[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-astar-thought-v2-latent-cot-compression":3,"topics-all":38,"news-related-3096df88-7158-4ffe-9356-1a83b829633b":57},{"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},"3096df88-7158-4ffe-9356-1a83b829633b","A*-Thought-V2:把思维链塞进隐空间,回复砍半,平均精度反升","arXiv 新论文把 CoT 推理轨迹投影到 3D PCA 空间,按每步与解题全局方向的夹角决定保留明文还是压成隐 token。在 Qwen3.5-9B\u002F3.6-27B 上,平均精度最高提升 2.6%,回复长度最多砍半,单位算力精度提升 2.29 倍。","长思维链(Chain-of-Thought)是这两年推理模型涨分的主要手段,但它的代价同样直白:每多想一步,就多烧一段 token,上下文窗口和推理账单同步膨胀。对部署侧来说,一段几百 token 的\"思考过程\"里,真正对答案有贡献的往往只是少数几步——问题是,哪些步骤可以扔?arXiv 9 月 7 日提交的新论文给出了一个相当几何化的答案。\n\n## 方法:给推理轨迹装一个\"方向罗盘\"\n\n论文的核心观察是:CoT 可以被看作隐藏状态空间里的一条轨迹。把问题、每一步推理、最终答案的表示都投影到一个 3D PCA 空间后,就能度量\"每一步的局部移动方向\"与\"从问题指向答案的全局方向\"之间的夹角。\n\n夹角小,说明这一步在直奔答案——保留为明文;夹角大,说明这一步在做检查、纠错或分支探索——压缩成连续的隐 token。模型最终输出的是\"明文与隐表示交错\"的序列,而不是把思考过程整个删掉或整个保留。论文还报告,夹角随时间的变化能反映探索、收敛、精化三个阶段:小夹角对应直接执行与答案形成,大夹角更常涉及检查、纠错和分支探索。\n\n## 训练:两个 forcing,一个管输入一个管标签\n\n要训动这个显隐交错的架构,论文引入了两个训练技巧:\n\n- **stepwise embedding forcing**:把每个冗余步骤池化成单个隐嵌入;\n- **label forcing**:对这个隐 token 用软的多峰词表分布做监督,而不是 one-hot 硬标签。\n\n论文的表征分析显示,隐状态形成了与文本状态明显区分的紧凑区域,而隐 token 位置更高的熵被认为反映了更宽的软目标,能鼓励更丰富的步骤级特征学习。\n\n## 数字:回复砍半,平均精度反升\n\n在 Qwen3.5-9B 和 Qwen3.6-27B 两个模型、六项域内加域外基准的实验里,论文报告:\n\n- 平均精度最高提升 2.6%;\n- 回复长度最多缩短一半;\n- 单位算力精度(Accuracy per Computation Unit)提升 2.29 倍;\n- 预处理时间下降 94.6%,训练时间最多下降 80.3%。\n\n代码开源于 GitHub(AI9Stars\u002FAStar-Thought),论文全文见 [arXiv:2609.07821](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.07821)。\n\n## 怎么看\n\n三点判断:\n\n1. **\"该留的留、该压的压\"比一刀切聪明。**硬剪枝丢中间信息,全保留烧钱;用方向一致性作为连续压缩的判据,思路干净,而且判据本身可解释——每一步\"偏航角\"有多大,是能算出来的量。\n2. **训练侧的数字值得注意。**预处理时间降 94.6%,说明数据构造流程在工程上是轻量的,不是只在论文里跑得通的玩具。\n3. **留意\"up to\"措辞。**2.6% 和\"砍半\"都是峰值数字,六项基准内部的分布要看论文细节;真要上生产,建议先拿自己的负载压测一遍再定。\n\n对推理成本敏感的团队,这类隐空间压缩方向值得放进观察列表:当模型\"想得太多\"成为账单主体时,能判断\"哪一步不值得明文展开\"的模型,就是省钱的模型。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.07821","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},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"75a3a8e2-15dc-4d69-81ec-d83ecc9998b7","en","A*-Thought-V2: Latent CoT Compression Cuts Responses in Half","arXiv paper compresses off-direction CoT steps into latent tokens. Qwen3.5-9B\u002F3.6-27B: accuracy up 2.6%, replies halved, 2.29x accuracy per compute unit.","Long Chain-of-Thought (CoT) is how reasoning models have been gaining score, but the cost is blunt: every extra thinking step burns tokens, and both the context window and the inference bill inflate together. For deployment teams, only a fraction of a several-hundred-token \"thinking process\" actually contributes to the answer. The question is which steps can be dropped. A paper submitted to arXiv on September 7 offers a distinctly geometric answer.\n\n## The method: a compass for the reasoning trajectory\n\nThe core observation is that CoT can be treated as a trajectory in hidden-state space. After projecting the representations of the question, each reasoning step, and the final solution into a 3D PCA space, the framework measures the angle between each step's local transition and the global question-to-solution direction.\n\nA small angle means the step is heading straight for the answer — it stays as explicit text. A large angle means the step is checking, correcting, or exploring a branch — it gets compressed into a continuous latent token. The model outputs an interleaved explicit-implicit sequence instead of deleting or keeping the whole thought process. The paper also reports that the temporal variation of these angles reveals exploration, convergence, and refinement stages: small angles correspond to direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration.\n\n## Training: two forcings, one for input and one for labels\n\nTo train this explicit-implicit architecture, the paper introduces two techniques:\n\n- **Stepwise embedding forcing** pools each redundant step into a single latent embedding;\n- **Label forcing** supervises that latent token with a soft multi-modal vocabulary distribution instead of a hard one-hot label.\n\nRepresentation analyses show latent states form a compact region distinct from textual states, and the higher entropy at latent-token positions is interpreted as reflecting broader soft targets that encourage richer step-level feature learning.\n\n## The numbers: half the response, higher average accuracy\n\nExperiments on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks report:\n\n- average accuracy improved by up to 2.6%;\n- response length reduced by up to half;\n- Accuracy per Computation Unit increased by 2.29x;\n- preprocessing time reduced by 94.6% and training time by up to 80.3%.\n\nCode is open-sourced on GitHub (AI9Stars\u002FAStar-Thought); the paper is at [arXiv:2609.07821](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.07821).\n\n## How to read it\n\nThree takeaways:\n\n1. **Keep-what-matters beats prune-everything.** Hard pruning loses intermediate information; keeping everything burns money. Using directional alignment as a continuous compression criterion is a clean, interpretable design — the \"off-course angle\" of each step is a computable quantity.\n2. **The training-side numbers matter.** A 94.6% cut in preprocessing time means the data-construction pipeline is lightweight in engineering terms, not a paper-only toy.\n3. **Watch the \"up to\" phrasing.** The 2.6% gain and the halved length are peak numbers; the distribution across the six benchmarks needs a closer read. Before production, stress-test on your own workloads.\n\nFor teams sensitive to inference cost, this latent-compression direction belongs on the watch list: when a model \"thinking too much\" becomes the bulk of the bill, a model that knows which step doesn't deserve explicit text is the money-saving model.","astar-thought-v2-latent-cot-compression","2026-09-09T15:10:00Z","2026-09-09T15:14:30.923380Z","2026-09-09T15:14:30.923390Z",true,"agent",135,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"5c75bd80-9f12-499a-898f-019615ac98ee","Prefix Sliding:让推理模型长思考提速3倍的免训练方案","prefix-sliding-efficient-test-time-scaling","2026-08-27T17:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"8a42c9c3-a1c7-40fb-8c75-8ac42977b5af","D-cut 把投机解码的「长草稿」剪掉一半：高并发推理平均提速 1.65×、MoE 跑出 3×","d-cut-speculative-draft-cut","2026-07-18T10:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"21be10ba-3364-4b55-a248-e6f878dae68b","QuasiMoTTo：quasi-Monte Carlo 进 test-time scaling","quasimotto-quasi-monte-carlo","2026-07-02T10:20:52+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"a453eb28-7fb0-4e07-adc1-0d0575850758","EntMTP 用熵信号给多 token 推测装上调速器：让 LLM 自适应匹配上下文可预测性","entmtp-entropy-speculative-decoding","2026-06-29T12:21:51+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9a3cd449-e29a-4730-814b-f1be5c2685c6","复旦FFD让Flash Attention退役？11.6× kernel提速把长上下文推到256K","fudan-ffd-long-context-attention-sparsity","2026-09-15T07:15:46+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"2638aeac-dc4d-4b73-b7fe-2b042015adee","OreoLook 开源:三层缓存把 AI 搜索搬进 8 核 CPU,重复问题 0.1 毫秒出答案","oreolook-three-layer-cpu-cache","2026-09-10T23:08:36+00:00"]