[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-pcomoe-moe-path-composition-inference":3,"news-related-4f745af8-2962-450f-a4b1-612b796d7dd7":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"4f745af8-2962-450f-a4b1-612b796d7dd7","PCoMoE:把 MoE 专家拆成路径组合,推理提速 1.31 倍还涨 10% 精度","EMNLP 2026 主会论文 PCoMoE 提出 MoE 推理新范式:不再把专家当原子单位,而是拆成细粒度路径动态组合。论文报告端到端推理最高提速 1.31 倍,模型精度反升 10%,代码仓库已上线、实现整理中。","MoE(混合专家)架构让大模型用稀疏激活换来了参数规模的自由,但推理侧一直有个没人挑战的默认假设:专家(expert)是最小的调度单位。路由器选中哪几个专家,整颗专家网络就完整跑一遍,不多不少。EMNLP 2026 主会接收的新论文 PCoMoE(arXiv:2609.01024)把这个假设拆了——作者主张,专家内部还有大量可复用的计算结构,推理时真正该调度的不是\"整颗专家\",而是专家内部的\"路径\"。\n\n## 问题:整颗专家是个\"集装箱\"\n\n论文对现状的诊断很直接:现代 MoE 推理框架无论做管理、调度还是剪枝,都把专家当作原子执行单元(atomic execution units)处理。这带来两个后果:一是优化边界被过早固定——在\"专家级\"粒度上做优化,专家内部的计算冗余就永远看不见;二是 MoE 的专家计算天然存在可共享的中间结构,但框架只认\"整颗专家\",这些复用机会全部浪费。\n\n用物流做类比,现在的 MoE 推理像按\"整个集装箱\"收费,哪怕两票货可以拼箱,系统也不支持。\n\n## 方案:把专家拆成可组合的路径\n\nPCoMoE 是一个路径组合式(path-compositional)执行框架,设计分三层:\n\n- **路径级建模**:把专家计算形式化为细粒度的组合路径。按官方仓库说明,专家被分解为可复用的 expansion 侧与 projection 侧组件;\n- **兼容性感知的逐层剪枝**:抑制低价值的路径组合,避免组合爆炸;\n- **硬件友好的执行引擎**:利用可复用的子专家结构,并把额外开销严格控制在有界范围内。\n\nGitHub 仓库还提到一个关键设计:source-grouped compute reuse,按来源分组做计算复用,让多个被激活的专家共享底层计算。\n\n## 结果:快了,还准了\n\n论文报告的数字:端到端推理最高提速 1.31 倍,同时模型精度提升 10%。加速与精度通常此消彼长——剪枝换速度、掉点作代价是常规操作,PCoMoE 报告的却是两个方向同时改善,机制上可解释为低价值路径被剪掉后,模型反而少受\"坏组合\"干扰。当然,这些是作者自报的实验数字,独立复现还要等社区验证。\n\n需要提醒:官方代码仓库(github.com\u002Fgzyyy0\u002FPCoMoE)已上线,但目前只有论文信息与致谢,实现代码标注\"整理中,即将发布\",想直接上手还得再等等。\n\n## 所以呢\n\nMoE 推理优化的竞争正从框架层打到结构层。既有工作大多在路由策略与专家并行上做文章,PCoMoE 把刀伸进了专家内部,给出一个更细粒度的执行抽象。这个方向的价值在于:MoE 用稀疏激活高效扩容的路子已被验证,模型规模越堆越大,专家内部的冗余浪费也会被成倍放大。对推理引擎团队,这是一条新战线;对研究者,\"执行单元的粒度\"本身就是一个值得重审的设计变量。论文详情见 arXiv:2609.01024。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.01024","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"6abfc454-54a9-4cb5-bcdc-f4b7e40cd997","en","PCoMoE: Decomposing MoE Experts into Composable Paths for 1.31x Faster Inference and 10% Higher Accuracy","PCoMoE, an EMNLP 2026 Main Conference paper, proposes a new MoE inference paradigm: experts are no longer treated as atomic units but decomposed into fine-grained composable paths. The paper reports up to 1.31x end-to-end speedup with a 10% accuracy gain; the code repository is online with the implementation still being prepared.","Mixture-of-Experts gave large language models the freedom to scale parameters via sparse activation, but inference has long rested on an unchallenged assumption: the expert is the smallest unit of scheduling. Whichever experts the router picks, each expert network runs in full—no more, no less. PCoMoE (arXiv:2609.01024), accepted to the EMNLP 2026 Main Conference, takes that assumption apart. The authors argue that experts contain substantial reusable computational structure, and what inference should really schedule is not the whole expert, but fine-grained paths inside it.\n\n## The Problem: The Whole Expert Is a \"Shipping Container\"\n\nThe paper's diagnosis is blunt. Modern MoE inference frameworks—whether they manage, schedule, or prune experts—treat experts as atomic execution units. This has two consequences. First, the optimization boundary gets fixed too early: optimizing at expert granularity leaves intra-expert computational redundancy permanently invisible. Second, expert computation in MoE naturally contains shareable intermediate structures, but since frameworks only recognize whole experts, those reuse opportunities are all wasted.