。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260129-1.png","author":"Yun Tang","link":"https://www.researchsquare.com/article/rs-8709232/v1","category":["序列/强化","地震数据处理"],"conference_name":null,"conference_url":null},{"id":26012301,"created_at":"2026-02-05T18:15:16.380275+00:00","title":"U-RWKV: Accurate and Efficient Volumetric Medical Image Segmentation via RWKV","date":"2026-01-23","content":"设计Tri-directional Spatial Enhancement RWKV (TSE-R) block，通过RWKV进行全局建模并结合空间偏移策略与三方向扫描机制，提升体素医学图像分割准确性（Dice达87.21%）且参数减少16.08倍。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260123-1.png","author":"Hongyu Cai","link":"https://ieeexplore.ieee.org/abstract/document/11360601","category":["3D/视频","体素分割"],"conference_name":"IEEE TIP | CCF A","conference_url":"https://ieeexplore.ieee.org/abstract/document/11360601"},{"id":26011701,"created_at":"2026-01-28T15:44:48.3253+00:00","title":"A Method for Detecting Spatio-temporal Correlation Anomalies of WSN Nodes Based on Topological Information Enhancement and Time-frequency Feature Extraction","date":"2026-01-17","content":"论文在时间域和频率域特征提取分支中，嵌入基于RWKV的CFE模块，以并行训练处理长距离依赖任务并降低计算复杂度，充分提取不同时间模态间的时序相关性特征。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260117-1.png","author":"Miao Ye","link":"https://arxiv.org/abs/2601.11951","category":["序列/强化","异常检测"],"conference_name":null,"conference_url":null},{"id":26011401,"created_at":"2026-02-05T18:13:45.954444+00:00","title":"ROSA-Tuning: Enhancing Long-Context Modeling via Suffix Matching","date":"2026-01-14","content":"将 RWKV-8 ROSA 与注意力机制并行部署在CPU端，识别长上下文中与当前查询相关的历史位置并将检索信息以可训练方式注入模型状态，显著恢复窗口注意力模型的长上下文建模能力且性能接近全局注意力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260114-1.png","author":"Yunao Zheng","link":"https://arxiv.org/abs/2602.02499","category":["通用","ROSA 训推"],"conference_name":null,"conference_url":null},{"id":26011201,"created_at":"2026-01-20T18:40:10.662002+00:00","title":"Hi-RWKV: Hierarchical RWKV Modeling for Hyperspectral Image Classification","date":"2026-01-12","content":"论文在分层编码器的每个阶段嵌入 Hi-RWKV 块作为核心处理模块，通过双向空间传播和边缘感知门控实现全局空间上下文建模，并引入可学习的波段嵌入和通道混合增强跨波段判别性，以线性复杂度实现高效的像素级分类。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260112-1.png","author":"Yunbiao Wang","link":"https://ieeexplore.ieee.org/document/11333959","category":["图像","高光谱图像分类"],"conference_name":"IEEE TIP | 中科院 1 区 TOP","conference_url":"https://ieeexplore.ieee.org/document/11333959"},{"id":26011001,"created_at":"2026-01-20T18:40:02.978102+00:00","title":"EmbeddingRWKV: State-Centric Retrieval with Reusable States","date":"2026-01-10","content":"论文使用 RWKV 作为检索和重排的统一骨干网络，通过状态表示学习生成可复用状态，实现离线状态缓存。在重排阶段仅处理查询令牌，将推理成本与文档长度解耦，获得 5.4×–44.8× 加速。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260110-1.png","author":"侯皓文","link":"https://arxiv.org/abs/2601.07861","category":["语言","检索增强生成"],"conference_name":null,"conference_url":null},{"id":26010901,"created_at":"2026-01-20T18:39:56.547725+00:00","title":"DyRSRNet: A Lightweight Super-Resolution Framework Based on Dynamic Recursive State-Space Networks","date":"2026-01-09","content":"在 VRSE 模块中嵌入 DyRWKV 机制，通过双向递归加权键值建模增强跨方向空间依赖性和通道交互，提升纹理细节恢复。在 RCSS 模块中堆叠方向感知状态路径，结合选择性扫描和动态调制实现长程上下文表示和结构恢复。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260109-1.png","author":"Sijia He","link":"https://link.springer.com/chapter/10.1007/978-981-95-5702-8_4","category":["图像","超分辨率"],"conference_name":null,"conference_url":null},{"id":26010702,"created_at":"2026-01-28T15:44:26.603458+00:00","title":"Graph fusion model for unimodal and multimodal fake news detection","date":"2026-01-07","content":"论文在单模态检测模型CMGN的文本编码模块中嵌入RWKV MLP-mixer，替换传统MLP以捕获新闻文本的长程依赖关系，生成高维向量用于后续特征融合，提升检测准确率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260107-2.png","author":"CUI SHAODONG","link":"https://rose-ibadai.repo.nii.ac.jp/records/2001365","category":["语言","假新闻检测"],"conference_name":"茨城大学博士学位论文","conference_url":"https://rose-ibadai.repo.nii.ac.jp/records/2001365"},{"id":26010701,"created_at":"2026-01-12T17:47:30.852349+00:00","title":"MFC-RFNet: A Multi-scale Guided Rectified Flow Network for Radar Sequence Prediction⋆","date":"2026-01-07","content":"该研究在 MFC-RFNet 中引入轻量级 Vision-RWKV 模块，用于高效捕获雷达序列的长程时空依赖。模型结合了多尺度特征通信、空间对齐和小波引导融合，并通过 Rectified Flow 训练实现快速采样。实验表明，该方法在多个降水临近预报数据集上超越了基线模型，尤其在高降雨率预测上表现更优。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260107-1.png","author":"Wenjie Luo","link":"https://arxiv.org/abs/2601.03633","category":["序列/强化","降水临近预报"],"conference_name":null,"conference_url":null},{"id":26010601,"created_at":"2026-01-28T15:44:40.721131+00:00","title":"Natural Cognizing Video: A Decoupling and Integration Network for General Event Boundary Captioning","date":"2026-01-06","content":"论文在事件分支的编码器中嵌入 RWKV 层，用于处理视频帧序列以建模动态变化，结合 Transformer 的并行训练和 RNN 的高效推理优势，提升状态前后描述的生成准确性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260106-1.png","author":"Yutao Liu","link":"https://ieeexplore.ieee.org/abstract/document/11329489","category":["3D/视频","视频描述"],"conference_name":"IEEE TMM | 中科院 1 区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11329489"},{"id":26010502,"created_at":"2026-01-12T17:46:11.899232+00:00","title":"HFRWKV: A High-Performance Fully On-Chip Hardware Accelerator for RWKV","date":"2026-01-05","content":"针对 RWKV 模型在 GPU 上并行性差与内存瓶颈的问题，本文提出了 HFRWKV 硬件加速器。该设计通过混合精度量化策略与全芯片计算架构，显著提升了模型的推理吞吐量与能效。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260105-2.png","author":"Shijie Liu","link":"https://arxiv.org/abs/2601.02135","category":["通用","硬件加速"],"conference_name":"FPGA 2026 | CCF B ","conference_url":"https://dl.acm.org/doi/abs/10.1145/3748173.3779553"},{"id":26010501,"created_at":"2026-01-12T17:45:25.457941+00:00","title":"Exploring Linear Attention in Underwater Image Enhancement with Retinex Theory","date":"2026-01-05","content":"本文提出 Retinex-RWKV 框架，将改进的 RWKV 线性注意力机制与 Retinex 理论结合，用于水下图像增强。该模型通过 SS2D-WKV 和 Omni-token shift 等设计，有效提升了特征提取能力。实验表明，该方法在定量和定性评估上均优于现有方法，能生成更高质量的图像。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20260105-1.png","author":"Long Yao","link":"https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6023431","category":["图像","图像增强"],"conference_name":null,"conference_url":null},{"id":25123101,"created_at":"2026-01-05T11:50:34.981178+00:00","title":"IESS-FusionNet: Physiologically Inspired EEG-EMG Fusion with Linear Recurrent Attention for Infantile Epileptic Spasms Syndrome Detection","date":"2025-12-31","content":"该研究基于 RWKV 的线性循环注意力，提出了 Cross Time-Mixing 模块，用于 EEG-EMG 信号的高效跨模态融合。IESS-FusionNet 框架结合了单模态编码器，实现了对婴儿癫痫性痉挛综合征的精准检测，性能优于基线且计算高效。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251231-1.png","author":"Junyuan Feng","link":"https://www.mdpi.com/2306-5354/13/1/57","category":["序列/强化","癫痫检测"],"conference_name":"Bioengineering","conference_url":"https://www.mdpi.com/journal/bioengineering"},{"id":25122401,"created_at":"2025-12-30T17:56:54.384093+00:00","title":"RWKV-SKF: A recurrent architecture with state-space and frequency-domain filtering for dissolved oxygen predicting and revealing influencing mechanisms","date":"2025-12-24","content":"该研究基于 RWKV 架构，提出了 RWKV-SKF 框架，通过融合状态空间卡尔曼滤波与频域傅里叶滤波，有效处理传感器噪声与周期性动态，以提升溶解氧预测精度。实验证明该模型性能优越，并能揭示关键影响机制。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251224-1.png","author":"Peijian Zeng","link":"https://www.sciencedirect.com/science/article/abs/pii/S0020025525011557","category":["序列/强化","时间序列预测"],"conference_name":"Information Sciences | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0020025525011557"},{"id":25122101,"created_at":"2025-12-30T17:58:48.874694+00:00","title":"DME-RWKV: An Interpretable Multimodal Deep Learning Framework for Predicting Anti-VEGF Response in Diabetic Macular Edema","date":"2025-12-21","content":"本研究基于 RWKV 架构提出了 DME-RWKV 模型，用于预测糖尿病黄斑水肿（DME）患者对抗 VEGF 治疗的反应。该模型融合了 OCT 和超广field成像，并结合因果注意力学习等方法，在生物标志物分割和治疗反应预测任务上均表现出色，具有高精度和可解释性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251221-1.png","author":"Yan Liu","link":"https://www.mdpi.com/2306-5354/13/1/12","category":["图像","医学图像分析"],"conference_name":"Bioengineering","conference_url":"https://www.mdpi.com/2306-5354/13/1/12"},{"id":25121701,"created_at":"2025-12-25T17:12:07.130359+00:00","title":"LADY: L inear A ttention for Autonomous D riving Efficiency without Transformers","date":"2025-12-17","content":"该研究基于 RWKV-7 模型提出首个完全线性注意力的端到端自动驾驶框架 LADY。它通过轻量级线性交叉注意力机制，高效融合多帧传感器数据，实现恒定计算与内存开销。实验证明 LADY 在提升规划性能的同时显著降低计算成本，并已在边缘设备部署验证。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251217-1.png","author":"李致远","link":"https://arxiv.org/abs/2512.15038","category":["3D/视频","端到端自动驾驶"],"conference_name":null,"conference_url":null},{"id":25121101,"created_at":"2025-12-25T17:12:03.380173+00:00","title":"SemanticBBV: A Semantic Signature for Cross-Program Knowledge Reuse in Microarchitecture Simulation","date":"2025-12-11","content":"该论文提出 SemanticBBV，使用轻量级 RWKV 编码器生成语义签名，以解决传统 BBV 无法跨程序重用知识的问题。它通过 Set Transformer 聚合嵌入并联合训练，实现了跨程序性能估计，显著加速了微架构模拟。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251211-1.png","author":"Zhenguo Liu","link":"https://arxiv.org/abs/2512.10231","category":["序列/强化","程序分析"],"conference_name":" ASP-DAC 2026","conference_url":"https://arxiv.org/abs/2512.10231"},{"id":25120902,"created_at":"2025-12-25T17:16:45.767279+00:00","title":"Fourier-RWKV: A Multi-State Perception Network for Efficient Image Dehazing","date":"2025-12-09","content":"Fourier-RWKV 通过将 WKV 注意力机制扩展至频域来解决图像去雾问题。该模型融合了空间、频域与语义三种感知状态，以线性复杂度实现了 SOTA 性能，有效平衡了恢复质量与计算效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251209-2.png","author":"Lirong Zheng","link":"https://arxiv.org/abs/2512.08161","category":["图像","图像去雾"],"conference_name":null,"conference_url":null},{"id":25120901,"created_at":"2025-12-11T11:12:54.865679+00:00","title":"FRWKV:Frequency-Domain Linear Attention for Long-Term Time Series Forecasting","date":"2025-12-09","content":"受 RWKV 线性注意力启发，本文提出了 FRWKV 模型，将线性注意力与频域分析相结合，用于长期时间序列预测。该方法在频域中实现 O(T) 复杂度的线性注意力，有效利用频谱信息增强特征表示，在多个基准测试中取得了领先性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251209-1.png","author":"Qingyuan Yang","link":"https://arxiv.org/abs/2512.07539","category":["序列/强化","时间序列预测"],"conference_name":"","conference_url":null},{"id":25120101,"created_at":"2025-12-11T11:12:43.687274+00:00","title":"EG-Net: Edge-Global aware network for accurate skin lesion segmentation","date":"2025-12-01","content":"该研究利用 RWKV 构建边缘-全局特征融合模块，以增强皮肤病变分割的全局上下文建模能力。结合边缘特征提取与通道增强解码器，EG-Net 有效解决了病变边界模糊问题，在多个公开数据集上实现了超越 SOTA 的精度与泛化性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251201-1.png","author":"Hongyu Cai","link":"https://www.sciencedirect.com/science/article/abs/pii/S1746809425018142","category":["图像","皮肤病变分割"],"conference_name":"BSPC | 中科院 2 区","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S1746809425018142"},{"id":25112003,"created_at":"2025-12-30T11:02:20+00:00","title":"Robin: RWKV Accelerator using Block Circulant Matrices based on FPGA","date":"2025-11-20","content":"论文介绍了 Robin，一种针对 RWKV 线性注意力模型的 FPGA 加速器软硬件协同设计方案。\n为了解决 RWKV 在 FPGA 上的存储和计算瓶颈，Robin 在算法层提出了部分块循环矩阵（PBCM）技术来压缩权重并保持精度 ；在硬件层设计了可配置循环计算核心，利用 DSP 封装策略高效支持循环和标准矩阵运算 。实验显示，相比 Tesla A100 GPU，Robin 实现了高达 3.09 倍的吞吐量提升和 7.31 倍的能效提升 。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251120-3.png","author":"Zeyu Li","link":"https://ieeexplore.ieee.org/document/11240845","category":["语言","AI 硬件推理加速"],"conference_name":"ICCAD | CCF B","conference_url":"https://ieeexplore.ieee.org/document/11240845"},{"id":25112002,"created_at":"2025-12-11T11:11:52.208937+00:00","title":"AFF-UNet-RWKV: A Lightweight Model for High-Quality Deblurring in Medical Imaging","date":"2025-11-20","content":"本文提出的 AFF-UNet-RWKV 模型，通过集成 RWKV-lite 空间混合器来捕获长程空间依赖，并结合 AFF 模块融合编解码器特征，实现了高效的医学图像去模糊。该轻量级模型在 PSNR 和 SSIM 指标上均优于传统方法，展现了优越的恢复性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251120-2.png","author":"Zhiyu Qin","link":"https://madison-proceedings.com/index.php/aetr/article/view/4338","category":["图像","图像去模糊"],"conference_name":null,"conference_url":null},{"id":25112001,"created_at":"2025-11-24T17:21:37.844401+00:00","title":"Evolution Strategies at the Hyperscale","date":"2025-11-20","content":"该研究在 RWKV-7 等大语言模型上应用了一种名为 EGGROLL 的新型演化策略算法，以实现高效微调。EGGROLL 通过引入低秩矩阵扰动代替传统 ES 的全秩扰动，显著降低了训练数十亿参数模型的计算与内存开销，从而能够支持超大规模的种群优化。该方法在强化学习和纯整数语言模型预训练等任务中也展示了竞争性的性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251120-1.png","author":"Bidipta Sarkar","link":"https://www.arxiv.org/abs/2511.16652","category":"[\"通用\",\"模型高效微调\"]","conference_name":"","conference_url":""},{"id":25111801,"created_at":"2025-11-24T17:20:23.955263+00:00","title":"基于动态邻接融合与通道混合的图神经网络社团检测方法","date":"2025-11-18","content":"该研究受 RWKV-style ChannelMix 架构启发，提出图通道混合器 (GCM) 模块以增强节点特征表达。结合动态邻接融合 (DAF) 模块，论文构建了时序-通道图注意力网络 (TC-GAT) 用于动态社团检测。该框架协同建模时空演化与通道交互，在多个动态图数据集上验证了其在检测精度与效率上的优越性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251118-1.png","author":"艾均","link":"https://www.arocmag.cn/abs/2025.07.0271","category":"[\"序列/强化\",\"社团检测\"]","conference_name":"计算机应用研究","conference_url":"https://www.arocmag.cn/abs/2025.07.0271"},{"id":25111701,"created_at":"2025-11-24T17:19:06.952766+00:00","title":"RawRWKV: An efficient raw image enhancement framework via RWKV architecture","date":"2025-11-17","content":"该研究提出 RawRWKV 框架，首次将 RWKV 架构用于低光照 raw 图像增强任务。