<![CDATA[qq_44681809的博客]]>

<![CDATA[qq_44681809的博客]]>https://blog.youkuaiyun.com/qq_44681809zh-cnhttps://blog.youkuaiyun.com/<![CDATA[从 LoRA 到 MoE-LoRA（七）：挑战二：专家粒度——从 “模块级” 到 “rank 级” 再到 “结构级”]]>https://blog.youkuaiyun.com/qq_44681809/article/details/170186086https://blog.youkuaiyun.com/qq_44681809/article/details/170186086qq_44681809Fri, 06 Nov 2026 09:39:02 +0800<![CDATA[从 LoRA 到 MoE-LoRA（六）：挑战二：专家粒度——从 “模块级” 到 “rank 级” 再到 “结构级”]]>https://blog.youkuaiyun.com/qq_44681809/article/details/170121115https://blog.youkuaiyun.com/qq_44681809/article/details/170121115qq_44681809Thu, 05 Nov 2026 08:57:30 +0800<![CDATA[从 LoRA 到 MoE-LoRA（五）：挑战一：路由机制——从 “静态选谁” 到 “动态怎么选”]]>https://blog.youkuaiyun.com/qq_44681809/article/details/170022480https://blog.youkuaiyun.com/qq_44681809/article/details/170022480qq_44681809Wed, 04 Nov 2026 09:17:46 +0800<![CDATA[从 LoRA 到 MoE-LoRA（四）：挑战一：路由机制——从 “静态选谁” 到 “动态怎么选”]]>https://blog.youkuaiyun.com/qq_44681809/article/details/169163139https://blog.youkuaiyun.com/qq_44681809/article/details/169163139qq_44681809Tue, 03 Nov 2026 08:57:22 +0800<![CDATA[从 LoRA 到 MoE-LoRA（三）：挑战一：路由机制——从 “静态选谁” 到 “动态怎么选”]]>https://blog.youkuaiyun.com/qq_44681809/article/details/169165160https://blog.youkuaiyun.com/qq_44681809/article/details/169165160qq_44681809Mon, 02 Nov 2026 10:36:41 +0800<![CDATA[从 LoRA 到 MoE-LoRA（二）：挑战一：路由机制——从 “静态选谁” 到 “动态怎么选”]]>https://blog.youkuaiyun.com/qq_44681809/article/details/169162929https://blog.youkuaiyun.com/qq_44681809/article/details/169162929qq_44681809Fri, 30 Oct 2026 09:25:10 +0800<![CDATA[从 LoRA 到 MoE-LoRA（一）：LoRA 与 MoE 是如何走到一起的]]>https://blog.youkuaiyun.com/qq_44681809/article/details/169160743https://blog.youkuaiyun.com/qq_44681809/article/details/169160743qq_44681809Mon, 26 Oct 2026 16:00:05 +0800<![CDATA[（2026|中科院自动化所，路由冗余，动态分配的价值在激进剪枝，多模态更脆弱）你只需要选定专家的 2/3：细粒度 MoE LLM 中的动态专家剪枝]]>https://blog.youkuaiyun.com/qq_44681809/article/details/168849846https://blog.youkuaiyun.com/qq_44681809/article/details/168849846qq_44681809Fri, 23 Oct 2026 14:31:25 +0800<![CDATA[（2026|COLM|Cornell，跨层/连续块共享专家池，跨层专家复用，扩大路由组合空间，深度 embedding）MoRE：专家复用 MoE]]>https://blog.youkuaiyun.com/qq_44681809/article/details/168745478https://blog.youkuaiyun.com/qq_44681809/article/details/168745478qq_44681809Thu, 22 Oct 2026 14:14:51 +0800<![CDATA[MoE 与推测解码（Speculative Decoding）加速：为什么慢、什么时候慢、怎么让它不慢？]]