<![CDATA[青年AI研究员idea加油站，AI开发者的新能源充电桩。]]>https://blog.youkuaiyun.com/QingKeLabzh-cnhttps://blog.youkuaiyun.com/<![CDATA[别再让计算等着搬权重了！KTransformers 让 CPU 直接参与 MoE 微调计算]]>https://blog.youkuaiyun.com/QingKeLab/article/details/169265073https://blog.youkuaiyun.com/QingKeLab/article/details/169265073QingKeLabTue, 27 Oct 2026 17:26:17 +0800<![CDATA[我们可能一直都低估了小模型！Select to Think：也许选择比生成更重要]]>https://blog.youkuaiyun.com/QingKeLab/article/details/167980977https://blog.youkuaiyun.com/QingKeLab/article/details/167980977QingKeLabThu, 15 Oct 2026 08:02:06 +0800<![CDATA[ICML 2026｜让 Agent 真正和谐作战：GoS 为多智能体推理构建共享信念状态]]>https://blog.youkuaiyun.com/QingKeLab/article/details/162547796https://blog.youkuaiyun.com/QingKeLab/article/details/162547796QingKeLabFri, 03 Jul 2026 12:27:01 +0800<![CDATA[ICML 2026 | 一段视频，同时听懂四道命令！OmniShow把多模态可控视频生成整合成统一「调音台」]]>https://blog.youkuaiyun.com/QingKeLab/article/details/162498699https://blog.youkuaiyun.com/QingKeLab/article/details/162498699QingKeLabThu, 02 Jul 2026 00:12:33 +0800<![CDATA[就从来没人质疑过 OPD 的 reward 设计吗？我们发现了 log 本身就是问题，也许就不该用！]]>https://blog.youkuaiyun.com/QingKeLab/article/details/162385493https://blog.youkuaiyun.com/QingKeLab/article/details/162385493QingKeLabSun, 28 Jun 2026 11:14:28 +0800<![CDATA[ACL 2026｜SFT 为何总是学习失败？也不是所有 SFT 失败都该加 epoch！教你修复 SFT 的五把「手术刀」]]>https://blog.youkuaiyun.com/QingKeLab/article/details/162094622https://blog.youkuaiyun.com/QingKeLab/article/details/162094622QingKeLabThu, 18 Jun 2026 12:11:51 +0800<![CDATA[从 LLM 的局限到世界模型：LeWorldModel 为何更接近 AI 的第一性原理？]]>https://blog.youkuaiyun.com/QingKeLab/article/details/162046975https://blog.youkuaiyun.com/QingKeLab/article/details/162046975QingKeLabTue, 16 Jun 2026 22:44:52 +0800<![CDATA[[特殊字符]什么？SFT、DAgger、离线 RL 和 OPD，竟然是同一张 2×2 表格上的四个格子！]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161984159https://blog.youkuaiyun.com/QingKeLab/article/details/161984159QingKeLabSun, 14 Jun 2026 21:24:39 +0800<![CDATA[直播预告！从 MLA 到 GQLA：无需从头训练，硬件自适应高效注意力机制]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161980004https://blog.youkuaiyun.com/QingKeLab/article/details/161980004QingKeLabSun, 14 Jun 2026 15:34:29 +0800<![CDATA[清华最新提出全新 VLM 架构 DPA：让多模态大模型先“对齐”，再“理解”！]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161931560https://blog.youkuaiyun.com/QingKeLab/article/details/161931560QingKeLabFri, 12 Jun 2026 19:30:32 +0800<![CDATA[还在用 2D 空间训具身？南洋理工 MMLab 提出 4D 具身世界模拟器，构建虚拟与现实的时空桥梁！]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161930816https://blog.youkuaiyun.com/QingKeLab/article/details/161930816QingKeLabFri, 12 Jun 2026 16:17:17 +0800<![CDATA[音频编辑模型全军覆没！多机构联合发布MMAE：首个通用音频编辑测评基准]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161903765https://blog.youkuaiyun.com/QingKeLab/article/details/161903765QingKeLabThu, 11 Jun 2026 23:08:16 +0800<![CDATA[视觉 FLOPs 减半！清华最新成果 LLaVA-UHD v4，突破高分辨率多模态大模型的视觉编码效率瓶颈]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161870823https://blog.youkuaiyun.com/QingKeLab/article/details/161870823QingKeLabWed, 10 Jun 2026 22:06:56 +0800<![CDATA[给 Agent RL 加个“事后复盘”：StepOPSD 让模型精准找到“哪一步走错了”？]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161728154https://blog.youkuaiyun.com/QingKeLab/article/details/161728154QingKeLabFri, 05 Jun 2026 21:51:05 +0800<![CDATA[直播预告！从 ARPO，到 AEPO，再到 Agent-World：探索通用智能体训练的可行路径]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161661009https://blog.youkuaiyun.com/QingKeLab/article/details/161661009QingKeLabWed, 03 Jun 2026 17:04:42 +0800<![CDATA[直播预告｜从 MiniLLM 开始，为什么 OPD 正在成为推理模型时代的“基础设施”？]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161462396https://blog.youkuaiyun.com/QingKeLab/article/details/161462396QingKeLabWed, 27 May 2026 21:39:48 +0800<![CDATA[长文本推理一定要改架构？阿里最新提出 RTPurbo：仅需百步训练，无损达到 97%+ 稀疏度与 9 倍加速]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161409530https://blog.youkuaiyun.com/QingKeLab/article/details/161409530QingKeLabTue, 26 May 2026 08:01:52 +0800<![CDATA[清华最新论文！动态 MoE 迁移算法 ZEDA：让 MoE 自己学会“摸鱼“，推理提速 20%]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161270884https://blog.youkuaiyun.com/QingKeLab/article/details/161270884QingKeLabThu, 21 May 2026 08:37:21 +0800<![CDATA[ICLR 2026｜Next-ToBE：让“自信而短视“的大模型看得更远、推理更准]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161214476https://blog.youkuaiyun.com/QingKeLab/article/details/161214476QingKeLabTue, 19 May 2026 10:28:23 +0800<![CDATA[ACL 2026 Findings | 浙大提出 GFT：On-Policy SFT 视角下的奖励微调]]>https://blog.youkuaiyun.com/QingKeLab/article/details/161197649https://blog.youkuaiyun.com/QingKeLab/article/details/161197649QingKeLabMon, 18 May 2026 16:37:50 +0800