[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-misa-sparse-attention-moe-routing-4-8x":3,"topics-all":33,"news-related-64119fe9-634a-4a57-b035-e88f536600a4":52},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":20,"news_slug":26,"published_at":27,"created_at":28,"modified_at":29,"is_published":30,"publish_type":31,"image_url":13,"view_count":32},"64119fe9-634a-4a57-b035-e88f536600a4","MISA：稀疏注意力引入MoE路由，长上下文LLM推理效率提升4-8倍","2026年5月11日，北大研究团队在arXiv发布MISA论文，将MoE路由机制引入稀疏注意力，解决了当前稀疏注意力方案（如DeepSeek DSA）在长上下文场景下indexer计算成为瓶颈的问题。\n\nMISA将indexer heads视为专家池，引入轻量级路由器基于块级统计信息动态选择少量活跃heads执行token级评分，而非每次让所有heads处理全部前缀token。实验数据显示，在仅使用8个活跃heads（DeepSeek-V3.2）和4个（GLM-5）的情况下，MISA即可在LongBench上与Dense DSA Indexer表现持平，同时推理效率提升4-8倍。该方法无需额外训练，可直接作为DSA替代方案部署。\n\n稀疏注意力是LLM长上下文推理优化的重要方向，MISA的价值在于既保留下原始indexer池的多样性，又显著降低了计算成本。随着LLM上下文窗口不断增长，这类将MoE路由思想融合进注意力机制的技术路线，正成为提升推理效率的新趋势。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.07363","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[21],{"id":22,"lang":23,"title":24,"summary":25,"content":13},"943968f9-15c9-4a67-bf5d-2f3edcc1d244","en","MISA: sparse attention meets MoE routing, 4-8x faster inference","On May 11, 2026, a Peking University research team published the MISA paper on arXiv, introducing the MoE routing mechanism into sparse attention, solving the problem where the indexer computation becomes a bottleneck in long-context scenarios with current sparse attention solutions (like DeepSeek DSA).\n\nMISA treats indexer heads as an expert pool, introducing a lightweight router that dynamically selects a small number of active heads to perform token-level scoring based on block-level statistics, rather than having all heads process the full prefix tokens every time. Experimental data shows that with only 8 active heads (DeepSeek-V3.2) and 4 (GLM-5), MISA can match Dense DSA Indexer performance on LongBench, while inference efficiency improves 4-8×. The method requires no additional training and can be deployed directly as a drop-in DSA replacement.\n\nSparse attention is an important direction for optimizing long-context LLM inference, and MISA's value lies in preserving the diversity of the original indexer pool while significantly reducing compute cost. As LLM context windows continue to grow, this kind of technical approach — fusing MoE routing thinking into the attention mechanism — is becoming a new trend in improving inference efficiency.","misa-sparse-attention-moe-routing-4-8x","2026-05-13T07:30:00Z","2026-05-13T07:24:52.710167Z","2026-08-19T02:08:40.142862Z",true,"agent",235,[34,43],{"slug":35,"tag_slug":35,"title_zh":36,"title_en":37,"intro_zh":38,"intro_en":39,"id":40,"is_active":30,"created_at":41,"modified_at":42},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":44,"tag_slug":44,"title_zh":45,"title_en":46,"intro_zh":47,"intro_en":48,"id":49,"is_active":30,"created_at":50,"modified_at":51},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":53},[54,59,64,69,74,79],{"id":55,"title":56,"news_slug":57,"published_at":58},"02c96622-1409-4867-9f25-772936edc6cd","滑动窗口反超SnapKV:NVIDIA重审投机解码KV压缩","rift-draft-kv-cache-compression","2026-10-23T15:20:00+00:00",{"id":60,"title":61,"news_slug":62,"published_at":63},"4d36e54d-ca37-4d34-b8d9-74ebc25762cf","PACE:递归状态量化损失主因是舍入频率","pace-four-bit-recurrent-state-quantization","2026-10-23T01:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"be76fca7-bc02-4a76-a58f-17a661b82b60","SparPR:消费级显卡的KV缓存回收术","sparpr-kv-cache-recovery-consumer-gpu","2026-10-22T01:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"8c579d8d-4114-4d7d-92c1-22f5fab0537b","连续流语言模型跑到 8B：Lafa 单模型自验证解码提速 2.54 倍","large-flow-language-models-lafa-self-speculative-decoding","2026-10-21T03:09:03+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"8c6d408b-2913-48b2-bc1f-4e958ad854c5","Space Bunny 谜底揭晓:MiniMax M3.1 Flash,匿名期吞下 200 万亿 token","minimax-m3-1-flash-space-bunny","2026-10-19T21:10:12+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"559df680-f13e-48f0-9263-73116aaff411","上交大DLite免历史投机解码:Agent提速2.74倍","dlite-history-free-speculative-decoding","2026-10-19T17:07:52+00:00"]