>https://blog.youkuaiyun.com/qq_44681809/article/details/168133662https://blog.youkuaiyun.com/qq_44681809/article/details/168133662qq_44681809Tue, 20 Oct 2026 16:07:26 +0800<![CDATA[（2026|NAVER AI，推测解码/SD，置信门控，Loop-aware Training，用更多 draft 前向换更少 target 前向）DLoop：循环推测解码]]>https://blog.youkuaiyun.com/qq_44681809/article/details/167882726https://blog.youkuaiyun.com/qq_44681809/article/details/167882726qq_44681809Fri, 16 Oct 2026 10:20:22 +0800<![CDATA[（2026|Cornell & IBM，共享 KV，首轮建记忆后续用记忆，共享记忆省内存提质量）LPT/H-LPT：共享记忆在循环 Transformer 中的有效性]]>https://blog.youkuaiyun.com/qq_44681809/article/details/167697622https://blog.youkuaiyun.com/qq_44681809/article/details/167697622qq_44681809Tue, 13 Oct 2026 18:17:26 +0800<![CDATA[Looped Transformer 最新进展：从循环 MoE、缩放定律到优化器与扩散生成]]>https://blog.youkuaiyun.com/qq_44681809/article/details/167221964https://blog.youkuaiyun.com/qq_44681809/article/details/167221964qq_44681809Sun, 11 Oct 2026 18:01:07 +0800<![CDATA[（2017|ICLR|Google Brain，主网络与超网络联合训练，静态/动态超网络，LSTM，RNN）超网络]]>https://blog.youkuaiyun.com/qq_44681809/article/details/164721496https://blog.youkuaiyun.com/qq_44681809/article/details/164721496qq_44681809Wed, 07 Oct 2026 11:32:40 +0800<![CDATA[（2026|清华 & 微软，专家空间探索，路由扰动，自适应噪声，锚定专家采样，Rollout 路由回放）ESRL：在 MoE 强化学习中的专家空间探索 ]]>https://blog.youkuaiyun.com/qq_44681809/article/details/165751577https://blog.youkuaiyun.com/qq_44681809/article/details/165751577qq_44681809Wed, 07 Oct 2026 11:32:13 +0800<![CDATA[（2026|斯坦福，MoE 比 Dense 更怕重复数据，退化程度主要由总参数 / 稀疏度决定，正则有效）数据稀缺与模型稀疏：MoE 对重复数据过拟合]]>https://blog.youkuaiyun.com/qq_44681809/article/details/165710911https://blog.youkuaiyun.com/qq_44681809/article/details/165710911qq_44681809Thu, 17 Sep 2026 15:14:51 +0800<![CDATA[（2026|清华，逼近-估计-计算的权衡，Router，门控与分配区域的几何对齐，稀疏 MoE 的逼近能力，共享专家）对 MoE 的统计学理解]]>https://blog.youkuaiyun.com/qq_44681809/article/details/165611459https://blog.youkuaiyun.com/qq_44681809/article/details/165611459qq_44681809Wed, 16 Sep 2026 16:36:01 +0800<![CDATA[（2026|MirroS，可执行世界表示/EWR，Agentic Discovery Loop，EWR 为 VLM 提供物理监督）代码作为世界：用于物理推理的 EWR 的 Agent 发现]]>https://blog.youkuaiyun.com/qq_44681809/article/details/164472258https://blog.youkuaiyun.com/qq_44681809/article/details/164472258qq_44681809Wed, 16 Sep 2026 16:35:49 +0800<![CDATA[IEEE Transactions on Industrial Informatics (TII) 投稿流程]]>https://blog.youkuaiyun.com/qq_44681809/article/details/163850080https://blog.youkuaiyun.com/qq_44681809/article/details/163850080qq_44681809Tue, 08 Sep 2026 14:39:42 +0800<![CDATA[（2026|IPI|我的论文投稿中，PSP 超轻量采样算法，轻量化架构探索/早融合+头压缩）LUMIN：面向工业异常检测的轻量级通用制造检测网络]]>https://blog.youkuaiyun.com/qq_44681809/article/details/164370960https://blog.youkuaiyun.com/qq_44681809/article/details/164370960qq_44681809Mon, 07 Sep 2026 10:06:20 +0800