通过结合 RWKV 的线性注意力机制与 U-Net 框架，该模型在显著降低计算复杂度的同时，实现了超越 CNN 和 Transformer 基线的 SOTA 图像增强效果，有效平衡了性能与效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251117-1.png","author":"Xingbo Dong","link":"https://link.springer.com/article/10.1007/s11760-025-04940-9","category":"[\"图像\",\"图像增强\"]","conference_name":"","conference_url":""},{"id":25111601,"created_at":"2025-11-24T17:17:55.186096+00:00","title":"ASALP: An Automatic Scaling Architecture for Edge Node Resources Based on Load Prediction","date":"2025-11-16","content":"该研究利用增强的 RWKV-EFE 模型进行负载预测，提出了一种边缘资源自动伸缩架构 ASALP。该架构在 Kubernetes-KubeEdge 框架中实现主动扩缩容，解决了原生 HPA 机制的延迟问题，从而显著提升了边缘环境的请求成功率和系统稳定性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251116-1.png","author":"Hui Liu","link":"https://link.springer.com/chapter/10.1007/978-3-032-10466-3_32","category":"[\"序列/强化\",\"负载预测\"]","conference_name":"NPC 2025 | CCF C","conference_url":"https://link.springer.com/chapter/10.1007/978-3-032-10466-3_32"},{"id":25111101,"created_at":"2025-11-14T16:09:25.454147+00:00","title":"Otter: Mitigating Background Distractions of Wide-Angle Few-Shot Action Recognition with Enhanced RWKV","date":"2025-11-11","content":"本文提出 Otter 模型，通过增强 RWKV 架构解决广角小样本动作识别的背景干扰问题。该方法设计复合分割模块 (CSM) 凸显主体，并引入时间重构模块 (TRM) 进行双向扫描以恢复时序关系，从而显著提升在复杂场景下的识别性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251111-1.png","author":"Wenbo Huang","link":"https://arxiv.org/abs/2511.06741","category":"[\"3D/视频\",\"小样本动作识别\"]","conference_name":"AAAI 2026 Oral | CCF A","conference_url":"https://arxiv.org/abs/2511.06741"},{"id":25111001,"created_at":"2025-11-14T16:08:05.47742+00:00","title":"MRT: Learning Compact Representations with Mixed RWKV-Transformer for Extreme Image Compression","date":"2025-11-10","content":"本文提出一种混合 RWKV-Transformer (MRT) 架构用于极端图像压缩。该架构结合 RWKV 的全局建模和 Transformer 的局部建模能力，将图像编码为更紧凑的 1D 隐式表示，并设计了专用的 RWKV 压缩模型 (RCM) 进一步提升压缩效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251110-1.png","author":"Han Liu","link":"https://arxiv.org/abs/2511.06717","category":"[\"图像\",\"图像压缩\"]","conference_name":"","conference_url":""},{"id":25103001,"created_at":"2025-12-11T11:11:40.49872+00:00","title":"RWKVSR: Receptance Weighted Key-Value Network for Hyperspectral Image Super-Resolution","date":"2025-10-30","content":"RWKVSR 引入了 RWKV 架构用于高光谱图像超分辨率，通过线性复杂度模块实现高效全局建模，结合方向可分离 3D 卷积和频域损失优化光谱一致性，在多个数据集上实现最佳性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251030-2.png","author":"Xiaofei Yang","link":"https://ieeexplore.ieee.org/document/11222729","category":["图像","图像超分辨率"],"conference_name":null,"conference_url":null},{"id":25103002,"created_at":"2025-11-10T14:08:17.567811+00:00","title":"SleepRWKVNet: A multimodal sleep staging network integrating bidirectional interactive RWKV and physiological prior-driven sequence-aware loss","date":"2025-10-30","content":"该研究提出 SleepRWKVNet，一种基于双向交互 RWKV 的多模态睡眠分期网络。模型通过创新的 Bi-IFM 模块高效融合长序列生理信号并解决模态贡献不一致问题，同时引入基于生理先验的序列感知损失函数 PS-Loss，有效缓解类别不平衡，提升了自动睡眠分期的准确性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251030-1.png","author":"Hang Zhang","link":"https://www.sciencedirect.com/science/article/abs/pii/S1746809425012248","category":"[\"序列/强化\",\"睡眠分期\"]","conference_name":"BSPC | 中科院 2 区","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S1746809425012248"},{"id":25102901,"created_at":"2025-11-10T14:08:06.240299+00:00","title":"WKV-sharing embraced random shuffle RWKV high-order modeling for pan-sharpening","date":"2025-10-29","content":"该研究提出一种基于 RWKV 的遥感图像全色锐化新范式 RS-RWKV。为解决 Vision RWKV 中的固定扫描偏差，该方法引入了贝叶斯启发的随机洗牌（Random Shuffle）扫描策略。同时，通过 WKV 共享机制实现高阶建模，有效降低延迟并提升了模型性能，在多个基准测试中表现优越。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251029-1.png","author":"Man Zhou","link":"https://openreview.net/forum?id=gqfQfqDQhx","category":"[\"图像\",\"全色锐化\"]","conference_name":"NeurIPS 2025 | CCF A","conference_url":"https://openreview.net/forum?id=gqfQfqDQhx"},{"id":25102704,"created_at":"2025-10-28T15:47:34.357263+00:00","title":"RWKV3D: An RWKV-Based Model with Multiple Training Strategies for Point Cloud Analysis","date":"2025-10-27","content":"本文提出一种基于 RWKV 架构的点云分析模型 RWKV3D，通过引入局部特征混合器 (LFM) 和双向多头移位 (BMS) 机制，有效提升了特征提取能力。该模型支持多种训练策略，在点云分类等任务中以更少的计算成本超越了 Transformer 和 Mamba 模型，取得了 SOTA 性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251027-4.png","author":"Chenglong Sun","link":"https://dl.acm.org/doi/abs/10.1145/3746027.3755658","category":"[\"3D/视频\",\"点云分析\"]","conference_name":"ACM MM 2025 | CCF A","conference_url":"https://dl.acm.org/doi/abs/10.1145/3746027.3755658"},{"id":25102701,"created_at":"2025-10-28T15:39:01.970356+00:00","title":"RWKV-PCSSC: Exploring RWKV Model for Point Cloud Semantic Scene Completion","date":"2025-10-27","content":"该研究提出 RWKV-PCSSC，一个受 RWKV 机制启发的轻量级点云语义场景补全网络。通过创新的 RWKV 种子生成器 (RWKV-SG) 和点反卷积 (RWKV-PD) 模块，该方法在多个数据集上实现了 SOTA 性能，同时显著减少了模型参数量和内存占用。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251027-3.png","author":"Wenzhe He","link":"https://dl.acm.org/doi/abs/10.1145/3746027.3754908","category":"[\"3D/视频\",\"场景补全\"]","conference_name":"ACM MM 2025 | CCF A","conference_url":"https://dl.acm.org/doi/abs/10.1145/3746027.3754908"},{"id":25102702,"created_at":"2025-10-28T15:38:32.034445+00:00","title":"Learning Structural Priors via Laplacian RWKV Diffusion with Light-Effect Dataset for Nighttime Visibility Enhancement","date":"2025-10-27","content":"该研究提出一种基于 RWKV 与扩散模型的夜间图像增强方法，通过设计的双环路拉普拉斯 RWKV (Lap-RWKV) 提取结构先验，以指导模型同时抑制光效并增强低光区域。作者还构建了首个包含光效的配对夜间数据集 NightLight。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251027-2.png","author":"Dirui Xie","link":"https://dl.acm.org/doi/abs/10.1145/3746027.3755510","category":"[\"图像\",\"图像增强\"]","conference_name":"ACM MM 2025 | CCF A","conference_url":"https://dl.acm.org/doi/abs/10.1145/3746027.3755510"},{"id":25102703,"created_at":"2025-10-28T15:38:20.009103+00:00","title":"Freq-RWKV: Granularity-Aware Spatial-Frequency Synergy via Dual-Domain Recurrent Scanning for Pan-sharpening","date":"2025-10-27","content":"本文为解决遥感图像全色锐化问题，首次提出 Freq-RWKV 框架。该框架改进了 RWKV 架构，通过小波引导的双域扫描策略克服其固有的局部一致性缺陷。其 U 形网络协调空间与频率域扫描，实现了从粗到细的特征增强，有效恢复高频细节。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251027-1.png","author":"Xueheng Li","link":"https://dl.acm.org/doi/abs/10.1145/3746027.3755521","category":"[\"图像\",\"全色锐化\"]","conference_name":"ACM MM 2025 | CCF A","conference_url":"https://dl.acm.org/doi/abs/10.1145/3746027.3755521"},{"id":25102201,"created_at":"2025-10-27T17:12:29.907519+00:00","title":"RS³-RWKV: Leveraging RWKV for Efficient Remote Sensing Semantic Segmentation","date":"2025-10-22","content":"该研究基于 RWKV 架构提出 RS3-RWKV 框架，用于高分辨率遥感图像语义分割。通过设计的邻近敏感 WKV 注意力 (PS-WKV) 和尺度自适应位移机制 (SA-Shift)，模型有效捕捉了空间连续性和多尺度特征，在提升分割精度的同时保持了计算效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251022-1.png","author":"Qingwang Wang","link":"https://ieeexplore.ieee.org/abstract/document/11214221","category":"[\"图像\",\"语义分割\"]","conference_name":"IEEE JSTARS | JCR Q1","conference_url":"https://ieeexplore.ieee.org/abstract/document/11214221"},{"id":25101001,"created_at":"2025-10-27T17:12:08.394545+00:00","title":"FS-RWKV: Leveraging Frequency Spatial-Aware RWKV for 3T-to-7T MRI Translation","date":"2025-10-10","content":"本研究提出基于 RWKV 的 FS-RWKV 框架，用于将低场 3T MRI 图像转换为高场 7T 图像。该模型引入频率空间全向位移 (FSO-Shift) 和结构保真度增强模块 (SFEB)，通过结合小波分解与多域特征融合，有效提升了合成图像的解剖细节与全局对比度，性能优于现有方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251010-1.png","author":"Yingtie Lei","link":"https://arxiv.org/abs/2510.08951","category":"[\"图像\",\"医学图像转换\"]","conference_name":"BIBM 2025 | CCF B","conference_url":"https://arxiv.org/abs/2510.08951"},{"id":25100801,"created_at":"2025-11-14T16:06:45.465017+00:00","title":"DREAMSTATE: Diffusing States and Parameters for Recurrent Large Language Models","date":"2025-10-08","content":"本文针对 RWKV 模型提出了 DREAMSTATE 框架，利用 Diffusion Transformer (DiT) 对其内部状态的概率流形进行建模，实现了可控的文本生成。研究进一步将静态 WKV 参数视为“结构性噪声”，设计了一种混合架构，通过 DiT 动态生成这些参数以适应全局上下文，验证了该设计的可行性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251008-2.png","author":"刘潇","link":"https://openreview.net/forum?id=HHsD970kdE","category":"[\"语言\",\"模型架构\"]","conference_name":"","conference_url":""},{"id":25100802,"created_at":"2025-11-10T14:07:53.870178+00:00","title":"Bridging Transformers and RWKV: Towards Efficient Multimodal Video Understanding","date":"2025-10-08","content":"该研究为解决长视频理解的效率瓶颈，提出一种 RWKV-Transformer 混合架构。通过将部分 Transformer 层替换为 RWKV 模块，并利用参数重用和渐进式蒸馏策略，模型在不进行令牌压缩的情况下显著提升了推理吞吐量，同时在多个视频理解基准上保持了与原模型相当甚至更优的性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251008-1.png","author":"","link":"https://openreview.net/forum?id=kmNqnwA4aV","category":"[\"3D/视频\",\"视频理解\"]","conference_name":"","conference_url":""},{"id":25100601,"created_at":"2025-10-10T18:23:09.619134+00:00","title":"GDSR: Global-Detail Integration through Dual-Branch Network with Wavelet Losses for Remote Sensing Image Super-Resolution","date":"2025-10-06","content":"论文首次将 RWKV 模型应用于遥感图像超分辨率，提出了 GDSR 双分支网络。该网络并行 RWKV 与 CNN 分别捕获全局和局部特征，通过特定模块进行融合，并结合小波损失函数增强细节恢复。实验表明，GDSR 在提升重建质量的同时，计算效率优于现有先进方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251006-1.png","author":"Qiwei Zhu","link":"https://ieeexplore.ieee.org/document/11192566","category":"[\"图像\",\"图像超分辨率\"]","conference_name":"IEEE TGRS | CCF B","conference_url":"https://ieeexplore.ieee.org/document/11192566"},{"id":25100101,"created_at":"2026-01-28T15:44:33.117846+00:00","title":"Pretrained Bi-RWKV Model","date":"2025-10-01","content":"论文在Bi-RWKV-at架构中，用双向Time Mixing模块替换单向RWKV的Time Mixing，实现双向上下文建模；并行处理前向和后向路径，通过加法合并输出，提升自然语言理解任务的性能；同时保持线性时间复杂度，相比Transformer获得1.95倍推理速度提升。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20251001-1.png","author":"Kyungbeen Cho","link":"https://koreascience.kr/article/CFKO202533772024748.page","category":["语言","语言模型"],"conference_name":null,"conference_url":null},{"id":25093001,"created_at":"2025-10-09T14:36:18.854822+00:00","title":"VRWKV-Editor: Reducing quadratic complexity in transformer-based video editing","date":"2025-09-30","content":"本文提出 VRWKV-Editor，一种基于 RWKV 架构的新型视频编辑模型，旨在解决传统 Transformer 注意力机制的二次方复杂度问题。该模型通过集成线性时空聚合模块，在处理长时高分辨率视频时，显著提升了运算速度并降低内存占用，同时保持了具有竞争力的编辑质量与时间一致性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250930-1.png","author":"Abdelilah Aitrouga","link":"https://arxiv.org/abs/2509.25998v2","category":"[\"3D/视频\",\"视频编辑\"]","conference_name":"","conference_url":""},{"id":25092701,"created_at":"2025-10-09T14:36:08.686808+00:00","title":"C3-OWD: A Curriculum Cross-modal Contrastive Learning Framework for Open-World Detection","date":"2025-09-27","content":"该论文基于 RWKV 架构提出 C3-OWD 框架，通过两阶段训练解决开放世界检测问题。第一阶段借助 RWKV 高效融合 RGB 与红外数据以增强模型鲁棒性；第二阶段通过视觉-语言对齐实现对新类别的泛化。该方法引入 EMA 机制防止灾难性遗忘，在鲁棒性和泛化性上均表现出色。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250927-1.png","author":"Siheng Wang","link":"https://arxiv.org/abs/2509.23316","category":"[\"图像\",\"开放世界检测\"]","conference_name":"","conference_url":""},{"id":25091901,"created_at":"2025-09-23T11:46:39.92139+00:00","title":"DPC-QA Net: A No-Reference Dual-Stream Perceptual and Cellular Quality Assessment Network for Histopathology Images","date":"2025-09-19","content":"该研究提出了用于组织病理学图像质量评估的双流网络 DPC-QA Net。其核心贡献是引入 Aggr-RWKV 模块，用于聚合细胞核与细胞膜的嵌入特征以进行精细的细胞结构质量感知。该模型结合了基于小波的全局差异感知与细胞级质量评估，有效解决了宏观与微观尺度上的质量不一致问题，并在多个数据集上展示了高准确性与泛化能力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250919-1.png","author":"Qijun Yang","link":"https://arxiv.org/abs/2509.15802","category":"[\"图像\",\"图像质量评估\"]","conference_name":"","conference_url":""},{"id":25091701,"created_at":"2025-09-23T11:46:28.333607+00:00","title":"Mastering Air Combat through Model-Based Reinforcement Learning","date":"2025-09-17","content":"本文将 RWKV 风格的线性注意力机制引入 DreamerV3 架构以实现并行训练，并提出一种用于自主空战的基于模型的强化学习智能体。该方法通过结合对比预测编码、Dyna 风格更新及课程学习自博弈流程，训练出单一端到端策略。实验证明，该智能体在样本效率和对新对手的快速适应性上显著优于基线模型。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250917-1.png","author":"Tianyu Lu","link":"https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5499593","category":"[\"序列/强化\",\"自主空战\"]","conference_name":"","conference_url":""},{"id":25091501,"created_at":"2025-09-19T11:11:30.012103+00:00","title":"Enhanced Traffic Sign Recognition via RWKV with Deformable Attention","date":"2025-09-15","content":"本文提出一种结合 RWKV 与可变形注意力 (Deformable Attention) 的新视觉编码器，以提升交通标志识别的性能。该模型利用 RWKV 的线性计算优势和可变形注意力的特征聚焦能力，在 GTSRB 数据集上实现了高精度与高效率，尤其在高分辨率图像处理上表现出显著的速度优势。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250915-2.png","author":"Jiahao Guo","link":"https://dl.acm.org/doi/10.1145/3757749.3757779","category":"[\"图像\",\"交通标志识别\"]","conference_name":"ICCMT","conference_url":"https://dl.acm.org/doi/10.1145/3757749.3757779"},{"id":25091502,"created_at":"2025-09-17T17:29:47.833835+00:00","title":"RWKV-VIO: An Efficient and Low-Drift Visual–Inertial Odometry Using an End-to-End Deep Network","date":"2025-09-15","content":"论文提出 RWKV-VIO ，一种基于 RWKV 架构的视觉惯性里程计，通过轻量级设计和线性计算复杂度，有效解决现有深度学习 VIO 方法在时间建模和计算效率上的挑战。该框架引入新型 IMU 编码器和并行编码策略，增强特征提取能力，并显著降低模型大小和推理时间，同时实现了领先的定位精度，适用于自主导航和机器人技术。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250915-1.png","author":"Jiaxi Yang","link":"https://www.mdpi.com/1424-8220/25/18/5737","category":"[\"3D/视频\",\"视觉惯性里程计\"]","conference_name":"Sensors | JCR Q1","conference_url":"https://www.mdpi.com/1424-8220/25/18/5737"},{"id":25091302,"created_at":"2025-10-09T14:35:55.058165+00:00","title":"Multi-modal dynamic brain graph representation learning for brain disorder diagnosis via temporal sequence model","date":"2025-09-13","content":"本文受 RWKV 架构启发，提出一种高效的时间多模态图神经网络 ET_MGNN，用于脑疾病诊断。