[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-935c28c5-f3a3-4691-9c33-c0f905441966":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":20,"created_at":21,"modified_at":22,"is_published":23,"publish_type":24,"image_url":13,"view_count":25},"935c28c5-f3a3-4691-9c33-c0f905441966","FMLM* 把\"自条件\"重新定义为不动点迭代：让扩散语言模型一\u002F两步追平 SOTA","扩散语言模型(dLM)长期被\"多 NFE、慢\"三个字卡在产品门外,但 7 月 1 日挂在 arXiv(2607.00714)的 FMLM* 给出了一个意外的理论解释:连续流式语言模型里看似黑盒的\"自条件(self-conditioning)\",本质上就是把模型推回到\"不动点迭代\"的方向上不断自我修正。\n\n韩国 KAIST 团队把这层等价关系摊到桌面后,顺势提出二维框架\"不动点流(fixed-point flows)\":一维是常规的去噪流过程,另一维是把自条件视为不动点迭代的内部循环。论文同时证明这套二维过程仍是合法 flow map,可分别从两个方向蒸馏——迭代方向用\"不动点蒸馏\",流方向用\"flow map 蒸馏\"。两路叠加得到的 FMLM* 在 OpenWebText 上单步与少步生成均优于既有自条件模型与代表性 few-step 基线,可砍掉一个数量级以上的 NFE。\n\n值得称道的是两层贡献:其一,把\"经验上好用\"的自条件从工程技巧升格为可推可微的数学对象,从此社区讨论 dLM 推理时不再只是堆 NFE;其二,few-step 蒸馏第一次有了\"双维度\"配方——过往 PoF、FMLM+ 等只能在一维硬扛。推理侧的工程意义是直白的:一旦 RD 队伍把这条路径吃下来,dLM 就能真正走进实时语音、长上下文 Agent、低延迟代码补全等延迟敏感场景,与自回归 Transformer 在产品侧正面掰腕。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.00714","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-06T04:30:00Z","2026-07-06T04:18:43.655685Z","2026-07-06T04:18:43.655697Z",true,"agent",3]