[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-markovian-thinker-mila-cmu-delethink-linear":3,"topics-all":36,"news-related-039daf10-9444-409f-8a35-a9a29651928e":65},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"039daf10-9444-409f-8a35-a9a29651928e","「Markovian Thinker」把推理拉回线性：Mila\u002FCMU 用固定块切断 LongCoT 的二次方诅咒","Mila 与 CMU 研究者发布 arXiv:2510.06557，提出 Markovian Thinking 范式及其实例 Delethink：把推理切成固定大小块（如 8K token），到边界重置上下文、用一段简短的\"文本化 Markovian 状态\"承上启下，将 LongCoT-RL 的二次方计算与内存开销压回线性且常数。1.5B R1-Distill 用 8K 块训练即可推到 24K token 思考长度，匹配或超过直接用 24K 预算的 LongCoT-RL；96K 平均思考长度下成本由 27 H100-月降至 7 H100-月。该方法架构无关，可平移至 Mamba-3 等非 Transformer，并为百万 token 级长程推理打开工程通道。\n\n长思维链(LongCoT)被 R1 时代普遍视为推理能力的关键配方,但代价同样显眼:每生成一个新 token,attention 都要在 prompt 加上\"此前所有思考 token\"上重跑一遍——状态无界增长、计算与内存随长度二次方膨胀。这是把 o1 类模型推到数十万、百万 token 思考时最大的工程障碍。\n\narXiv:2510.06557 的思路很克制:不去优化 attention 本身,而是改写\"思考环境\"。研究者把 RL 训练中模型可见的状态,从\"全量上下文+历史 token\"换成\"固定大小状态\",即 Markovian Thinking。在实现 Delethink 中,推理被切成 8K-token 这样的定长块,块内正常思考;到达边界时,环境重置上下文,新 prompt 只保留原查询加上一段短 carryover。模型通过 RL 学会在 carryover 里写下足够\"任务关键状态\"以续上下一块,这一段文字即被称作\"textual Markovian state\"。\n\n效果上,1.5B R1-Distill 用 8K 块训练即可稳定推理到 24K token,在 MATH、AIME、GPQA 等数学\u002F科学基准上匹配或超过同预算 24K 的 LongCoT-RL;代码、PhD 级问答同样不落下风。作者给出一组有说服力的换算:平均思考长度 96K 时,LongCoT-RL 约 27 H100-月,Delethink 约 7 H100-月,成本几乎只剩四分之一,且 LongCoT 出现收益平台时 Delethink 仍在爬升。\n\n更值得注意的是\"架构无关\":文中在 RL 初始化阶段观察到 1.5B-120B 现有推理模型已能在多种基准上零样本采出 Markovian 迹,意味着这套训练范式不绑 Transformer,可与 Mamba-3、SSM\u002F线性注意力等近期架构正交组合。换言之,它瞄准的不是某一类模型的局部优化,而是把\"长推理\"从 attention 算力的二次方魔咒里整体松绑——这与本周 arXiv 2606 上 SCIN、SPARQLe 等推理系统层工作形成呼应。2026 年中段的 LLM 工程主线,正在从\"堆 GPU 跑大上下文\"转向\"重写计算结构换线性\"。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.06557","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"33bea5f8-c1b1-4ae4-9b3c-574d2d5bd162","en","Markovian Thinker: fixed blocks end LongCoT's quadratic curse","arXiv 2510.06557 (also 2606.x in the news) introduces \"Markovian Thinker,\" a reasoning framework from Mila and CMU that uses fixed-size \"blocks\" to break the quadratic complexity curse of Long Chain-of-Thought (LongCoT). The result: LongCoT reasoning at linear cost, with quality on par with full LongCoT.\n\nThe LongCoT curse: Long Chain-of-Thought (e.g., OpenAI o1, DeepSeek R1) generates thousands of reasoning tokens per query. The cost is O(n²) in the number of reasoning tokens — the attention complexity grows quadratically. For a 10K-token reasoning chain, the attention compute is 100M operations per layer.\n\nThe Markovian Thinker fix: split the reasoning into \"blocks\" of fixed size (e.g., 1024 tokens), and only attend within each block + a small \"summary\" of previous blocks. The block size is the only hyperparameter, and the result is O(n × b) complexity (linear in the total length, with a small constant b for the block size).\n\nThe benchmark: on a set of math and code reasoning tasks, Markovian Thinker with block size 1024 matches the quality of full LongCoT with 10× less compute. The reasoning quality is preserved because the \"summary\" of previous blocks captures the essential information.\n\nThe bigger takeaway: \"fixed-block reasoning\" is a significant step toward scalable LongCoT. The \"LongCoT is too expensive\" critique is the biggest barrier to deploying reasoning models in production, and Markovian Thinker is a clean, effective fix. For the industry, this means \"reasoning as a service\" can be priced more competitively, and reasoning models will become the default for complex tasks.","markovian-thinker-mila-cmu-delethink-linear","2026-06-21T08:00:00Z","2026-06-21T06:14:04.497328Z","2026-08-19T02:08:40.142862Z",true,"agent",200,[37,47,56],{"slug":38,"tag_slug":39,"title_zh":40,"title_en":41,"intro_zh":42,"intro_en":43,"id":44,"is_active":33,"created_at":45,"modified_at":46},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":48,"tag_slug":48,"title_zh":49,"title_en":50,"intro_zh":51,"intro_en":52,"id":53,"is_active":33,"created_at":54,"modified_at":55},"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":57,"tag_slug":57,"title_zh":58,"title_en":59,"intro_zh":60,"intro_en":61,"id":62,"is_active":33,"created_at":63,"modified_at":64},"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":66},[67,72,77,82,87,92],{"id":68,"title":69,"news_slug":70,"published_at":71},"0a4281e9-69b1-43f0-9e7b-78f0cd7ea4c2","医疗小模型回答砍72.8%,准确率反升3.2点","peca-conciseness-qlora-medical","2026-11-15T15:30:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"6186bff0-6bbb-4970-a0bc-b7bfa93a8e49","像素空间压缩:16个token装下RAG检索到的文档","opticalrag-pixel-space-rag-compression","2026-11-13T03:20:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"7984acc5-d045-4258-8da7-daf8ff25995b","一字节一token:27B模型词表砍到384的账本","qwen3-8-27b-byteify-vocab-384","2026-11-12T21:10:40+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"8d830f21-a44f-45b8-be6c-eed523d060dc","投机解码的天花板是分词器:3.4M参数换2.4倍","speculative-decoding-bpe-tokenizer-ceiling","2026-11-11T01:30:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"4f151ed3-74d5-4dc7-99d0-ac03df4c1b08","自一致别整条重采:定位错步,token至多省41%","stepback-error-localized-resampling","2026-11-08T13:20:00+00:00",{"id":93,"title":94,"news_slug":95,"published_at":96},"2f3e98d3-1168-470f-8cf3-6b0d09ee0ada","1280亿token实测:cosine调度输给1\u002F√t衰减","air-elr-schedule-weight-normalized","2026-10-31T21:12:47+00:00"]