该模型利用 RWKV 模块捕捉动态脑图序列的长短期时间依赖性，并有效融合动态功能连接与结构连接信息，在多个数据集上提升了分类准确性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250923-1.png","author":"Jinwei Lang","link":"https://www.sciencedirect.com/science/article/abs/pii/S0925231225021812","category":"[\"图像\",\"脑疾病诊断\"]","conference_name":"Neurocomputing | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0925231225021812"},{"id":25091301,"created_at":"2025-09-18T10:59:40.428665+00:00","title":"A Traditional Approach to Symbolic Piano Continuation","date":"2025-09-13","content":"论文基于 RWKV-7 架构构建了一个 20M 参数的小型模型，专注于符号化钢琴音乐续写。通过在精选数据上进行简单的下一词元预测训练，该模型在特定任务上的表现可与大 39 倍的 Transformer 模型相媲美，验证了小型专用模型的竞争力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250913-1.png","author":"Christian Zhou-Zheng","link":"https://arxiv.org/abs/2509.12267","category":"[\"音频/语音\",\"音乐续写\"]","conference_name":"","conference_url":""},{"id":25091201,"created_at":"2025-09-17T17:29:22.806506+00:00","title":"ITC-RWKV: Interactive Tissue–Cell Modeling with Recurrent Key-Value Aggregation for Histopathological Subtyping","date":"2025-09-12","content":"论文文提出 ITC-RWKV 模型，引入基于 Receptance Weighted Key-Value (RWKV) 架构的 Aggr-RWKV 机制，以线性复杂度高效聚合组织病理图像中的细胞特征。该模型采用双流架构，整合宏观组织特征与微观细胞表示，并设计双向组织-细胞交互模块。实验结果表明，ITC-RWKV 在多种组织病理学亚型分类任务中超越现有方法，证明了细胞级聚合与组织-细胞交互的关键作用。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250912-1.png","author":"Yating Huang","link":"https://research.manchester.ac.uk/en/publications/itc-rwkv-interactive-tissuecell-modeling-with-recurrent-key-value","category":"[\"图像\",\"组织病理分类\"]","conference_name":"BMVC 2025 | CCF C","conference_url":"https://research.manchester.ac.uk/en/publications/itc-rwkv-interactive-tissuecell-modeling-with-recurrent-key-value"},{"id":25091001,"created_at":"2025-09-17T17:29:05.738999+00:00","title":"EfficientIML: Efficient High-Resolution Image Manipulation Localization","date":"2025-09-10","content":"论文提出以 RWKV 骨干网络为核心的 EfficientIML 模型，该模型旨在高效处理高分辨率图像篡改定位任务。研究发布了首个超高分辨率 SIF 数据集，以应对扩散生成式伪造挑战。EfficientIML 采用轻量级三阶段 EfficientRWKV 骨干，结合混合状态空间与注意力机制，实现全局与局部细节并行捕捉，并在性能和推理速度上优于现有方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250910-1.png","author":"Jinhan Li","link":"https://arxiv.org/abs/2509.08583","category":"[\"图像\",\"图像篡改定位\"]","conference_name":"","conference_url":""},{"id":25090801,"created_at":"2025-11-14T16:05:28.795381+00:00","title":"Robotic control optimization based on receptance-weighted reinforcement learning","date":"2025-09-08","content":"本文将改进的 RWKV 架构应用于强化学习和机器人控制。研究优化了 RWKV 的通道混合模块，并用其替代 Decision Transformer 中的自注意力模块。在 D4RL 数据集上的实验表明，该方法相比基线模型具有更快的速度和更高的准确性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250908-1.png","author":"Zhaomin Zhu","link":"https://www.spiedigitallibrary.org/conference-proceedings-of-spie/13801/138011N/Robotic-control-optimization-based-on-receptance-weighted-reinforcement-learning/10.1117/12.3076952.short","category":"[\"序列/强化\",\"机器人控制\"]","conference_name":"CVAR 2025","conference_url":"https://www.spiedigitallibrary.org/conference-proceedings-of-spie/13801/138011N/Robotic-control-optimization-based-on-receptance-weighted-reinforcement-learning/10.1117/12.3076952.short"},{"id":25090501,"created_at":"2025-09-17T17:28:44.858744+00:00","title":"Spectral Channel Mixing Transformer with Spectral-Center Attention for Hyperspectral Image Classification","date":"2025-09-05","content":"论文提出 TC-Former 框架，创新性地将 RWKV 线性注意力机制与 Transformer 结合，用于高光谱图像分类。通过 TimeMixFormer 和 HyperMixFormer 模块，该方法优化了计算复杂度，提高了长序列处理效率，同时增强了光谱特征的判别表示能力。实验结果显示，该框架在分类精度上超越了现有先进算法，显著降低了计算开销。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250905-1.png","author":"Zhenming Sun","link":"https://www.mdpi.com/2072-4292/17/17/3100","category":"[\"图像\",\"高光谱图像分类\"]","conference_name":"Remote Sensing | JCR Q1","conference_url":"https://www.mdpi.com/2072-4292/17/17/3100"},{"id":25090201,"created_at":"2025-09-08T15:03:07.127404+00:00","title":"AudioRWKV: Efficient and Stable Bidirectional RWKV for Audio Pattern Recognition","date":"2025-09-02","content":"AudioRWKV 针对 Transformer (O(L²) 复杂度) 和 Mamba (稳定性差) 在音频建模中的挑战，提出高效稳定的 AudioRWKV (A-RWKV) 架构。它基于 RWKV7，引入 2D 深度可分离卷积以捕获局部谱-时模式，并采用双向 WKV (Bi-WKV) 内核实现全局上下文建模。A-RWKV 保持线性复杂度，展现出卓越的稳定性与性能，尤其在长音频处理中效率显著提升。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250902-1.png","author":"Jiayu Xiong","link":"https://arxiv.org/abs/2509.02167","category":"[\"音频/语音\",\"音频模式识别\"]","conference_name":"","conference_url":""},{"id":25083001,"created_at":"2025-09-01T15:26:39.633752+00:00","title":"Hybrid CNN-RWKV with high-frequency enhancement for real-world chinese-english scene text image super-resolution","date":"2025-08-30","content":"本文提出 Hybrid CNN-RWKV 结合高频增强（HCR-HFE）模型，用于真实场景中文-英文文本图像超分辨率（STISR）。该模型引入循环双向 WKV 注意力捕获全局依赖，并设计高频增强模块、多尺度大核卷积块与多频通道注意力。实验证明，HCR-HFE 在 Real-CE 数据集上性能优越，具有广泛适用性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250830-1.png","author":"Yanbin Liu","link":"https://link.springer.com/article/10.1007/s10489-025-06785-8","category":"[\"图像\",\"场景文本图像超分辨率\"]","conference_name":"Applied Intelligence | JCR Q2","conference_url":"https://link.springer.com/article/10.1007/s10489-025-06785-8"},{"id":25082803,"created_at":"2025-09-01T15:26:25.160075+00:00","title":"Finch-LIC: Learned Image Compression with Gated Multihead Linear Attention","date":"2025-08-28","content":"该论文提出了 FinchLIC，引入了 Multihead Bi-RWKV 块，通过扩展内部状态规模增强特征提取能力。同时，提出的 K-Manhattan 距离 token 移位 (KMshift) 方法有效建模邻近上下文，扩展了模型的感受野，提升了率失真 (RD) 性能。实验证明 FinchLIC 在保持线性复杂度的同时，实现了有竞争力的 RD 性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250828-3.png","author":"Fangzhou Yi","link":"https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5413219","category":"[\"图像\",\"图像压缩\"]","conference_name":"","conference_url":""},{"id":25082802,"created_at":"2025-09-01T15:26:19.212151+00:00","title":"DSM-Seg: A CNN-RWKV Hybrid Framework for Forward-Looking Sonar Image Segmentation in Deep-Sea Mining","date":"2025-08-28","content":"该研究提出了 DSM-Seg，一个结合了 CNN 和 RWKV 模型的混合架构，用于解决深海采矿前视声纳 (FLS) 图像分割中的挑战。RGFSC 模块利用 RWKV 机制进行高效的全局融合，并引入语义约束以抑制噪声。在深海地形和海洋垃圾数据集上的实验表明，DSM-Seg 在复杂条件下显著提高了分割精度，同时满足了实时性能要求。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250828-2.png","author":"Xinran Liu","link":"https://www.mdpi.com/2072-4292/17/17/2997","category":"[\"图像\",\"图像分割\"]","conference_name":"Remote Sensing | JCR Q1","conference_url":"https://www.mdpi.com/2072-4292/17/17/2997"},{"id":25082801,"created_at":"2025-08-29T13:43:21.096349+00:00","title":"PointDGRWKV: Generalizing RWKV-like Architecture to Unseen Domains for Point Cloud Classification","date":"2025-08-28","content":"本文介绍了 PointDGRWKV，首个针对点云分类域泛化 (DG PCC) 的 RWKV-based 框架。它解决了 RWKV 直接应用于点云时，固定方向 Token Shift 导致的几何扭曲和 Bi-WKV 注意力机制易受跨域差异影响的问题。\r\n\r\n在 DG PCC 基准上实现了先进性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250828-1.png","author":"Hao Yang","link":"https://arxiv.org/abs/2508.20835","category":"[\"3D/视频\",\"点云域泛化\"]","conference_name":"","conference_url":""},{"id":25082101,"created_at":"2025-09-01T15:26:12.151545+00:00","title":"VAFTrack: asynchronous feature fusion via visual receptive weighted key-value perceptual for visual tracking","date":"2025-08-21","content":"该论文提出了 VAFTrack，一个基于 RWKV 模型双向注意力机制的异步融合视觉跟踪模型。它通过引入视觉感受加权键值感知融合模块（VPFM），有效解决了模板与搜索特征融合不足和冗余问题，显著提升了特征匹配精度、目标感知能力及对复杂场景的适应性，实现了在 TrackingNet、LaSOT 和 GOT-10k 数据集上的领先跟踪性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250821-1.png","author":"Zhongmin Liu","link":"https://link.springer.com/article/10.1007/s00530-025-01913-3","category":"[\"图像\",\"目标跟踪\"]","conference_name":"Multimedia Systems | JCR Q2","conference_url":"https://link.springer.com/article/10.1007/s00530-025-01913-3"},{"id":25081701,"created_at":"2025-08-22T15:53:56.916917+00:00","title":"REB-former: RWKV-enhanced E-branchformer for Speech Recognition","date":"2025-08-17","content":"本文提出了 REB-former，一种基于 RWKV 增强的 E-Branchformer 模型，用于自动语音识别 (ASR)。该模型交错结合了 E-Branchformer 和 RWKV 层，并引入 GroupBiRWKV 模块克服 RWKV 的单向性，旨在降低计算复杂度并提升语音建模能力。REB-former 在 LibriSpeech 100h 数据集上实现了先进的性能，显著降低了词错误率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250817-1.png","author":"Jie Song","link":"https://www.isca-archive.org/interspeech_2025/song25b_interspeech.html","category":"[\"音频/语音\",\"语音识别\"]","conference_name":"Interspeech 2025 | CCF C","conference_url":"https://www.isca-archive.org/interspeech_2025/song25b_interspeech.html"},{"id":25080501,"created_at":"2025-08-11T15:37:28.992677+00:00","title":"A Multimodal Bone Stick Matching Approach Based on Large-Scale Pre-Trained Models and Dynamic Cross-Modal Feature Fusion","date":"2025-08-05","content":"本文基于 RWKV 模型及其变体，提出了一种多模态骨签匹配方法。该方法利用 Vision-RWKV 提取视觉特征、RWKV 分析铭文、BERT 编码考古元数据，并通过动态跨模态特征融合机制整合信息。实验表明，该方法在骨签碎片匹配中达到94.73%的准确率，显著优于传统方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250805-1.png","author":"Tao Fan","link":"https://www.mdpi.com/2076-3417/15/15/8681","category":"[\"图像\",\"多模态骨签匹配\"]","conference_name":" Applied Sciences | JCR Q2","conference_url":"https://www.mdpi.com/2076-3417/15/15/8681"},{"id":25073001,"created_at":"2025-08-04T13:55:25.852869+00:00","title":"Monthly Service Prediction for 4G/5G Systems: A Short Time Series Based Neural Network Solution","date":"2025-07-30","content":"本文基于 RWKV 模型开发了一个短时间序列预测框架，包括深度时间聚类表示（DTCR）和递减时间差网络（DTD-Net）。DTCR 使用 RWKV 编码器聚类数据以提升内部逻辑，DTD-Net 通过裁剪月度特征块减少注意力操作防止过拟合。实验在中国移动服务数据上验证了框架的有效性，平均 MAPE 为 0.126 和 0.120。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250730-1.png","author":"Hailiang Lu","link":"https://ieeexplore.ieee.org/abstract/document/11104274","category":"[\"序列/强化\",\"时间序列预测\"]","conference_name":"IEEE TCCN | 中科院 1 区 TOP","conference_url":"https://ieeexplore.ieee.org/abstract/document/11104274"},{"id":25072901,"created_at":"2025-08-04T13:54:42.379779+00:00","title":"RWKV-Receptance Recurrent Key Value in the field of Speaker Diarization","date":"2025-07-29","content":"本文基于 RWKV 模型提出了一种高效说话人分离方法，通过将 RWKV 架构整合到端到端神经分离模型中，解决了传统注意力机制的计算复杂性问题。该方法在保持上下文建模能力的同时，显著提升了推理速度和内存效率，并在基准数据集上降低了分离错误率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250729-1.png","author":"Uthayasanker Thayasivam","link":"https://openreview.net/forum?id=5WG3x1hgdN","category":"[\"音频/语音\",\"说话人分离\"]","conference_name":"","conference_url":""},{"id":25072701,"created_at":"2025-10-09T14:35:43.088017+00:00","title":"SpikeRWKV:Energy-efficient Large Language Model with Spiking Neural Network","date":"2025-07-27","content":"该论文提出了 SpikeRWKV，一个基于 RWKV 架构并结合脉冲神经网络 (SNN) 的高能效大语言模型。为平衡性能与能耗，论文设计了一种创新的多头脉冲编码方案，支持并行处理和正负脉冲表示。实验证明，该模型在大幅降低能耗的同时，在自然语言理解 (NLU) 任务上保持了优异性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250727-1.png","author":"Yulu Zhang","link":"http://poster-openaccess.com/files/ICIC2025/4202.pdf","category":"[\"语言\",\"类脑语言模型\"]","conference_name":"ICIC 2025 | CCF C","conference_url":"http://poster-openaccess.com/files/ICIC2025/4202.pdf"},{"id":25072601,"created_at":"2025-07-29T14:58:46.491011+00:00","title":"LowKeyEMG: Electromyographic typing with a reduced keyset","date":"2025-07-26","content":"本文基于 RWKV 模型开发了 LowKeyEMG 系统，用于通过表面肌电信号实现高效文本输入。该系统仅使用 7 个手势键，结合语言模型优化文本重建，在实时实验中达到平均 23.3 词/分钟的输入速度，手势效率提升 17%，top-3 单词准确率达 99.2%。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250726-1.png","author":"Johannes Y. Lee","link":"https://arxiv.org/abs/2507.19736","category":"[\"序列/强化\",\"辅助人机交互\"]","conference_name":"","conference_url":""},{"id":25072501,"created_at":"2025-08-04T13:54:13.399821+00:00","title":"Smooth Reading: Bridging the Gap of Recurrent LLM to Self-Attention LLM on Long-Context Tasks","date":"2025-07-25","content":"本文基于 RWKV 模型提出 Smooth Reading 方法，这是一种分块推理策略，灵感来源于人类阅读方式。该方法将长上下文分成小块处理，迭代总结信息以减少内存需求，优化循环模型性能。实验显示，该方法在 LongBench 上提升性能高达 3.61%，并保持高效训练和推理速度，显著缩小与自注意力模型的差距。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250725-1.png","author":"Kai Liu","link":"https://arxiv.org/abs/2507.19353","category":"[\"通用\",\"推理优化\"]","conference_name":"","conference_url":""},{"id":25072401,"created_at":"2025-07-29T14:57:59.609777+00:00","title":"DRWKV: Focusing on Object Edges for Low-Light Image Enhancement","date":"2025-07-24","content":"本文基于 RWKV 模型提出了 DRWKV 模型，整合全局边缘 Retinex (GER) 理论以解耦光照与边缘结构，引入演化 WKV 注意力机制增强空间连续性。实验表明，该模型在多个低光增强基准测试中取得领先的 PSNR、SSIM 和 NIQE 分数，同时保持低计算复杂度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250724-1.png","author":"白雪城","link":"https://arxiv.org/abs/2507.18594","category":"[\"图像\",\"低光图像增强\"]","conference_name":"WACV 2026","conference_url":"https://api.rwkv.cn/project/default/editor/18489?sort=created_at%3Aasc"},{"id":25072301,"created_at":"2025-07-29T14:57:13.066363+00:00","title":"An Efficient Image Fusion Network Exploiting Unifying Language and Mask Guidance","date":"2025-07-23","content":"本文基于 RWKV 模型提出 RWKVFusion 框架，利用语言描述和语义掩码指导图像融合过程。该方法通过高效扫描策略将 RWKV 适配为双向版本，并引入多模态融合模块促进信息交换，构建轻量级网络以降低计算成本。在多种图像融合任务中实现了最先进的性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250723-1.png","author":"Zi-Han Cao","link":"https://ieeexplore.ieee.org/abstract/document/11091495","category":"[\"图像\",\"图像融合\"]","conference_name":"IEEE TPAMI | CCF A","conference_url":"https://ieeexplore.ieee.org/abstract/document/11091495"},{"id":25071503,"created_at":"2025-08-08T17:23:18.779635+00:00","title":"MSFF-RWKV : Single-Structure Multi-stage Feature Fusion Lightweight Super-Resolution Network","date":"2025-07-15","content":"本文基于 RWKV 模型提出了一种轻量级图像超分辨率网络 MSFF-RWKV。通过单块多阶段特征融合策略，结合 LPP-Shift 和 ME-Shift 模块，显著减少了参数和计算复杂度，同时提升了图像重建质量。实验表明，该模型在 PSNR 和 SSIM 指标上优于现有方法，参数减少 26.6%。