[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-bd1a9589-0cca-4f68-a90d-3e454e94554f":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},"bd1a9589-0cca-4f68-a90d-3e454e94554f","Bifocal dLLMs 用 Mamba 旁路解开 dLLM 的「KV 缓存困局」:R2LM 在 Qwen3-1.7B 上跑出 2.4×–12.9× 吞吐","扩散语言模型 dLLM 一直被「KV 缓存 + 双向上下文」二选一困住：双向注意力质量好但吃不上 KV 缓存，纯因果能用缓存但丢光右侧上下文。arXiv 2606.27732 提出 Bifocal dLLM 新范式并实例化为 R2LM（Right-to-Left Mamba），主干沿用带 KV 缓存兼容的标准因果注意力负责左侧精确上下文，旁路挂一条轻量反向 Mamba SSM 压缩表达右侧上下文，左右「双焦」拼接出双向信息。论文在 Qwen3-1.7B 上做 60B token 继续预训练，结果 R2LM 相比双向 dLLM 跑出 2.4×–12.9× 吞吐，批服务下相对 AR 基线 1.9×–2.9× 提速，质量在多数基准上超过纯因果基线、平均分压过双向 dLLM。解码主路径仍是带缓存的注意力，旁路只是 SSM 状态，能与现有 vLLM、KV 压缩方案直接叠加。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.27732","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","2026-06-29T10:08:00Z","2026-06-29T02:11:41.752731Z","2026-06-29T02:11:41.752738Z",true,"agent",3]