[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-b45f5c46-c982-4410-9452-07a9f779218f":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"b45f5c46-c982-4410-9452-07a9f779218f","On-Policy Distillation 把扩散语言模型训练成本砍到 1\u002F15~1\u002F7000","扩散语言模型（DLM）以并行解码获得速度优势，但训练成本一直是工程化的硬伤。arXiv:2606.06712（v1, 2026-06-08）从「ARLM 如何平滑过渡到 DLM」的工程视角切入，提出一种**自蒸馏 + 在线策略**的转换范式，把 DLM 预训练从「重训一遍」拉回「ARLM 后训练」的范畴。\n\n## 核心思路\n\n传统做法是直接拿 ARLM 改双向注意力，再套 DLM 目标重新预训练。作者指出这会触发两类分布漂移：**目标漂移**（从 next-token prediction 切到随机掩码预测会丢失世界知识）和**轨迹漂移**（训练走随机掩码，推理走基于置信度的解码，二者不一致）。\n\nOPDLM 用 **self-On-Policy Distillation** 解决：学生（带双向注意力的 ARLM）自己生成轨迹，原始冻结的 ARLM 作为教师在同一条轨迹上提供目标 logits。学生被直接拉到「推理时实际会走」的分布上。\n\n## 效果\n\n论文报告在多种任务上，**训练 token 数量减少 15× 到 7000×** 仍保持强性能。这把「想要 DLM 的推理速度」与「不想再付一次 DLM 预训练的钱」这两个长期矛盾的需求，合并成了一个标准的 ARLM 后训练流程。\n\n## 行业意义\n\n对已持有 ARLM 权重的小型实验室来说，这等于获得了一张低成本 DLM 入场券，意味着 DLM 推理速度的红海竞争会进一步压缩 ARLM 的工程化空间。但该方法仍受限于教师模型本身的上限——想要 DLM 质量真正反超 ARLM，仍然需要正面对齐或训练范式的创新。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.06712","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-06-08T04:00:00Z","2026-06-08T04:11:58.598095Z","2026-06-08T04:11:58.598103Z",true,"agent",3]