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250715-3.png","author":"Xiang Ren","link":"https://link.springer.com/chapter/10.1007/978-981-96-9949-0_35","category":"[\"图像\",\"图像超分辨率\"]","conference_name":"ICIC 2025 | CCF C","conference_url":"https://link.springer.com/chapter/10.1007/978-981-96-9949-0_35"},{"id":25071502,"created_at":"2025-07-21T14:25:08.891942+00:00","title":"DEVR: Train an Efficient Vision-RWKV Model with Improved Knowledge Distillation","date":"2025-07-15","content":"本文基于 RWKV 模型提出了一种高效的视觉模型 DEVR，通过改进知识蒸馏方法。重新设计了 RWKV 块以增强通道特征和空间信息捕获，并引入结合对比学习和蒸馏的损失函数，分阶段对齐特征空间。实验表明，DEVR 在图像分类、检测和分割任务中优于现有模型，计算成本更低、速度更快。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250715-2.png","author":"Xuhong Li","link":"https://link.springer.com/chapter/10.1007/978-981-96-9794-6_29","category":"[\"图像\",\"高效视觉模型\"]","conference_name":"ICIC 2025 | CCF C","conference_url":"https://link.springer.com/chapter/10.1007/978-981-96-9794-6_29"},{"id":25071501,"created_at":"2025-07-16T15:43:30.002703+00:00","title":"U-RWKV: Lightweight medical image segmentation with direction-adaptive RWKV","date":"2025-07-15","content":"本文基于 RWKV 模型提出 U-RWKV 框架，用于轻量级医学图像分割。该框架引入方向自适应 RWKV 模块 (DARM) 和阶段自适应挤压-激励模块 (SASE)，高效建模长距离依赖并减少方向偏差，在资源受限环境中实现高性能分割。实验验证了其优越的计算效率和分割精度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250715-1.png","author":"Hongbo Ye","link":"https://arxiv.org/abs/2507.11415","category":"[\"图像\",\"医学图像分割\"]","conference_name":"MICCAI 2025 | CCF B","conference_url":"https://link.springer.com/chapter/10.1007/978-3-032-05141-7_59"},{"id":25070601,"created_at":"2025-07-09T16:14:46.125454+00:00","title":"Scaling Context Requires Rethinking Attention","date":"2025-07-06","content":"本文基于 RWKV 模型提出了 power attention 优化方案，一种线性成本的序列建模层，通过独立调整状态大小解决长上下文训练问题。实验显示其在上下文学习中优于指数和线性注意力，并开发高效 GPU 内核实现速度提升。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250706-2.jpg","author":"Carles Gelada","link":"https://arxiv.org/abs/2507.04239","category":"[\"通用\",\"架构优化方案\"]","conference_name":"","conference_url":""},{"id":25070301,"created_at":"2025-07-07T11:29:00.809655+00:00","title":"AuroraLong: Bringing RNNs Back to Efficient Open-Ended Video Understanding","date":"2025-07-03","content":"本文基于 RWKV 模型提出了 AURORA LONG，通过将 MLLMs 中的 LLM 组件替换为 RWKV 模型，以恒定大小的隐藏状态处理任意长度的输入序列。研究结合视觉令牌合并，通过按大小升序重新排列视觉令牌，显著提高了处理效率。AURORA LONG 在多个视频基准测试中表现出与基于 Transformer 的模型相当的性能，同时降低了算力消耗。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250703-1.png","author":"Weili Xu","link":"https://arxiv.org/abs/2507.02591","category":"[\"3D/视频\",\"视频理解\"]","conference_name":"ICCV 2025 | CCF A","conference_url":"https://arxiv.org/abs/2507.02591"},{"id":25070101,"created_at":"2025-07-09T16:13:51.368987+00:00","title":"EvRWKV: A Continuous Interactive RWKV Framework for Effective Event-Guided Low-Light Image Enhancement","date":"2025-07-01","content":"本文基于 RWKV 模型提出 EvRWKV 框架，通过双域处理实现事件与图像的连续跨模态交互。该框架采用 Cross-RWKV 模块进行细粒度时空融合，结合 EISFE 模块实现自适应频域噪声抑制与空域形变卷积对齐。在真实低光数据集上的实验表明，该方法能有效抑制噪声、恢复结构细节并提升视觉清晰度，达到 SOTA 性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250701-1.png","author":"WenJie Cai","link":"https://arxiv.org/abs/2507.03184","category":"[\"图像\",\"低光增强\"]","conference_name":"","conference_url":""},{"id":25062601,"created_at":"2025-06-30T11:38:27.427844+00:00","title":"Out-of-Distribution Semantic Occupancy Prediction","date":"2025-06-26","content":"论文为解决自动驾驶中的“意外”物体识别难题，创新性地引入 RWKV 架构来强化模型的特征感知力，并提出了 OccOoD 框架。它巧妙融合了精细的 3D 体素和全局的鸟瞰图视角，能更准确地判断异常。在不影响常规物体识别性能的前提下，实现了对未知风险的精准捕获。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250626-1.png","author":"Yuheng Zhang","link":"https://arxiv.org/abs/2506.21185","category":"[\"3D/视频\",\"3D 语义占用预测\"]","conference_name":"","conference_url":""},{"id":25062401,"created_at":"2025-06-26T14:04:12.276472+00:00","title":"Accurate, fast, cheap: Choose three. Replacing Multi-Head-Attention with Bidirectional Recurrent Attention for Long-Form ASR","date":"2025-06-24","content":"本文研究了将多头注意力（MHA）替换为双向循环注意力（RA）在长语音识别（ASR）中的应用，发现双向 RWKV-Conformer 模型在保持相似准确率的同时，效率更高。通过引入 Direction Dropout 方法，进一步提升了模型的灵活性和性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250624-1.png","author":"Martin Ratajczak","link":"https://arxiv.org/abs/2506.19761","category":"[\"音频/语音\",\"语音识别\"]","conference_name":"Interspeech 2025 | CCF C ","conference_url":"https://arxiv.org/abs/2506.19761"},{"id":25061801,"created_at":"2025-06-26T14:04:05.380543+00:00","title":"SMNet: A Semantic Guided Mamba Network for Remote Sensing Change Detection","date":"2025-06-18","content":"本文基于 RWKV 模型和 Mamba 架构提出了一种新的遥感变化检测模型 SMNet，该模型通过整合多层次特征表示，有效解决了当前方法在变化检测任务中性能有限和特征表达能力不足的问题。实验结果表明，SMNet 在多个遥感变化检测基准数据集上表现出色，显著优于现有方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250618-1.png","author":"Guoxia Xu","link":"https://ieeexplore.ieee.org/abstract/document/11039697","category":"[\"图像\",\"遥感变化检测\"]","conference_name":"IEEE TAES | JCR Q1","conference_url":"https://ieeexplore.ieee.org/abstract/document/11039697"},{"id":25061701,"created_at":"2025-06-18T14:41:47.950818+00:00","title":"Exploring Diffusion with Test-Time Training on Efficient Image Restoration","date":"2025-06-17","content":"论文基于 RWKV 模型提出了 DiffRWKVIR 框架，该框架将测试时训练（TTT）与高效扩散相结合，通过 Omni-Scale 2D 状态演化扩展 RWKV 的位置依赖参数化，实现全局上下文感知，并通过块优化闪存处理加速计算，最终在图像修复任务中超越现有方法，显著提升了效率和效果。该论文还提出了先验引导的高效扩散方法，通过提取紧凑的图像先验表示，加速了训练和推理过程，同时解决了传统扩散模型中的计算低效问题。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250617-1.png","author":"Rongchang Lu","link":"https://arxiv.org/abs/2506.14541","category":"[\"图像\",\"图像修复\"]","conference_name":"","conference_url":""},{"id":25061601,"created_at":"2025-06-26T14:03:53.215823+00:00","title":"Blind Identification of Collective Motion Criticality using Sequence Model Predictive Entropy Variance","date":"2025-06-16","content":"本文基于 RWKV-7 序列模型提出了一种无参数的集体运动临界性识别方法，通过分析单智能体轨迹数据来检测 Vicsek 模型中的临界区域。该方法利用预测香农熵的方差作为指标，无需系统控制参数或全局信息，成功在 L=32 和 L=64 系统中识别出临界噪声水平，且结果符合有限尺寸缩放原理。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250616-3.png","author":"Tianyi Wu","link":"https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5297784","category":"[\"序列/强化\",\"集体运动临界性盲识别\"]","conference_name":"","conference_url":""},{"id":25061602,"created_at":"2025-06-19T16:23:28.639065+00:00","title":"A Parallel Processing Architecture for Long-Term Power Load Forecasting","date":"2025-06-16","content":"本文基于 RWKV-TS 模型提出了 MP-RWKV，通过并行处理路径解决长期电力负荷预测中不同预测范围的挑战。MP-RWKV 通过上下文状态机制和位置感知注意力机制，在短期和长期预测场景中均表现出色。实验结果表明，MP-RWKV 在 24 小时至 432 小时的预测范围内均优于现有基准模型，尤其在传统模型性能下降的长期预测中表现突出。显著提升了长期电力负荷预测的准确性和稳定性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250616-2.png","author":"Adil Rizki","link":"https://www.mdpi.com/2673-4591/97/1/26","category":"[\"序列/强化\",\"时间序列预测\"]","conference_name":"","conference_url":""},{"id":25061603,"created_at":"2025-06-18T14:41:38.237916+00:00","title":"Personalizable Long-Context Symbolic Music Infilling with MIDI-RWKV","date":"2025-06-16","content":"论文基于 RWKV 模型提出了 MIDI-RWKV ，一个用于个性化、多轨道、长上下文和可控符号音乐填充的新型模型。该模型采用 RWKV-7 线性架构，能够在边缘设备上实现高效且连贯的音乐协同创作。MIDI-RWKV 通过微调初始状态实现了在极小样本条件下的个性化。实验结果表明，MIDI-RWKV 在多项定量和定性指标上均优于现有方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250616-1.png","author":"Christian Zhou-Zheng","link":"https://arxiv.org/abs/2506.13001","category":"[\"音频/语音\",\"音乐生成\"]","conference_name":"","conference_url":""},{"id":25061401,"created_at":"2025-06-19T16:23:18.175438+00:00","title":"RWKV-IF: Efficient and Controllable RNA Inverse Folding via Attention-Free Language Modeling","date":"2025-06-14","content":"本文基于 RWKV 模型提出了一种名为 RWKV-IF 的高效可控 RNA 逆折叠框架，通过将结构到序列的生成建模为条件语言建模任务，以线性复杂度捕获长程依赖关系。研究引入了一种解码策略，结合 Top-k 采样、温度控制和 G-C 含量偏向，生成结构准确且具有生物物理意义的序列。显著优于传统搜索基线方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250614-1.png","author":"Gaoyuan Ji","link":"https://www.biorxiv.org/content/10.1101/2025.06.13.659654v1","category":"[\"序列/强化\",\"RNA 逆折叠\"]","conference_name":"","conference_url":""},{"id":25061201,"created_at":"2025-06-13T17:20:35.635299+00:00","title":"Med-URWKV: Pure RWKV With ImageNet Pre-training For Medical Image Segmentation","date":"2025-06-12","content":"论文基于 RWKV 模型提出了一种名为 Med-URWKV 的纯 RWKV 架构，该架构基于 U-Net 框架构建，并融入了基于 ImageNet 的预训练，以进一步探索 RWKV 在医学图像分割任务中的潜力。研究通过在七个数据集上的实验，验证了 Med-URWKV 在医学图像分割任务中的有效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250612-1.png","author":"Zhenhuan Zhou","link":"https://arxiv.org/abs/2506.10858","category":"[\"图像\",\"医学图像分割\"]","conference_name":"","conference_url":""},{"id":25060701,"created_at":"2025-06-10T11:34:40.466412+00:00","title":"Vision-QRWKV: Exploring Quantum-Enhanced RWKV Models for Image Classification","date":"2025-06-07","content":"论文基于 RWKV 模型提出了一种量子增强的混合架构 Vision-QRWKV，用于图像分类任务。通过将变分量子电路（VQC）集成到 RWKV 的通道混合组件中，模型提升了非线性特征转换能力。实验表明，该模型在多个医疗和标准图像数据集上表现优于经典模型。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250607-1.png","author":"Chi-Sheng Chen","link":"https://arxiv.org/abs/2506.06633","category":"[\"图像\",\"量子增强图像分类\"]","conference_name":"","conference_url":""},{"id":25060601,"created_at":"2025-06-11T15:05:32.847883+00:00","title":"VisualRWKV-HM: Enhancing linear visual-language models via hybrid mixing","date":"2025-06-06","content":"论文基于 RWKV 模型提出了 VisualRWKV-HM，这是一种具有线性复杂度的视觉语言模型，在单图像、多图像和多视图基准测试中均达到了 SOTA 性能。与基于 Transformer-Mamba 架构的混合模型 LongLLaVA 相比，它在上下文长度为 16K 时消耗的内存更少，吞吐量提高了 24%。此外，VisualRWKV-HM 具有良好的可扩展性，通过扩展状态编码器和解码器，可以进一步提高性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250606-1.png","author":"侯皓文【元始智能】","link":"https://www.sciencedirect.com/science/article/abs/pii/S1566253525004099","category":"[\"图像\",\"视觉语言模型\"]","conference_name":"Information Fusion | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S1566253525004099"},{"id":25060501,"created_at":"2025-06-10T11:34:33.605464+00:00","title":"FEAT: Full-Dimensional Efficient Attention Transformer for Medical Video Generation","date":"2025-06-05","content":"论文基于 RWKV 模型架构中的 WKV 注意力机制，提出了 FEAT 模型，通过统一的空间-时间-通道注意力机制解决医疗视频生成中通道交互不足、计算复杂度高和去噪指导粗糙的问题。在多个数据集上实现了高效高质量的医疗视频生成。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250605-1.png","author":"王绘涵【北航】","link":"https://arxiv.org/abs/2506.04956","category":"[\"3D/视频\",\"医学视频生成\"]","conference_name":"MICCAI 2025 early accepted | CCF B","conference_url":"https://link.springer.com/chapter/10.1007/978-3-032-05114-1_26"},{"id":25060401,"created_at":"2025-06-10T15:42:56.469311+00:00","title":"Pan-Sharpening via Causal-Aware Feature Distribution Calibration","date":"2025-06-04","content":"论文基于 RWKV 模型提出了一种新的全色锐化方法，通过因果推断解决网络优化中的频率不平衡问题。该方法在训练阶段利用 RWKV 架构的全局感受野，有效学习高频分量的长尾分布，并量化特征偏差的累积方向。实验结果表明，该方法在多个基准数据集上均优于现有先进方法，展示了其在全色锐化任务中的有效性和鲁棒性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250604-1.png","author":"Xueheng Li","link":"https://ieeexplore.ieee.org/abstract/document/11023855","category":"[\"图像\",\"全色锐化\"]","conference_name":"IEEE TGRS | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/11023855"},{"id":25060301,"created_at":"2025-06-10T11:34:25.18971+00:00","title":"Diet-Seg: Dynamic Hardness-Aware Learning for Enhanced Brain Tumor Segmentation","date":"2025-06-03","content":"论文基于 RWKV 模型提出了一种新型脑肿瘤分割框架 Diet-Seg，通过将基于熵的像素级难度估计与动态学习率调节策略结合，有效提升了脑肿瘤分割的准确性。Diet-Seg 框架采用 RWKV-UNet 作为主干网络，以捕捉全局空间依赖性。实验结果表明，Diet-Seg 在 BraTS2018–2021 数据集上表现优于现有方法，特别是在肿瘤子区域的分割上取得了显著提升。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250603-1.png","author":"Boya Lv","link":"https://www.biorxiv.org/content/10.1101/2025.05.31.657149v1","category":"[\"图像\",\"医疗图像分割\"]","conference_name":"","conference_url":""},{"id":25060101,"created_at":"2025-06-13T17:20:27.869286+00:00","title":"Relational Context Modeling for Improved Knowledge Graph Completion","date":"2025-06-01","content":"论文基于 RWKV 模型和 TuckER 模型，提出了一种名为 RCME 的混合架构，用于改进知识图谱补全。RCME 结合了 RWKV 的序列建模能力和动态嵌入生成，以及 TuckER 的关系解码鲁棒性，在链接预测和三元组分类任务中表现优于多种先进模型。实验结果表明，该架构在多个基准数据集上均取得了显著的性能提升。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250601-1.png","author":"Guoqi Lin","link":"https://www.engineeringletters.com/issues_v33/issue_6/EL_33_6_28.pdf","category":"[\"语言\",\"知识图谱补全\"]","conference_name":"","conference_url":""},{"id":25052902,"created_at":"2025-06-30T11:38:15.275911+00:00","title":"融合接收加权键值架构和球面几何特征的甲状腺结节分割方法","date":"2025-05-29","content":"本文基于 RWKV 模型提出了一种融合接收加权键值架构（RWKV）和球面几何特征（SGF）采样技术的甲状腺结节分割方法。该方法通过二维偏移预测和像素级采样位置调整，结合区块注意力模块（PAM），实现精确分割。实验表明，该方法在甲状腺结节区域分割数据集（TN3K）和甲状腺影像数字数据库（DDTI）上取得了优异的分割性能，且计算复杂度较低，为甲状腺结节精确分割提供了一种高效解决方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250529-2.png","author":"朱丽程","link":"https://www.biomedeng.cn/article/10.7507/1001-5515.202412009","category":"[\"图像\",\"图像分割\"]","conference_name":"生物医学工程学杂志","conference_url":"https://www.biomedeng.cn/article/10.7507/1001-5515.202412009"},{"id":25052901,"created_at":"2025-06-03T11:43:31.355996+00:00","title":"URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration","date":"2025-05-29","content":"论文基于 RWKV 模型，提出了一种统一的多状态视角模型 URWKV，用于低光照图像恢复。该模型通过定制化的 URWKV 块感知和分析复杂退化，利用多阶段状态实现自适应场景感知的亮度调制。显著提升了性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250529-1.png","author":"Rui Xu","link":"https://arxiv.org/abs/2505.23068","category":"[\"图像\",\"低光照图像恢复\"]","conference_name":"CVPR 2025 | CCF A","conference_url":"https://cvpr.thecvf.com/virtual/2025/poster/34219"},{"id":25052401,"created_at":"2025-05-26T14:30:43.077571+00:00","title":"RainRWKV: a deep RWKV model for video deraining","date":"2025-05-24","content":"论文提出了一种基于 RWKV 模型的 RainRWKV 框架，用于视频去雨任务。