\n\nA logistics analogy: today's MoE inference bills by the whole container, even when two shipments could share one box.\n\n## The Approach: Decompose Experts into Composable Paths\n\nPCoMoE is a path-compositional execution framework with three design layers:\n\n- **Path-level formulation**: expert computation is formalized as fine-grained compositional paths. Per the official repository, experts are decomposed into reusable expansion-side and projection-side components;\n- **Compatibility-aware layer-wise pruning**: suppresses low-value path combinations and keeps combinatorial explosion in check;\n- **A hardware-friendly execution engine**: exploits reusable sub-expert structures while keeping overhead strictly bounded.\n\nThe GitHub repository also highlights a key design element: source-grouped compute reuse, which groups computation by source so that multiple activated experts share underlying work.\n\n## The Results: Faster and More Accurate\n\nThe numbers reported in the paper: up to 1.31x end-to-end inference speedup, alongside a 10% improvement in model accuracy. Speed and accuracy usually trade off—pruning for speed at the cost of quality is the norm—yet PCoMoE reports improvement in both directions. A mechanistic explanation: once low-value paths are pruned, the model suffers less interference from \"bad combinations.\" That said, these are author-reported numbers; independent replication awaits community verification.\n\nOne caveat: the official code repository (github.com\u002Fgzyyy0\u002FPCoMoE) is live, but currently contains only paper information and acknowledgments. The implementation is marked as \"being organized, will be released soon\"—practical adoption will have to wait.\n\n## So What\n\nThe battle over MoE inference optimization is moving from the framework layer into the structural layer. Prior work mostly focused on routing strategy and expert parallelism; PCoMoE cuts into the expert itself and offers a finer-grained execution abstraction. The value of this direction: sparse activation has proven to be a viable way to scale MoE capacity, and as models keep growing, the redundancy wasted inside each expert gets multiplied accordingly. For inference engine teams, this is a new front; for researchers, the granularity of the execution unit is itself a design variable worth re-examining. Full paper: arXiv:2609.01024.","pcomoe-moe-path-composition-inference","2026-09-02T21:10:00Z","2026-09-02T21:09:27.790167Z","2026-09-02T21:09:27.790177Z",true,"agent",44,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"199cd4ef-f092-45a5-8635-91778dd2bce2","编译即训练：一句规约炼出 83.6% 准确率的本地神经函数，教师模型只用一次","compile-by-training-neural-functions","2026-09-04T23:08:03+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"7a6d28b6-65da-4a29-96b1-dedb9894de97","随机驱逐追平最强打分器:Salesforce 重写 KV Cache 压缩常识","random-attention-kv-cache-eviction","2026-09-04T19:08:26+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"b4f270b3-db43-4586-a0e5-a062320c6d1b","让模型自己声明看哪里:Declarative Attention 零训练砍 52% KV 读取","declarative-attention-kv-cache-declare","2026-09-03T23:07:03+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"44a035c8-b8a3-48e5-af4f-c76323dac7b5","RWKV7-G1j 13.3B 开源:不用注意力,每 token 推理成本是常数","rwkv7-g1j-13b-attention-free","2026-09-03T13:14:19+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"fdbe1ee2-131c-4632-baf6-03109d7c1814","Qwen3.8-Next 架构论文:125B 参数 6B 激活,1\u002F9 训练 FLOPs 对标 397B 前辈","qwen3-8-flash-next-architecture","2026-09-01T23:15:00+00:00"]