通过引入小波变换移位机制和管状嵌入机制，分别增强了模型对低频特征和高频细节的捕捉能力，从而在视频去雨任务中实现了卓越的性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250524-1.png","author":"Xijun Wang","link":"https://link.springer.com/article/10.1007/s00371-025-03965-y","category":"[\"3D/视频\",\"视频去雨\"]","conference_name":"The Visual Computer","conference_url":"https://link.springer.com/article/10.1007/s00371-025-03965-y"},{"id":25052201,"created_at":"2025-05-29T14:41:40.971554+00:00","title":"DualComp: End-to-End Learning of a Unified Dual-Modality Lossless Compressor","date":"2025-05-22","content":"论文提出了 DualComp，一种 RWKV-7 的统一双模态无损压缩器，首次实现了图像和文本数据的统一无损压缩。DualComp 在图像和文本数据集上的压缩性能实现 SOTA，且参数更少，支持桌面 CPU 上的近实时推理。其单模态变体在 Kodak 数据集上以仅 1.2% 的模型大小超越了之前的最佳图像压缩器约 9%。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250522-1.png","author":"Yan Zhao","link":"https://arxiv.org/abs/2505.16256","category":"[\"通用\",\"数据压缩\"]","conference_name":"","conference_url":""},{"id":25052001,"created_at":"2025-05-26T14:30:33.864035+00:00","title":"ModRWKV: Transformer Multimodality in Linear Time","date":"2025-05-20","content":"论文提出了一种基于 RWKV-7 架构的 ModRWKV 框架，探索了现代 RNN 架构在多模态场景下的应用。ModRWKV 通过动态自适应的异构模态编码器实现多源信息融合，并通过广泛的实验确定了性能与计算效率之间的最佳平衡。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250520-1.png","author":"Jiale Kang","link":"https://aclanthology.org/2025.emnlp-main.204/","category":"[\"图像\",\"多模态融合\"]","conference_name":"EMNLP 2025 Oral | CCF B","conference_url":"https://aclanthology.org/2025.emnlp-main.204/"},{"id":25051801,"created_at":"2025-05-26T14:30:26.934367+00:00","title":"Quantum-Enhanced Channel Mixing in RWKV Models for Time Series Forecasting","date":"2025-05-18","content":"论文提出了 QuantumRWKV 模型，将 RWKV 模型中的前馈网络部分替换为变分量子电路，以增强非线性表示能力。实验证明，该模型在处理非线性或混沌动力学的时间序列任务中表现更优。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250518-1.png","author":"Chi-Sheng Chen","link":"https://arxiv.org/abs/2505.13524","category":"[\"序列/强化\",\"量子增强时间序列预测\"]","conference_name":"","conference_url":""},{"id":25051601,"created_at":"2025-05-27T11:04:58.792837+00:00","title":"Maximizing Asynchronicity in Event-based Neural Networks","date":"2025-05-16","content":"论文提出了一种新的异步到同步框架 EVA，用于实时事件相机数据处理。该框架基于 RWKV-6 构建了高效的异步编码器，实现了逐事件的表示更新，并采用自监督学习方法获得具有高度泛化能力的事件表示。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250516-1.png","author":"Haiqing Hao","link":"https://arxiv.org/abs/2505.11165","category":"[\"图像\",\"事件相机异步检测\"]","conference_name":"","conference_url":""},{"id":25051201,"created_at":"2025-05-26T14:30:19.317435+00:00","title":"Spatio-Temporal Weighted Graph Reason Learning for Multivariate Time-Series Anomaly Detection","date":"2025-05-12","content":"论文提出了 STWGRL 框架，用于多元时间序列异常检测。其核心贡献包括基于 D-RWKV 模块高效捕获长期序列信息，结合 TaGAA 模块自适应聚合信号特征，从而在检测精度、时间成本和可靠性间取得平衡。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250512-1.png","author":"Huaqi Zhang","link":"https://ieeexplore.ieee.org/abstract/document/11002535","category":"[\"序列/强化\",\"时间序列异常检测\"]","conference_name":"IEEE IoT | 中科院 1 区","conference_url":"https://ieeexplore.ieee.org/abstract/document/11002535"},{"id":25050502,"created_at":"2025-05-12T13:36:10.557413+00:00","title":"RADLADS: Rapid Attention Distillation to Linear Attention Decoders at Scale","date":"2025-05-05","content":"论文提出 RADLADS 框架，通过注意力蒸馏将传统 softmax attention 的 Transformer 高效转换为线性注意力模型。基于 RWKV 架构开发了两种新型变体 RAD-RWKV6 和 RAD-RWKV7，显著改善了现有 RWKV 架构在模型转换中的兼容性问题，并在 7B 至 72B 参数量级上实现了接近原模型的推理质量。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250505-2.png","author":"Daniel Goldstein","link":"https://arxiv.org/abs/2505.03005","category":"[\"通用\",\"线性注意力模型转换\"]","conference_name":"COLM 2025","conference_url":"https://openreview.net/forum?id=38GehGepDd#discussion"},{"id":25050501,"created_at":"2025-05-06T11:04:49.186756+00:00","title":"Multi-View Learning with Context-Guided Receptance for Image Denoising","date":"2025-05-05","content":"论文基于 RWKV 模型提出 CRWKV 架构，通过引入双向 BiWKV 机制突破因果约束，实现线性复杂度的像素序列交互。结合 Context-guided Token Shift (CTS) 机制增强噪声分布建模，并通过 Frequency Mix 模块整合频域特征，在图像去噪任务中取得 SOTA 效果，推理时间减少 40%。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250505-1.png","author":"Binghong Chen","link":"https://arxiv.org/abs/2505.02705","category":"[\"图像\",\"图像去噪\"]","conference_name":"IJCAI 2025 | CCF A","conference_url":"https://arxiv.org/abs/2505.02705"},{"id":25050201,"created_at":"2025-05-12T13:35:48.754985+00:00","title":"RWKVQuant: Quantizing the RWKV Family with Proxy Guided Hybrid of Scalar and Vector Quantization","date":"2025-05-02","content":"论文提出了 RWKVQuant，一种专门针对 RWKV 模型的训练后量化框架。通过结合标量量化和向量量化技术，并设计基于信息熵的代理策略与码本优化算法，该框架成功将 RWKV-14B 模型压缩至约 3 位宽，在精度损失小于 1% 的同时实现 2.14 倍加速。实验证明了该方法在语言和视觉任务上的有效性，是首个针对 RWKV 家族的完整量化解决方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250502-1.png","author":"许晨（后摩智能）","link":"https://arxiv.org/abs/2505.03803","category":"[\"通用\",\"模型量化\"]","conference_name":"ICML 2025 | CCF A","conference_url":"https://openreview.net/forum?id=UOw6Qt0qYU&noteId=Ac66N39Wuc"},{"id":25043002,"created_at":"2025-05-06T16:19:41.814432+00:00","title":"Multiple Span Bidirectional RWKV Network for Infrared Image Super-Resolution","date":"2025-04-30","content":"论文提出了一种基于 RWKV 模型的多跨度双向 MSB-RWKV 网络用于红外图像超分辨率。通过改进 RWKV 的注意力机制，设计了 MSB-WKV 线性复杂度全局注意力模块和 Wide Shift 局部特征增强层，实现了红外图像长程依赖建模与局部细节恢复的高效平衡。实验表明该方法在红外图像超分辨率任务中优于现有技术。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250430-2.png","author":"Zhaofei Xu","link":"https://link.springer.com/article/10.1007/s13042-025-02644-7","category":"[\"图像\",\"红外图像超分辨率\"]","conference_name":"","conference_url":""},{"id":25043001,"created_at":"2025-05-06T14:50:51.311339+00:00","title":"RWKV-X: A Linear Complexity Hybrid Language Model","date":"2025-04-30","content":"论文提出了 RWKV-X 混合语言模型，通过将 RWKV 的短程建模效率与新型稀疏注意力机制结合，显著提升了长上下文处理能力。该模型在 64K token 序列上持续预训练后，在长上下文基准测试中超越前期 RWKV-7 模型，同时保持线性训练时间复杂度和恒定推理解码复杂度，支持百万级 token 序列解码。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250430-1.png","author":"侯皓文【元始智能】","link":"https://arxiv.org/abs/2504.21463","category":"[\"通用\",\"混合语言模型\"]","conference_name":"","conference_url":""},{"id":25041701,"created_at":"2025-04-21T14:38:50.960752+00:00","title":"Zig-RiR: Zigzag RWKV-in-RWKV for Efficient Medical Image Segmentation","date":"2025-04-17","content":"论文提出了一种基于 RWKV 模型的嵌套结构 Zig-RiR，用于高效医学图像分割。通过将图像块分解，结合 Outer Zig-RWKV 和 Inner Zig-RWKV 模块分别捕获全局和局部特征，并引入 Zigzag-WKV 注意力机制保持空间连续性。该方法在医学图像数据集上实现了 14.4 倍的速度提升和 89.5% 的内存优化。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250417-1.png","author":"Tianxiang Chen","link":"https://ieeexplore.ieee.org/document/10969076","category":"[\"图像\",\"医学图像分割\"]","conference_name":" IEEE TMI | 中科院 1 区","conference_url":"https://ieeexplore.ieee.org/document/10969076"},{"id":25041401,"created_at":"2025-04-18T17:52:52.054672+00:00","title":"RGB-Event based Pedestrian Attribute Recognition: A Benchmark Dataset and An Asymmetric RWKV Fusion Framework","date":"2025-04-14","content":"论文提出了首个基于 RGB-Event 的多模态行人属性识别基准数据集 EventPAR，并提出了一种新型的 RWKV 融合框架。通过结合 RWKV 视觉编码器和非对称 RWKV 融合模块，有效整合 RGB 帧的空间特征与事件流的时间信息。实验表明该方法在三个基准数据集上达到了最优性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250414-1.png","author":"Xiao Wang","link":"https://arxiv.org/abs/2504.10018","category":"[\"图像\",\"多模态行人属性识别\"]","conference_name":"","conference_url":""},{"id":25041001,"created_at":"2025-04-16T11:46:40.334216+00:00","title":"MolRWKV: Conditional Molecular Generation Model Using Local Enhancement and Graph Enhancement","date":"2025-04-10","content":"论文提出 MolRWKV 模型，基于 RWKV 架构创新性地结合 CNN 与 GCN 技术，实现对条件分子生成的精准控制。RWKV 模型为 SMILES 序列处理提供长程依赖学习能力。MolRWKV 在无条件生成任务中达到基准模型水平，在条件生成任务中显著提升支架相似性控制精度，并能生成具有特定靶蛋白亲和力的分子。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250410-1.png","author":"Xihan Li","link":"https://onlinelibrary.wiley.com/doi/abs/10.1002/jcc.70100","category":"[\"序列/强化\",\"条件分子生成\"]","conference_name":"Journal of Computational Chemistry","conference_url":"https://onlinelibrary.wiley.com/doi/abs/10.1002/jcc.70100"},{"id":25040701,"created_at":"2025-04-18T17:52:14.942197+00:00","title":"Kinematic Modeling of a 7-DOF Tendon-Like-Driven Robot Based on Optimization and Deep Learning","date":"2025-04-07","content":"论文提出了一种结合循环网络与自注意力机制的深度学习微调模型 RWKV-TDR7，用于 7-DOF 肌腱驱动冗余机器人的复杂轨迹规划。通过融合 RWKV 模型的时间序列处理能力和自注意力机制，在保持轨迹拟合精度的同时显著降低了计算复杂度（O (N)），支持长序列输出并提升实时控制效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250407-1.png","author":"SaiXuan Chen","link":"https://onlinelibrary.wiley.com/doi/abs/10.1002/rob.22544","category":"[\"3D/视频\",\"逆运动学与轨迹规划\"]","conference_name":"","conference_url":""},{"id":25032801,"created_at":"2025-04-07T07:06:54.030206+00:00","title":"DREMnet: An Interpretable Denoising Framework for Semi-Airborne Transient Electromagnetic Signal","date":"2025-03-28","content":"论文使用 RWKV 架构进行数据处理，引入了上下文 - WKV 机制，并进行双向信号建模。通过堆叠嵌入保留了卷积网络的强大局部感知。测试数据集上的实验结果表明，DREMnet 方法优于现有技术，处理后的现场数据更准确地反映了理论信号，提高了对地下电气结构的识别能力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250328-1.png","author":"Shuang Wang【成都理工大学】","link":"https://arxiv.org/abs/2503.22223","category":"[\"序列/强化\",\"电磁信号去噪\"]","conference_name":"COLM 2025","conference_url":"https://openreview.net/group?id=colmweb.org/COLM/2025/Conference"},{"id":25032701,"created_at":"2025-04-07T07:17:27.100908+00:00","title":"Geometry-Aware RWKV for Heterogeneous Light Field Spatial Super-Resolution","date":"2025-03-27","content":"论文基于 RWKV 设计了一种具有通道相关性的纹理转移模块和一个空间角度校正模块；同时，采用具有几何感知的 RWKV 来捕获广场的内在集合结构。实验结果表明，所提出的方法在定量和定性比较中均优于最先进的方法，同时在推理时间和内存使用方面实现了更高的效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250327-1.png","author":"Zean Chen【宁波大学】","link":"https://ieeexplore.ieee.org/abstract/document/10943155","category":"[\"图像\",\"异构空间光场超分辨率\"]","conference_name":"IEEE SLP | JCR Q2","conference_url":"https://ieeexplore.ieee.org/abstract/document/10943155"},{"id":25032601,"created_at":"2025-04-07T06:46:23+00:00","title":"RSRWKV: A Linear-Complexity 2D Attention Mechanism for Efficient Remote Sensing Vision Task","date":"2025-03-26","content":"论文提出了 RSRWKV，它具有新颖的二维 WKV 扫描机制，在保持线性复杂度的同时，连接了序列处理和二维空间推理。实现了多方向的各向同性上下文聚合。实验结果表明，在多种数据集上的分类、检测和分割任务中，RSRWKV 优于卷积神经网络和 Transformer 基线，为高分辨率遥感分析提供了一种可扩展的解决方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250326-1.png","author":"Chunshan Li【哈尔滨工业大学】","link":"https://arxiv.org/abs/2503.20382","category":"[\"图像\",\"高分辨率遥感分析\"]","conference_name":"","conference_url":""},{"id":25031901,"created_at":"2025-03-24T02:51:16+00:00","title":"RWKV-7 \"Goose\" with Expressive Dynamic State Evolution","date":"2025-03-19","content":"论文提出 RWKV-7 \"Goose\"，一种新的序列建模架构。通过引入广义 Delta Rule 等一系列优化，RWKV-7 的语言建模能力在所有开源 3B 规模模型中达到 SoTA 水平，计算效率、任务表现和模型表达力全面超越 Transformer 和过去的 RWKV-6 架构。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250319-1.png","author":"Bo Peng【元始智能】","link":"https://arxiv.org/abs/2503.14456","category":"[\"通用\",\"RWKV架构论文\"]","conference_name":"COLM 2025","conference_url":"https://colmweb.org/AcceptedPapers.html"},{"id":25031201,"created_at":"2025-03-12T09:21:14.429968+00:00","title":"CMGN: Text GNN and RWKV MLP-mixer combined with cross-feature fusion for fake news detection","date":"2025-03-12","content":"论文提出了一种新型跨特征融合网络 CMGN，结合文本图神经网络（GNN）与 RWKV MLP-mixer 用于假新闻检测。RWKV MLP-mixer 通过 MLP 层替代自注意力机制处理新闻文本以提取深层语义特征，而 Text GNN 将附加文本（如标题、地点）建模为图节点关系。跨特征融合机制动态整合多模态特征。在 LIAR、FA-KES、IFND 和 CHEF 数据集上的实验表明，CMGN 优于现有方法。焦点损失函数解决类别不平衡问题，消融实验验证了 RWKV 在特征提取中的关键作用。该模型通过图关系建模与 RWKV 的高效序列处理，推动了假新闻检测的进步。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250312-1.png","author":"Shaodong Cui【茨城大学】","link":"https://www.sciencedirect.com/science/article/abs/pii/S0925231225004837","category":"[\"语言\",\"图神经网络\",\"假新闻检测\"]","conference_name":"Neurocomputing | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0925231225004837"},{"id":25031001,"created_at":"2025-03-20T06:43:12.996771+00:00","title":"Linear attention based spatiotemporal multi graph GCN for traffic flow prediction","date":"2025-03-10","content":"论文提出 LASTGCN 模型，用于交通流预测，结合多因素融合单元（MFF-unit）动态整合气象数据、多图卷积网络捕捉空间关联，以及线性注意力机制 RWKV 模 块。RWKV 替代传统 Transformer 注意力，通过线性计算降低复杂度，高效捕获交通序列的长期依赖，兼具可并行训练与类 RNN 推理优势，适用于中短期交通管理。在真实数据集 （PeMSD） 实验中，模型精度与鲁棒性优于现有方法，长期预测表现突出，气象等外部因素集成进一步提升了预测效果。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250310-1.png","author":"Yanping Zhang【齐鲁工业大学】","link":"https://www.nature.com/articles/s41598-025-93179-y","category":"[\"序列/强化\",\"时序预测\"]","conference_name":"","conference_url":""},{"id":25030801,"created_at":"2025-03-27T05:48:08+00:00","title":"BlackGoose Rimer: Harnessing RWKV-7 as a Simple yet Superior Replacement for Transformers in Large-Scale Time Series Modeling","date":"2025-03-08","content":"论文提出将 RWKV-7 架构融入时间序列模型 Timer 中，通过其时间混合和通道混合组件提升性能。实验表明，新方法在多个数据集上实现了 1.13x 至 43.3x 的性能提升，训练时间减少 4.5 倍，且参数仅为原模型的 1/23。该成果为大规模时间序列建模提供了高效、轻量级的解决方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250308-1.png","author":"李韦乐【元始智能】","link":"https://arxiv.org/abs/2503.06121","category":"[\"序列/强化\",\"时间序列分析\"]","conference_name":"","conference_url":""},{"id":25030701,"created_at":"2025-03-20T06:27:56.76048+00:00","title":"Flare-Aware RWKV for Flare Removal","date":"2025-03-07","content":"论文提出 Flare-RWKV，一种基于 RWKV 的新型架构，用于去除图像中的镜头光晕。该方法通过结合轻量级光晕检测网络与基于 RWKV 的修复网络（利用其线性复杂度的扫描机制捕捉全局依赖关系，以及增强局部上下文感知的令牌移位机制），解决了光晕消除任务中的特定挑战。核心创新包括 Flare-Aware 特征选择 （FAFS）， 利用检测到的光晕掩码优先重建背景区域。相比 UNet 和 Transformer 变体，Flare-RWKV 在合成与真实数据集上表现更优，同时保持参数高效性，验证了 RWKV 在光晕去除任务中的有效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250307-1.png","author":"Wanying Zhang【哈工大】","link":"https://ieeexplore.ieee.org/abstract/document/10888487","category":"[\"图像\",\"图像去光晕\"]","conference_name":"ICASSP 2025 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/10888487"},{"id":25030702,"created_at":"2025-03-20T06:26:04.208949+00:00","title":"RWKVMatch: Vision RWKV-based Multi-scale Feature Matching Network for Unsupervised Deformable Medical Image Registration","date":"2025-03-07","content":"论文提出 RWKVMatch，一种基于 Vision-RWKV 的可变形医学图像配准框架，融合全局注意力与跨模态特征融合机制。通过将 RWKV 扩展为 3D 视觉模块，在保持线性计算复杂度的同时有效捕捉三维医学图像空间特征。结合弹性形变数据增强策略，该模型在脑部 MRI 数据集 （LPBA40/IXI） 上取得最优性能， LPBA40 数据集 DSC 达 0.704，雅可比负值率仅 0.154%，验证了 RWKV 在配准精度与计算效率方面的优势。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250307-4.png","author":"Zeyu Gong【华中科技大学】","link":"https://ieeexplore.ieee.org/abstract/document/10888484","category":"[\"图像\",\"医学影像配准\"]","conference_name":"ICASSP 2025 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/10888484"},{"id":25030703,"created_at":"2025-03-20T06:15:13.46782+00:00","title":"HFE-RWKV: High-Frequency Enhanced RWKV Model for Efficient Left Ventricle Segmentation in Pediatric Echocardiograms","date":"2025-03-07","content":"论文提出 HFE-RWKV， 通过改造 RWKV 的高效循环架构并增强高频特征，用于儿科超声心动图的左心室分割。该模型重新设计 RWKV 的空间混合模块以强化边界相关的高频成分，并引入空间 - 频率一致性损失函数，在保持计算效率的同时实现更精准的形状感知分割。相比 U-Mamba，HFE-RWKV 以仅 67% 的参数和 26% 计算量将 Dice 分数提升 2%，证明了 RWKV 在需要精度与资源效率的医学影像任务中的适应性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250307-2.png","author":"Zi Ye【广州软件院】","link":"https://ieeexplore.ieee.org/abstract/document/10888300","category":"[\"图像\",\"医学图像分割\"]","conference_name":"ICASSP 2025 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/10888300"},{"id":25030704,"created_at":"2025-03-20T06:12:14.95647+00:00","title":"ID-RWKV: Image Deraining RWKV\n","date":"2025-03-07","content":"论文提出 ID-RWKV， 一种基于线性复杂度 RWKV 架构的图像去雨方法，以解决 Transformer 二次计算复杂度的局限性。通过用 RWKV 块替代自注意力机制，结合局部 - 全局双向 WKV（LG-WKV） 增强空间特征建模，并设计多阶段 U 型网络渐进去雨。引入傅里叶增强模块和深浅层特征融合模块 （DSFFM） 减少背景信息丢失。实验表明，该方法在合成与真实数据集上优于主流 Transformer 模型，且参数量 （12.38M） 和计算量 （60.2G FLOPs） 更低，验证了 RWKV 在二维视觉任务中的高效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250307-3.png","author":"Yong Yang【天津工业计科学院】","link":"https://ieeexplore.ieee.org/abstract/document/10889384","category":"[\"图像\",\"图像去雨\"]","conference_name":"ICASSP 2025 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/10889384"},{"id":25030705,"created_at":"2025-03-20T06:09:09.219269+00:00","title":"Toward Comprehensive Semantic Prompt for Region Contrastive Learning Underwater Image Enhancement","date":"2025-03-07","content":"论文提出 SRCNet，一种融合语义引导和区域对比学习的水下图像增强网络。该方法创新性地设计了语义感知 RWKV 模块，在保留 RWKV 架构全局感知能力的同时，通过语义提示机制维护区域色彩一致性和结构细节。通过将 RWKV 的高效注意力机制与语义感知约束相结合，有效减少了水下不同区域间的无效像素干扰。创新的区域对比学习策略通过多视角负样本利用，增强了退化敏感特征的学习能力。实验结果表明该方法在色彩还原和细节恢复方面优于现有最优方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250307-5.png","author":"Xiao Wang【武汉科技大学】","link":"https://ieeexplore.ieee.org/abstract/document/10888780","category":"[\"图像\",\"水下图像增强\"]","conference_name":"ICASSP 2025 | CCF B","conference_url":"https://ieeexplore.ieee.org/abstract/document/10888780"},{"id":68,"created_at":"2025-03-12T09:16:06.838959+00:00","title":"PathRWKV: Enabling Whole Slide Prediction with Recurrent-Transformer","date":"2025-03-05","content":"论文提出 PathRWKV，一种用于计算病理学中全玻片图像（WSI）分析的新型循环 - Transformer 混合模型。针对可变切片规模处理、模型复杂性和训练 - 推理平衡的挑战，PathRWKV 融合动态循环结构以实现全切片建模，并采用 RWKV 的线性注意力机制降低计算成本并缓解过拟合。多任务学习联合优化异构临床指标以提升训练效率，异步推理设计则支持预测阶段对所有切片的序列化处理。在七大 WSI 数据集评估中，PathRWKV 在癌症分型、转移检测及生存预测等任务中表现优于现有方法，展现了其在病理学应用中的卓越泛化性和可扩展性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250305-1.png","author":"Sicheng Chen【新加坡国立大学】","link":"https://arxiv.org/abs/2503.03199","category":"[\"图像\",\"全玻片图像（WSI）分析\"]","conference_name":"","conference_url":""},{"id":67,"created_at":"2025-03-04T06:27:07.127045+00:00","title":"Delta-WKV: A Novel Meta-in-Context Learner for MRI Super-Resolution","date":"2025-02-28","content":"论文提出 Delta-WKV，一种新型线性 Transformer 模型，用于 MRI 超分辨率。该模型结合元上下文学习（MiCL）和 Delta 规则，在推理过程中动态调整权重，以高效识别局部和全局模式。受 RWKV 启发，Delta-WKV 采用四向扫描机制，并利用通道混合网络替代传统 MLP，增强长程依赖捕获能力并保留高频细节。在 IXI 和 fastMRI 数据集上的实验表明，Delta-WKV 的 PSNR/SSIM 指标优于现有方法，且训练和推理速度比 SwinIR 和 MambaIR 快 15%，展现了其在临床大规模应用中的高效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250228-1.png","author":"卢荣昌【青海大学】","link":"https://arxiv.org/abs/2502.20852","category":"[\"图像\",\"磁共振成像超分辨率 \"]","conference_name":"MICCAI 2025 | CCF B","conference_url":"https://arxiv.org/abs/2502.20852"},{"id":64,"created_at":"2025-03-03T07:14:50.229416+00:00","title":"TabulaTime: A Novel Multimodal Deep Learning Framework for Advancing Acute Coronary Syndrome Prediction through Environmental and Clinical Data Integration","date":"2025-02-24","content":"论文提出 TabulaTime，一种新型多模态深度学习框架，整合临床和环境时序数据以提升急性冠状动脉综合征（ACS）预测。核心创新包括 PatchRWKV 模块，该模块结合循环神经网络（RNN）和注意力机制，以线性计算复杂度高效提取时序特征，在捕捉时序依赖上优于 Transformer、LSTM 等先进模型。实验显示，其准确率较传统方法提升 20.5%，突显了整合空气污染数据的重要性。框架通过注意力机制增强可解释性，识别出收缩压、PM₁₀等关键预测因子。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250224-1.png","author":" Liangxiu Han【英国曼彻斯特都会大学】","link":"https://arxiv.org/abs/2502.17049","category":"[\"序列/强化\",\"急性冠状动脉综合征预测\"]","conference_name":"","conference_url":""},{"id":63,"created_at":"2025-02-27T09:52:46.609711+00:00","title":"Rwkv-vg: visual grounding with RWKV-driven encoder-decoder framework","date":"2025-02-21","content":"论文提出 RWKV-VG，一种完全基于 RWKV 架构的视觉定位框架。不同于传统的 CNN 或 Transformer 方法，RWKV-VG 利用 RWKV 结合 RNN 的顺序建模与 Transformer 注意力的混合设计，高效建模模态内和跨模态交互。该框架包含 RWKV 驱动的视觉 / 语言编码器、跨模态解码器及可学习的 [REG] 令牌用于边界框回归。在 ReferItGame 和 RefCOCO 等基准测试中，其性能超越 TransVG 等 Transformer 方法，精度更高且收敛更快。消融实验验证了 RWKV 模块和 [REG] 令牌位置的关键作用。该工作证实了 RWKV 在视觉 - 语言任务中的竞争力，兼具高效计算与高精度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250221-1.png","author":" Fanding Li【中国科学技术大学】","link":"https://link.springer.com/article/10.1007/s00530-025-01720-w","category":"[\"图像\",\"视觉-语言\"]","conference_name":"Multimedia Systems | JCR Q2","conference_url":"https://link.springer.com/article/10.1007/s00530-025-01720-w"},{"id":59,"created_at":"2025-02-21T09:44:00.84504+00:00","title":"Substation equipment non-rigid defect detection via receptance weighted key value-based causality-aware networks","date":"2025-02-13","content":"论文提出了一种基于 RWKV 架构的因果感知设备缺陷检测框架，以解决变电站设备中的非刚性缺陷检测和长尾分布问题。RWKV 架构具有全局感受野，可增强缺陷特征提取能力。它与框架中的其他模块相结合。实验表明，该框架优于基线方法，验证了其有效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250213-1.png","author":"Jie Zhang【中科院合肥机械所】","link":"https://link.springer.com/article/10.1007/s11760-025-03852-y","category":"[\"图像\",\"设备缺陷检测\"]","conference_name":"Signal, Image and Video Processing","conference_url":"https://link.springer.com/article/10.1007/s11760-025-03852-y"},{"id":66,"created_at":"2025-03-04T06:23:19.349274+00:00","title":"Linear Attention Modeling for Learned Image Compression","date":"2025-02-09","content":"论文提出 LALIC，一种基于线性注意力的学习图像压缩框架，利用 Bi-RWKV 块实现高效紧凑的特征提取。通过结合 Spatial Mix 和 Channel Mix 模块以及基于卷积的 Omni-Shift 机制，LALIC 将 RWKV 的线性复杂度注意力适配至二维潜在表示。此外，新型 RWKV-SCCTX 上下文模型增强了空间 - 通道依赖建模，改进了熵编码性能。实验表明，该方法在 Kodak、Tecnick 和 CLIC 数据集上的 BD-rate 分别超越 VTM-9.1 达 - 14.84%、-15.20% 和 - 17.32%，性能达到前沿水平。该研究首次将 RWKV 的双向注意力引入图像压缩，在全局上下文建模与低计算开销间取得平衡。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250209-2.png","author":"Li Song【上海交通大学】","link":"https://arxiv.org/abs/2502.05741","category":"[\"图像\",\"图像压缩\"]","conference_name":"CVPR 2025 | CCF A","conference_url":"https://cvpr.thecvf.com/virtual/2025/poster/33192"},{"id":56,"created_at":"2025-02-12T06:13:31.694355+00:00","title":"Training Language Models for Social Deduction with Multi-Agent Reinforcement Learning","date":"2025-02-09","content":"论文提出利用多智能体强化学习训练语言模型，使其在无需人类示范的社交推理游戏中实现自然语言沟通。通过结合“听”（从讨论中推理内鬼身份）和“说”（奖励能改变他人观点的信息），该方法采用 RWKV 模型——一种基于线性注意力的循环架构，以高效处理长游戏序列并降低计算负担。实验表明，基于 RWKV 的智能体胜率是标准强化学习方法的两倍，并展现出基于证据指控等类人策略。 RWKV 的选择解决了扩展性和长上下文处理的挑战，对实时多智能体交互至关重要。\r\n\r\n","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/RWKV-AMONG-US-43.png","author":"Bidipta Sarkar【斯坦福大学】","link":"https://arxiv.org/abs/2502.06060","category":"[\"序列/强化\",\"多智能体强化学习\"]","conference_name":"AAMAS 2025 | CCF B","conference_url":"https://dl.acm.org/doi/10.5555/3709347.3743819"},{"id":55,"created_at":"2025-02-07T07:25:14.685825+00:00","title":"RWKV-UI: UI Understanding with Enhanced Perception and Reasoning","date":"2025-02-06","content":"论文提出 RWKV-UI，一种基于 RWKV 架构的视觉语言模型，专为高分辨率用户界面（UI）理解设计。针对现有视觉语言模型在高分辨率 UI 图像处理中的信息丢失和推理能力不足，该模型集成三种视觉编码器（SIGLIP、DINO、SAM），采用分块编码策略处理 4096×4096 图像并保留细节。结合 RWKV 高效的 RNN 结构，模型引入布局检测和思维链（CoT）视觉提示，增强空间推理和多步交互预测能力。实验表明其在 UI 任务中表现卓越，在动作定位和元素识别等任务上优于更大规模模型，凸显了 RWKV 在多模态场景中的适应性和高效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250206-1.png","author":"杨家希（侯皓文实习生）","link":"https://arxiv.org/abs/2502.03971","category":"[\"图像\",\"UI设计\"]","conference_name":"","conference_url":""},{"id":54,"created_at":"2025-02-05T06:08:37.467134+00:00","title":"Exploring Linear Attention Alternative for Single Image Super-Resolution","date":"2025-02-01","content":"论文提出 OmniRWKVSR 模型用于单图像超分辨率，结合 RWKV 架构与新型特征提取技术（VRSM 和 VRCM），以解决计算复杂性和重建质量问题。通过利用 RWKV 的线性计算效率及RNN-Transformer 混合优势，该模型避免了二次注意力计算成本，同时增强多尺度特征捕捉。实验结果表明其性能优于 MambaIR 和 SwinIR，在4倍超分辨率任务中 PSNR 提升 0.26%、SSIM 提升 0.16%，且训练速度加快 15%。研究突显了 RWKV 在平衡效率与图像恢复质量（尤其在遥感应用）中的有效性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250201-1.png","author":"卢荣昌【青海大学】","link":"https://arxiv.org/abs/2502.00404","category":"[\"图像\",\"单图像超分辨率\"]","conference_name":"IJCNN 2025 | CCF C","conference_url":"https://2025.ijcnn.org/"},{"id":37,"created_at":"2024-12-22T04:52:49.774628+00:00","title":"RWKV-Lite: Deeply Compressed RWKV for Resource-Constrained Devices","date":"2025-01-31","content":"论文提出 RWKV-Lite，旨在解决资源受限设备运行 RWKV 模型的难题。其采用低秩近似、稀疏性预测和聚类头等技术，将 RWKV 模型压缩了 4.95-3.8 倍，而准确度仅损失 2.95pp。RWKV-edge 为在边缘设备部署 RWKV 模型提供有效方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241214-1.png\r\n","author":"Wonkyo Choe,Yangfeng Ji【弗吉尼亚大学】","link":"https://arxiv.org/abs/2412.10856","category":"[\"通用\",\"边缘设备部署\"]","conference_name":"","conference_url":""},{"id":57,"created_at":"2025-02-13T02:20:56.222172+00:00","title":"ARWKV: Pretrain is not what we need, an RNN-Attention-Based Language Model Born from Transformer","date":"2025-01-26","content":"论文提出 ARWKV，一种基于 RWKV 架构的 RNN - 注意力语言模型，旨在超越 Transformer 的表达能力和状态跟踪能力。通过从 Qwen2.5 等 Transformer 模型蒸馏知识到 RNN 中，ARWKV 用 RWKV-7 的时间混合模块替代自注意力机制，实现在有限资源（如单块 A100 GPU 训练 7B 模型）下的高效训练。方法包含注意力对齐、知识蒸馏和监督微调三个阶段。评估显示模型在基准测试中表现良好，但师生模型规模差异可能导致性能下降。该工作融合了 Transformer 的效率与 RNN 优势，凸显了 RWKV 在混合架构中的潜力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250126-1.png","author":"林玥煜【元始智能】","link":"https://arxiv.org/abs/2501.15570","category":"[\"通用\",\"蒸馏\"]","conference_name":"","conference_url":""},{"id":53,"created_at":"2025-01-24T03:03:08.046438+00:00","title":"Rate-Aware Learned Speech Compression","date":"2025-01-21","content":"论文提出了一种基于通道感知熵模型的学习语音压缩方案，该方案通过替换传统的量化器来增强率失真性能。它利用多尺度卷积和 RWKV 混合块来提高编码器和解码器的表示能力。实验结果表明，与现有编解码器相比，提出的方法在比特率节省和声学质量指标方面取得了显著改善。这一研究成果对于解决实时通信中的语音压缩问题具有重要意义，并且为未来的研究提供了新的思路和方向。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250121-1.png","author":"Jun Xu【上海交通大学】","link":"https://arxiv.org/abs/2501.11999","category":"[\"音频/语音\",\"语音压缩\"]","conference_name":"","conference_url":""},{"id":62,"created_at":"2025-02-21T10:06:55.703116+00:00","title":"Learnable Sparsification of Die-to-Die Communication via Spike-Based Encoding","date":"2025-01-15","content":"论文提出了 SNAP，一种结合了脉冲神经网络（SNNs）和人工神经网络（ANNs）的混合神经网络架构。为了评估 SNAP，论文将 RWKV 作为代表性的语言模型架构进行集成。实验表明，SNAP 优于传统的 SNN 和非脉冲模型，实现了高达 5.3 倍的能源效率提升和 15.2 倍的推理延迟降低，凸显了其在大规模人工智能系统中的潜力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250115-1.png","author":"Ruijie Zhu【美国加利福尼亚大学圣克鲁兹分校】","link":"https://arxiv.org/abs/2501.08645","category":"[\"通用\",\"脉冲神经网络\",\"人工神经网络\"]","conference_name":"","conference_url":""},{"id":52,"created_at":"2025-01-17T03:05:02.716652+00:00","title":"RWKV-UNet: Improving UNet with Long-Range Cooperation for Effective Medical Image Segmentation","date":"2025-01-14","content":"论文提出 RWKV-UNet，它将 RWKV 结构融入 U-Net 用于医学图像分割。通过 IR-RWKV 模块增强长距离依赖捕获能力，结合 CCM 模块改善跳跃连接。实验表明，在多个数据集上取得 SOTA 性能，其变体平衡了性能和效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250114-1.png\r\n","author":"蒋峻涛【浙江大学】","link":"https://arxiv.org/abs/2501.08458","category":"[\"图像\",\"医学图像分割\"]","conference_name":"","conference_url":""},{"id":65,"created_at":"2025-03-03T07:22:13.814212+00:00","title":"ChemRB: a novel generative model based on bidirectional molecular ring constraints","date":"2025-01-10","content":"论文提出了一种名为ChemRB的新型生成模型，用于药物发现中的分子设计，通过双向分子环约束解决现有单向编码器的局限性。通过整合RWKV机制，ChemRB将RNN的线性计算效率与Transformer的上下文感知相结合，有效捕获SMILES序列中的长程依赖性。该模型引入两个预训练任务——环级特征预测和全局跨度闭合预测，以提升分子有效性，尤其针对复杂环系统。实验结果表明，ChemRB在生成有效、唯一且新颖的分子方面表现卓越，优于基准数据集上的先进模型。此外，其在EGFR抑制剂重新设计中的应用验证了其实用性，展现出高结合亲和力与结构保真度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250110-1.png","author":" 辜丽川【安徽农业大学】","link":"https://jsnu.magtech.com.cn/CN/10.15983/j.cnki.jsnu.2025005","category":"[\"序列/强化\",\"分子生成模型\"]","conference_name":"Journal of Shaanxi Normal University","conference_url":"https://jsnu.magtech.com.cn/CN/10.15983/j.cnki.jsnu.2025005"},{"id":60,"created_at":"2025-02-21T09:55:53.654692+00:00","title":"Explore Activation Sparsity in Recurrent LLMs for Energy-Efficient Neuromorphic Computing","date":"2025-01-09","content":"论文提出了一种低成本、无需训练的算法，用于稀疏循环大语言模型（R-LLMs）的激活，以实现高效能的神经形态计算。论文以 RWKV 为例展示了该方法的有效性。通过在 RWKV 中添加阈值函数，平均激活稀疏度得以提升。硬件模拟显示出显著的节能和延迟改善，并且该方法还可以扩展到其他模型。\n","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250109-1.png","author":"Ivan Knunyants【荷兰埃因霍芬理工大学】","link":"https://arxiv.org/abs/2501.16337","category":"[\"通用\",\"激活稀疏性\"]","conference_name":"AICAS 2025","conference_url":"https://arxiv.org/abs/2501.16337"},{"id":50,"created_at":"2025-01-13T08:46:28.860191+00:00","title":"Reducing Cross-Sensor Domain Gaps in Tactile Sensing via Few-Sample-Driven Style-to-Content Unsupervised Domain Adaptation","date":"2025-01-05","content":"FSSC 是一种全新的 few-sample-driven style-to-content 无监督域适应方法。它采用基于 RWKV 架构的设计来应对跨传感器域适应的难题，例如传感器差异导致的域差距。借助 GLAB 层、FST 模块等重要组件，它达成了有效减少触觉传感跨传感器域差距的目标。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20250105-1.png\r\n","author":"Xingshuo Jing【南京东南大学】","link":"https://www.mdpi.com/1424-8220/25/1/256","category":"[\"3D/视频\",\"无监督领域\",\"触觉传达\"]","conference_name":"Sensors 2025 | JCR Q2","conference_url":"https://www.mdpi.com/1424-8220/25/1/256"},{"id":51,"created_at":"2025-01-13T08:52:06.658861+00:00","title":"Efficient Relational Context Perception for Knowledge Graph Completion","date":"2024-12-31","content":"论文提出一种受 RWKV 启发的知识图谱补全新方法，整合了三元接受感知（TRP）架构和塔克分解模块。TRP 通过时间和通道混合模块有效建模顺序信息以学习动态嵌入，实验证明该方法优于现有模型。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241231-1.png\r\n","author":"Wenkai Tu【武汉大学】","link":"https://arxiv.org/abs/2501.00397","category":"[\"语言\",\"知识图谱补全\"]","conference_name":"","conference_url":""},{"id":43,"created_at":"2025-01-06T03:27:27.450323+00:00","title":"Injecting Explainability and Lightweight Design into Weakly Supervised Video Anomaly Detection Systems","date":"2024-12-28","content":"论文介绍了一种名为 TCVADS 的视频异常检测系统。该系统采用两个阶段的运行模式。在第一阶段，系统使用增强的 RWKV 模块来进行高效的时间序列分析。通过结合知识蒸馏和跨模态学习技术，TCVADS 在性能上优于现有的方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241228-1.png\r\n","author":"Wen-Dong Jiang【台湾淡江大学】","link":"https://arxiv.org/abs/2412.20201","category":"[\"3D/视频\",\"异常检测\"]","conference_name":"","conference_url":""},{"id":40,"created_at":"2024-12-31T09:19:52.733145+00:00","title":"StyleRWKV: High-Quality and High-Efficiency Style Transfer with RWKV-like Architecture","date":"2024-12-27","content":"StyleRWKV 是一种新的视频风格迁移方法。它采用受 RWKV 启发的架构来解决先前方法的缺点，如高计算复杂度等问题，并通过 Re-WKV 注意力机制等关键要素实现了高效且高质量的风格迁移。实验证明，StyleRWKV 在风格化质量、模型复杂性和推理效率方面都优于最先进的方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241227-1.png\r\n","author":"Miaomiao Dai【上海交通大学】","link":"https://arxiv.org/abs/2412.19535","category":"[\"图像\",\"视频风格迁移\"]","conference_name":"","conference_url":""},{"id":39,"created_at":"2024-12-26T03:42:16.485131+00:00","title":"L3TC: Leveraging RWKV for Learned Lossless Low-Complexity Text Compression","date":"2024-12-21","content":"L3TC 是一种新的学习型无损低复杂度文本压缩方法。它采用 RWKV 作为基础架构，并提出了异常值感知分词器和高秩重参数化策略，在不增加推理复杂度的前提下提升了模型学习能力。实验结果验证，L3TC 相较于 gzip 压缩节省了 48% bit。与同性能的其他学习型压缩器相比，L3TC 模型的参数减少 50 倍，且实时解码速度高达 MB/s 量级。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241221-1.png\r\n","author":"程正雪【上海交通大学】","link":"https://arxiv.org/abs/2412.16642","category":"[\"语言\",\"文本压缩方法\"]","conference_name":"AAAI 2025 | CCF A","conference_url":"https://ojs.aaai.org/index.php/AAAI/article/view/33446"},{"id":41,"created_at":"2024-12-31T09:46:46.297048+00:00","title":"A Survey of RWKV","date":"2024-12-19","content":"论文深入探究 RWKV，从架构方面剖析融合机制，从应用上梳理多领域成果，分析 RWKV 长序列和多模态等挑战，探讨其处理能力、安全、硬件等未来方向。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241219-1.png\r\n","author":"Zhiyuan Li【吉林大学】","link":"https://arxiv.org/abs/2412.14847","category":"[\"通用\",\"RWKV发展概览\"]","conference_name":"","conference_url":""},{"id":36,"created_at":"2024-12-22T04:45:10.001103+00:00","title":"PCF-RWKV: Product Carbon Footprint Estimation System Based on Large Language Model","date":"2024-12-18","content":"PCF-RWKV 是一款基于 RWKV 架构的产品碳足迹评估模型，具有多个堆叠残差块和三个任务专用的低等级适配器 (LoRA)。通过集成 Multi-Agents 技术，PCF-RWKV 可自动构建生产流程的 LCI，将生产流程与排放因子相匹配以计算碳足迹，从而提高企业碳足迹评估的效率和安全性，并解决传统方法的局限性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241218-1.png\r\n","author":"Zhen Li,Peihao Tang【清华大学】","link":"https://www.preprints.org/manuscript/202412.1705/v1","category":"[\"语言\",\"碳足迹评估\"]","conference_name":"Sustainability 2025 | JCR Q2","conference_url":"https://www.mdpi.com/2071-1050/17/3/1321"},{"id":42,"created_at":"2024-12-31T10:11:24.48023+00:00","title":"Linear Attention Based Channel Estimation Scheme for V2X Communications","date":"2024-12-13","content":"RWKV-DPA 是一种用于车联网 （V2X） 通信的创新信道估计方案，它使用线性注意力机制捕捉输入符号之间的相关性，并结合数据导频辅助（DPA）估计来动态跟踪信道变化，从而提升信道估计性能。仿真结果表明，RWKV-DPA 评估器在误码率（BER）和归一化均方误差（NMSE）方面性能上优于其他基于深度学习的估计器，同时保持了较低的计算复杂度。在涉及高速运动的场景中，RWKV-DPA 估计器表现出卓越的性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241213-1.png\r\n","author":"Qianyao Fu【中国传媒大学】","link":"https://ieeexplore.ieee.org/document/10779439","category":"[\"序列/强化\",\"信道估计\"]","conference_name":"ICCIS 2024","conference_url":"https://ieeexplore.ieee.org/document/10779439"},{"id":35,"created_at":"2024-12-16T06:50:49.982016+00:00","title":"Exploring Real&Synthetic Dataset and Linear Attention in Image Restoration","date":"2024-12-11","content":"论文提出 RWKV-IR：基于 RWKV 的新颖图像恢复模型，支持全局和局部感受野。RWKV-IR 在 Urban100 x4 的实验上比 SwinIR 好 0.08dB，比 MambaIR 好 0.03dB，展示了 RWKV-IR 的先进图像恢复能力和快速收敛能力。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241211-1.jpg","author":"Yuzhen Du【上海交通大学】","link":"https://arxiv.org/abs/2412.03814","category":"[\"图像\",\"图像恢复模型\"]","conference_name":"","conference_url":""},{"id":44,"created_at":"2025-01-06T03:30:49.876225+00:00","title":"Voice dialog system based on RWKV model","date":"2024-11-28","content":"论文提出开发一个面向老年人的智能语音对话系统，采用经 LoRA 微调的 RWKV 模型。实验结果表明它提高了答案的流畅性和合理性，在老年护理方面有应用潜力，未来工作会对模型进行优化。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241128-1.png\r\n","author":"Tianxi Zhao【东北大学】","link":"https://ieeexplore.ieee.org/abstract/document/10762107","category":"[\"音频/语音\",\"语音对话系统\"]","conference_name":"EIT 2024","conference_url":"https://ieeexplore.ieee.org/abstract/document/10762107"},{"id":34,"created_at":"2024-11-19T03:43:07.515378+00:00","title":"DFT: A Dual-branch Framework of Fluctuation and Trend for Stock Price Prediction","date":"2024-11-09","content":"论文提出了用于股票价格预测的 DFT 框架，将股票分解为趋势和波动双分支，在波动分支中引入 RWKV 模型有效建模时间相关性。DFT  在多个股票数据集上展现出卓越性能，为股票价格预测提供了更有效的途径。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241109-1.png\r\n","author":"Chengqi Dong, Zhiyuan Cao【吉林大学】","link":"https://arxiv.org/abs/2411.06065","category":"[\"序列/强化\",\"股票预测模型\"]","conference_name":"","conference_url":""},{"id":33,"created_at":"2024-11-15T09:29:36.253943+00:00","title":"Video RWKV: Video Action Recognition Based RWKV","date":"2024-11-08","content":"论文提出了用于视频理解的 LSTM-CrossRWKV（LCR）框架，该框架将 RWKV 引入视频领域，通过创新的 Cross RWKV 门和 LSTM 循环执行机制，有效捕捉时空特征，利用边缘信息减少冗余。在多个数据集上表现出色，为视频分析提供了高效解决方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241108-1.png\r\n","author":" Zhuowen Yin【沈阳东北大学】","link":"https://arxiv.org/abs/2411.05636","category":"[\"3D/视频\",\"视频理解\"]","conference_name":"","conference_url":""},{"id":32,"created_at":"2024-11-08T06:14:47.347998+00:00","title":"From Explicit Rules to Implicit Reasoning in an Interpretable Violence Monitoring System","date":"2024-10-29","content":"论文提出了新的弱监督暴力监控框架 RuleVM。RuleVM 使用 RWKV 架构作为其轻量级事件序列模块，并使用相对距离代替特征相似性。这种机制使模型更加轻量级，因为它只考虑画面帧间的相对距离，无需评估高维特征相似性，从而有效降低计算复杂性，有助于降低计算成本，提高训练和推理效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241029-1.png\r\n","author":"Wen-Dong Jiang【台湾淡江大学】","link":"https://arxiv.org/abs/2410.21991","category":"[\"3D/视频\",\"暴力监控\",\"智慧城市\"]","conference_name":"","conference_url":""},{"id":31,"created_at":"2024-11-04T07:39:57.598217+00:00","title":"Modern Sequence Models in Context of Multi-Agent Reinforcement Learning","date":"2024-10-28","content":"论文提出 MARWKV（Multi-Agent-RWKV）架构，以实现多智能体强化学习（MARL）。实验结果表明： MARWKV 在 MARL 基准测试中的表现与 MAT（Transformer） 相当，并且具有类似的少样本学习能力，且在 Agent 数量较多时拥有更好的计算效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241028-1.png\r\n","author":"Jenish Thapa【奥地利约翰内斯开普勒林茨大学】","link":"https://epub.jku.at/obvulihs/content/titleinfo/10580112","category":"[\"序列/强化\",\"Multi-Agent\"]","conference_name":"","conference_url":""},{"id":46,"created_at":"2025-01-06T03:38:02.346853+00:00","title":"AutoGMM-RWKV: A Detecting Scheme Based on Attention Mechanisms Against Selective Forwarding Attacks in Wireless Sensor Networks","date":"2024-10-23","content":"论文提出了 AutoGMM-RWKV 用于检测无线传感器网络中的选择性转发攻击。它聚焦于节点单轮转发率时间序列，通过将自编码器、高斯混合模型和 K - 均值与 RWKV 相结合，提高了检测精度。模拟结果显示误检率和漏检率较低，提供了一个可靠的解决方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241023-1.png\r\n","author":"Shihao Xie【成都电子科技大学】","link":"https://ieeexplore.ieee.org/abstract/document/10729884","category":"[\"序列/强化\",\"攻击检测\"]","conference_name":"IEEE IoT | 中科院 1 区","conference_url":"https://ieeexplore.ieee.org/document/10729884"},{"id":30,"created_at":"2024-11-04T07:18:48.552315+00:00","title":"MATCC: A Novel Approach for Robust Stock Price Prediction Incorporating Market Trends and Cross-time Correlations","date":"2024-10-21","content":"论文提出了基于 RWKV 架构的 “MATCC” 框架，可用于解决股价预测中的市场趋势和跨时间关联问题。实验表明，MATCC 在预测准确性和投资组合表现方面显著优于其他模型，验证了股价趋势和跨时间关联在股票分析中的重要性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241021-1.png\r\n","author":"Zhiyuan Cao, Jiayu Xu【吉林大学&上海锐天】","link":"https://dl.acm.org/doi/abs/10.1145/3627673.3679715","category":"[\"序列/强化\",\"股票预测模型\"]","conference_name":"CIKM 2024 | CCF B","conference_url":"https://dl.acm.org/doi/10.1145/3627673.3679715"},{"id":28,"created_at":"2024-10-16T03:00:23.140692+00:00","title":"VisualRWKV-HD and UHD: Advancing High-Resolution Processing for Visual Language Models","date":"2024-10-15","content":"VisualRWKV 的高分辨率版本，VisualRWKV-HD 支持 1024x1024 的图片，VisualRWKV-UHD 支持 4096x4096 的图片。这两个新模型更加适合文档理解和 Text-rich 的任务。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241015-1.png\r\n","author":"侯皓文【元始智能】","link":"https://arxiv.org/abs/2410.11665","category":"[\"图像\",\"高分辨率图像\"]","conference_name":"","conference_url":""},{"id":29,"created_at":"2024-11-04T07:02:22.118308+00:00","title":"AttnInput: Revolutionizing Pinyin Input with Context-Aware RWKV Language Models","date":"2024-10-13","content":"AttnInput 利用 RWKV 语言模型的优势来增强拼音输入法，通过轻量级端侧网络将拼音信息直接集成到 RWKV 的内部状态中，有效解决了之前基于LLM的输入法所面临的语义不连续问题。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20241013-1.png\r\n","author":"桂之瑜【中国科学技术大学】","link":"https://openreview.net/forum?id=9OxTqscUwi","category":"[\"语言\",\"拼音输入法\",\"端侧应用\"]","conference_name":"","conference_url":""},{"id":1,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"OccRWKV: Rethinking Efficient 3D Semantic Occupancy Prediction with Linear Complexity","date":"2024-09-26","content":"论文提出了 “OccRWKV” ：基于 RWKV 的 3D 语义占用预测，可用于自动驾驶、具身智能等领域。实验表明，OccRWKV 在 SemanticKITTI 数据集达到 25.1 的 mIoU，比最佳基线 Co-Occ 快 20 倍，使其适合在机器人上实时部署，以增强自主导航效率。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240926-1.jpg\r\n","author":"Junming Wang,Wei Yin 地平线&香港大学","link":"https://www.arxiv.org/abs/2409.19987","category":"[\"3D/视频\",\"3D语义分割\"]","conference_name":"ICRA 2025 | CCF B","conference_url":"https://jmwang0117.github.io/OccRWKV/"},{"id":38,"created_at":"2024-12-26T03:19:29.056683+00:00","title":"MiSS: Revisiting the Trade-off in LoRA with an Efficient Shard-Sharing Structure","date":"2024-09-19","content":"论文提出MiSS框架，通过分片预训练权重并共享可训练矩阵，解决LoRA收敛慢的问题，提升大语言模型的高效微调。MiSS包含高效计算的Bone结构和解决共线性更新的Bat方法。实验表明，MiSS在自然语言理解和生成任务中优于LoRA及其变体，尤其在RWKV-7B和RWKV6-3B模型上，MiSS以相同或更少参数实现更快收敛和更强泛化。该框架降低了计算和内存开销，为RWKV等架构提供了参数共享和非线性更新的高效适配方案。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240919-1.png","author":"Jiale Kang【元始智能】","link":"https://arxiv.org/abs/2409.15371","category":"[\"通用\",\"RWKV微调方法\"]","conference_name":"","conference_url":""},{"id":2,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Multi-scale RWKV with 2-dimensional temporal convolutional network for short-term photovoltaic power forecasting","date":"2024-09-06","content":"论文提出了“多尺度 RWKV 二维时间卷积网络”（MSRWKV-2DTCN），将 FFT 和 2D TCN 与 RWKV 架构相结合并应用于光伏发电预测。研究证实，对比其他光伏发电功率预测模型，MSRWKV-2DTCN 在短期光伏发电功率预测方面具有更高的准确性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240906-1.png\r\n","author":"Jianhua Hao【山东师范大学】","link":"https://www.sciencedirect.com/science/article/abs/pii/S0360544224028433","category":"[\"序列/强化\",\"光伏发电预测\"]","conference_name":"Energy | JCR Q1","conference_url":"https://www.sciencedirect.com/science/article/abs/pii/S0360544224028433"},{"id":3,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Experimentation in Content Moderation using RWKV","date":"2024-09-05","content":"论文提出了 Mod-RWKV ，研究了 RWKV 模型在内容审核方面的效果。通过使用一个包含图像、视频、声音和文本的数据集对 RWKV 模型进行 SFT 微调，使其适用于各种内容的审查场景。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240905-1.png\r\n","author":"Rohan Dutta【Meta】","link":"https://arxiv.org/abs/2409.03939","category":"[\"语言\",\"内容审核\"]","conference_name":"","conference_url":""},{"id":5,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation","date":"2024-08-30","content":"论文提出了时间交互式建模（Temporal and Interactive Modeling，TIM），将 RWKV 模型应用于生成人类之间的交互动作。实验数据显示：TIM 仅使用 InterGen 数据集中 32% 的可训练参数，就取得了 SOTA 效果。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240830-1.png\r\n","author":"Yabiao Wang, Shuo Wang【浙江大学&腾讯优图】","link":"https://arxiv.org/abs/2408.17135","category":"[\"3D/视频\",\"动作生成\",\"人类交互建模\"]","conference_name":"CVPR 2025 | CCF A","conference_url":"https://cvpr.thecvf.com/virtual/2025/poster/32570"},{"id":4,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"OnlySportsLM: Optimizing Sports-Domain Language Models with SOTA Performance under Billion Parameter","date":"2024-08-30","content":"论文提出了 OnlySportsLM ：针对体育运动相关任务优化 RWKV-v6 架构，并训练了一个196M 的 OnlySportsLM 。Benchmark 显示，与 sota 体育运动模型相比 OnlySportsLM 的精度提高了 37.62%/34.08% 。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240830-2.png\r\n","author":"Zexin Chen【纽约大学】","link":"https://arxiv.org/abs/2409.00286","category":"[\"语言\",\"体育领域模型\"]","conference_name":"","conference_url":""},{"id":47,"created_at":"2025-01-06T03:47:05.354203+00:00","title":"Revenge of the Fallen? Recurrent Models Match Transformers at Predicting Human Language Comprehension Metrics","date":"2024-08-26","content":"论文提出，在语言任务中Transformer 一直占据主导地位，但近期 RWKV 等循环模型出现。本文表明像 RWKV 这样的当代循环模型在模拟人类语言理解方面能够与 Transformer 相媲美甚至超越它们，开启了新的研究方向。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240826-1.png\r\n","author":"James A. Michaelova【加州大学】","link":"https://arxiv.org/abs/2404.19178","category":"[\"通用\",\"循环神经网络\"]","conference_name":"COLM 2024","conference_url":"https://openreview.net/group?id=colmweb.org/COLM/2024/Conference#tab-accept"},{"id":6,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Why Perturbing Symbolic Music is Necessary: Fitting the Distribution of Never-used Notes through a Joint Probabilistic Diffusion Model","date":"2024-08-04","content":"论文提出了 Music-Diff 架构，该架构引入了 Joint Semantic Pre-training 方法来执行多变量扰动，并引入了多分支降噪器 “Symb-RWKV” 模型来恢复联合分布的噪声（通过 Pareto 优化来适应多个噪声目标）。实验表明，与语言模型相比，在音符和语义层面进行扰动的联合概率扩散模型可以提供更多样本多样性和组成规律性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240804-1.png\r\n","author":"Shipei Liu【大连理工大学】","link":"https://arxiv.org/abs/2408.01950","category":"[\"音频/语音\",\"音乐生成\",\"概率扩散模型\"]","conference_name":"","conference_url":""},{"id":7,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Optimizing Robotic Manipulation with Decision-RWKV: A Recurrent Sequence Modeling Approach for Lifelong Learning","date":"2024-07-23","content":"南方科技大学的研究团队提出了 Decision-RWKV (DRWKV) 模型，并将经验回放（experience replay）的概念与 Decision-RWKV 模型相结合，设计出适合机器人的终身学习算法实验结果显示： DRWKV 模型在单任务测试和终身学习场景中拥有先进的性能。与此同时， Decision-RWKV 相比 DT（Decision-Transformer）显著地减少了推理时间和内存占用，使其成为现实应用（尤其是机器人领域）的更佳选择。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240723-1.png\r\n","author":"董榆健【南方科技大学】","link":"https://arxiv.org/abs/2407.16306","category":"[\"序列/强化\",\"终身学习\"]","conference_name":"","conference_url":""},{"id":8,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"BSBP-RWKV: Background Suppression with Boundary Preservation for Efficient Medical Image Segmentation","date":"2024-07-21","content":"BSBP-RWKV 是中科大研究团队提出来的一种高效的医学图像分割模型，该模型结合了 Perona-Malik Diffusion (PMD）噪声抑制方面的优点和 RWKV 高效模型结构的优点，有效地抑制图像噪声干扰并保留病变区域的边界细节。与 SOTA 相比，BSBP-RWKV 的复杂度降低了 5.8 倍，在推理过程中，每张 1024×1024 图像的显存使用量减少了 62.7% 以上。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240721-1.png\r\n","author":"Xudong Zhou【合肥中国科学技术大学】","link":"https://openreview.net/pdf?id=ULD5RCk0oo","category":"[\"图像\",\"医学图像分割\",\"噪声抑制\"]","conference_name":"ACM MM 2024 | CCF A","conference_url":"https://dl.acm.org/doi/abs/10.1145/3664647.3681033"},{"id":9,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"GoldFinch: High Performance RWKV/Transformer Hybrid with Linear Pre-Fill and Extreme KV-Cache Compression","date":"2024-07-16","content":"GoldFinch 是一种 RWKV/Transformer 混合序列模型，将新的 GOLD transformer 叠加在 Finch（RWKV-6）架构的增强版本之上，有效地在线性时间和空间中生成高度压缩和可重用的 KV-Cache。相对于1.5B 参数的 Finch 和 Llama 模型而言，GoldFinch 的建模性能显着提高。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240716-1.png\r\n","author":"Daniel Goldstein【EleutherAI, Recursal AI】","link":"https://arxiv.org/abs/2407.12077","category":"[\"通用\",\"混合架构\"]","conference_name":"","conference_url":""},{"id":10,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Restore-RWKV: Efficient and Effective Medical Image Restoration with RWKV","date":"2024-07-14","content":"Restore-RWKV 是首个基于 RWKV 的医学图像修复模型。文章提出了一种循环 WKV（Re-WKV）注意力机制，该机制以线性计算复杂度捕获全局依赖关系。Restore-RWKV 在各种医学图像修复任务中均具有卓越的性能，包括 MR I图像超分辨率、CT 图像去噪、PET 图像合成和一体化医学图像修复。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240714-1.png\r\n","author":"杨植文【北航大学】","link":"https://arxiv.org/abs/2407.11087","category":"[\"图像\",\"医学图像修复\"]","conference_name":"","conference_url":""},{"id":48,"created_at":"2025-01-06T03:59:02.905409+00:00","title":"Enhancing Transformer RNNs with Multiple Temporal Perspectives","date":"2024-07-11","content":"论文提出多时间视角概念以增强循环神经网络（RNN）。将其应用于 RWKV 模型时，能以极少的参数增加丰富上下文理解。实证结果验证了其有效性，在基准测试中表现提升且保持线性推理复杂度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240711-1.png\r\n","author":"Razvan-Gabriel Dumitru【美国亚利桑那大学】","link":"https://arxiv.org/abs/2402.02625","category":"[\"通用\",\"循环神经网络\",\"多时间角度\"]","conference_name":"ICML 2024 workshop","conference_url":"https://icml.cc/virtual/2024/36134"},{"id":11,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Mamba or RWKV: Exploring High-Quality and High-Efficiency Segment Anything Model","date":"2024-06-27","content":"RWKV-SAM（Segment Anything Model）是基于 RWKV 的图像分段切割方法。与 Transformer 模型相比，RWKV-SAM（图像分割模型） 实现了 2 倍以上的加速，且可以在各种数据集上实现更好的图像分割性能。此外，RWKV-SAM 的分类和语义分割结果优于最新的视觉 Mamba 模型。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240627-1.png\r\n","author":"原昊博【南洋理工大学】","link":"https://arxiv.org/abs/2406.19369","category":"[\"图像\",\"图像分割\"]","conference_name":"","conference_url":""},{"id":12,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"VisualRWKV: Exploring Recurrent Neural Networks for Visual Language Models","date":"2024-06-19","content":"VisualRWKV 是基于 RWKV 的可视化语言模型，能够处理各种可视化任务。与基于 Transformer 的模型（如 LLaVA-1.5）相比，VisualRWKV 在各种基准测试中实现了具有竞争力的性能。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240619-1.png\r\n","author":"侯皓文【元始智能】","link":"https://arxiv.org/abs/2406.13362","category":"[\"图像\",\"视觉语言\",\"多模态\"]","conference_name":"COLING 2025 | CCF B","conference_url":"https://openreview.net/forum?id=DiMETwbmke"},{"id":13,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"RWKV-CLIP: A Robust Vision-Language Representation Learner","date":"2024-06-11","content":"RWKV-CLIP 是一个 RWKV 驱动的视觉语言表示学习模型，该框架在多个下游任务中实现了最先进的性能，包括线性探测、零样本分类，以及零样本图像文本检索。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240611-1.png\r\n","author":"杨铠成【格灵深瞳】","link":"https://arxiv.org/abs/2406.06973","category":"[\"图像\",\"视觉-语言表示（CLIP）\"]","conference_name":"EMNLP 2024 | CCF B","conference_url":"https://aclanthology.org/2024.emnlp-main.276/"},{"id":14,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning","date":"2024-05-24","content":"PointRWKV 项目是一种基于 RWKV 的 3D 点云学习框架，在下游点云任务上性能优于基于 Transformer 和 Mamba 的同类工作，显著节省了约 46% 的 FLOPS。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240524-1.png\r\n","author":"何庆东【腾讯优图】","link":"https://arxiv.org/abs/2405.15214","category":"[\"3D/视频\",\"3D点云\",\"3D学习\"]","conference_name":"AAAI 2025 | CCF A","conference_url":"https://ojs.aaai.org/index.php/AAAI/article/view/32353"},{"id":15,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence","date":"2024-04-08","content":"RWKV-5/6 架构的论文，在 RWKV-4 架构基础上行进行了优化，包括多头矩阵值状态和动态递归机制。这些优化可提高表达能力，同时保持 RNN 的推理效率特征。该论文还引入了一个包含 1.12 万亿个 tokens 的新多语言语料库和一个基于贪婪匹配的快速分词器。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240408-1.png\r\n","author":"Bo Peng【元始智能】","link":"https://arxiv.org/abs/2404.05892","category":"[\"通用\",\"RWKV架构论文\"]","conference_name":"COLM 2024","conference_url":"https://openreview.net/forum?id=soz1SEiPeq#discussion"},{"id":16,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Diffusion-RWKV: Scaling RWKV-Like Architectures for Diffusion Models","date":"2024-04-06","content":"Diffusion-RWKV 用于图像生成任务，实验结果表明，Diffusion-RWKV 在 FID 和 IS 指标上达到了与现有基于 CNN 或 Transformer 的扩散模型相当甚至超越的性能，同时显著减少了总体计算 FLOP 的使用量。其优势在降低了空间聚合的复杂性，因此在处理高分辨率图像方面特别出色。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240406-1.png\r\n","author":"费政聪【昆仑万维】","link":"https://arxiv.org/abs/2404.04478","category":"[\"图像\",\"图像生成\",\"扩散模型\"]","conference_name":"","conference_url":""},{"id":17,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Onboard deep lossless and near-lossless predictive coding of hyperspectral images with line-based attention","date":"2024-03-26","content":"LineRWKV是一个应用于在航天器上压缩高光谱图像的预测神经网络。 LineRWKV 对每个像素执行预测，然后对残差进行熵编码，并逐行递归地工作以限制内存消耗。在 HySpecNet-11k 数据集和 PRISMA 图像上的实验表明，LineRWKV 是第一个在无损和近无损压缩下优于 CCSDS-123.0-B-2 的学习方法。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240326-1.png\r\n","author":"Diego Valsesia【意大利都灵理工大学】","link":"https://arxiv.org/abs/2403.17677","category":"[\"图像\",\"高光谱图像压缩\",\"航天应用\"]","conference_name":"IEEE TGRS | CCF A","conference_url":"https://ieeexplore.ieee.org/document/10684791"},{"id":18,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Vision-RWKV: Efficient and Scalable Visual Perception with RWKV-Like Architectures","date":"2024-03-07","content":"Vision-RWKV（VRWKV）是在 RWKV 模型的基础上改进而成的一种适用于视觉任务的模型。VRWKV 可以作为 ViT 的低成本替代方案，其线性RNN架构具有较低的计算成本，并保留了 ViT 级别的性能。该算法在处理高分辨率图像时具有较低的空间复杂度，能平滑处理高分辨率图像。在图像分类任务中，VRWKV 优于 ViT，在处理高分辨率输入时速度更快，内存效率更高。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240307-1.png\r\n","author":"Yuchen Duan【上海人工智能实验室】","link":"https://arxiv.org/abs/2403.02308","category":"[\"图像\",\"视觉任务\",\"高效计算\"]","conference_name":"ICLR 2025 | CCF A","conference_url":"https://iclr.cc/media/iclr-2025/Slides/28412.pdf"},{"id":19,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"TLS-RWKV: Real-Time Online Action Detection with Temporal Label Smoothing","date":"2024-02-19","content":"论文提出了一种基于带有时间标签平滑（Temporal Label Smooth）的 RWKV 模型的 OAD 方法。RWKV 模型既能捕捉时间依赖性又能高效计算，这使其非常适合实时应用。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240219-1.png\r\n","author":"朱子琪、邵武昌、焦东东【电子六所】","link":"https://link.springer.com/article/10.1007/s11063-024-11540-0","category":"[\"3D/视频\",\"实时动作检测\",\"时间标签平滑\"]","conference_name":"Neural Processing Letters | JCR Q2","conference_url":"https://link.springer.com/journal/11063"},{"id":20,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"SDiT: Spiking Diffusion Model with Transformer","date":"2024-02-18","content":"论文提出了 Spiking Diffusion Transformer （SDiT - 基于 Transformer 的新型 SNN 扩散模型架构），但采用 RWKV 作为 Transformer 自注意力机制的替代。通过将 RWKV 有效地与 SNN 集成，SDiT 方法提高了重建图像的质量。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240218-1.png\r\n","author":"杨舒、马涵志【浙江大学】","link":"https://arxiv.org/abs/2402.11588","category":"[\"图像\",\"脉冲神经网络\",\"扩散模型\"]","conference_name":"","conference_url":""},{"id":21,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"RWKV-TS: Beyond Traditional Recurrent Neural Network for Time Series Tasks","date":"2024-01-17","content":"RWKV-TS 是一种基于时间序列的模型，可以在时间序列任务上同时实现强大的性能和效率。该模型具有三个显著特点：1.新颖的 RNN 架构，具有 O(L) 的时间复杂度和内存使用特性；2.相比传统 RNN，增强了捕获长期序列信息的能力；3.具备高度计算效率并能够有效地进行扩展。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20240117-1.png\r\n","author":"侯皓文【元始智能】","link":"https://arxiv.org/abs/2401.09093","category":"[\"序列/强化\",\"时间序列分析\"]","conference_name":"","conference_url":""},{"id":22,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Advancing VAD Systems Based on Multi-Task Learning with Improved Model Structures","date":"2023-12-19","content":"这篇论文将 RWKV 应用于语音活动检测 （VAD），实验表明，基于 RWKV 的实时语义 VAD 系统在多个指标上均优于基于 SAN-M 的离线语义 VAD 系统。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20231219-1.png\r\n","author":"Lingyun Zuo【阿里巴巴】","link":"https://arxiv.org/abs/2312.14860","category":"[\"音频/语音\",\"语音活动检测\"]","conference_name":"","conference_url":""},{"id":23,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"RWKV-based Encoder-Decoder Model for Code Completion","date":"2023-11-17","content":"这篇论文将 RWKV 应用于智能代码补全，帮助程序员减少错误和提高编程效率。实验证明，RWKV 代码补全模型在三个指标上优于基线模型：准确率、EM 和 Edit Sim。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20231117-1.png\r\n","author":"Lu Zhou【重庆西南大学】","link":"https://ieeexplore.ieee.org/abstract/document/10442108","category":"[\"语言\",\"代码补全\",\"编程辅助\"]","conference_name":"EIECT 2023","conference_url":"https://ieeexplore.ieee.org/xpl/conhome/10441785/proceeding"},{"id":24,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"RWKV: A Linear Attention Mechanism for Temperature and Humidity Compensation for Gas Sensors","date":"2023-10-25","content":"这篇论文将 RWKV 应用于用气体传感器，结合线性注意机制来优化传感器的响应能力。实验证明， RWKV 在检测气态物质领域取得了极低的误差值。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20231025-1.png\r\n","author":"Shihao Xie【成都电子科技大学】","link":"https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4612708","category":"[\"序列/强化\",\"气体传感器\",\"环境监测\"]","conference_name":"","conference_url":""},{"id":25,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"Exploring RWKV for Memory Efficient and Low Latency Streaming ASR","date":"2023-09-26","content":"这篇论文将 RWKV 应用于流式 ASR（自动语音识别），实验表明，在最小化延迟和推理内存成本的同时，RWKV 实现了与分块 Conformer 转录器相当甚至更好的准确度。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20230926-1.png\r\n","author":"Keyu An, Shiliang Zhang【阿里巴巴】","link":"https://arxiv.org/abs/2309.14758","category":"[\"音频/语音\",\"语音识别\"]","conference_name":"","conference_url":""},{"id":27,"created_at":"2024-10-09T09:03:03.911847+00:00","title":"RWKV: Reinventing RNNs for the Transformer Era","date":"2023-05-22","content":"RWKV-4 架构的论文，首次提出 Receptance Weighted Key Value (RWKV) 架构，实现了在训练期间并行计算，并在推理期间保持恒定的计算和内存复杂性。","img":"https://api.rwkv.cn/storage/v1/object/public/cnImg/paper-renamed/img-20230522-1.png\r\n","author":"Bo Peng【元始智能】","link":"https://arxiv.org/abs/2305.13048","category":"[\"通用\",\"RWKV架构\"]","conference_name":"EMNLP 2023 | CCF B","conference_url":"https://2023.emnlp.org/program/accepted_findings/"}]}