[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-cot-hidden-state-directions-direct-answers":3,"topics-all":38,"news-related-4544ed60-b5fb-41b6-9fee-dd2f59012547":67},{"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},"4544ed60-b5fb-41b6-9fee-dd2f59012547","隐状态转个方向:直答46到85,超CoT省九成token","Heriot-Watt团队把CoT隐状态按patch分成recovering与failed两类,用两类均值构造每题专属方向,把直答提示最后一个隐状态旋转一次:25个任务上LLaMA 3.1 8B从46到85(CoT为69),Qwen 2.5 7B从52到86(CoT为77),每题19个token,约为CoT十分之一。","8B 级开源模型答推理题,现在有个默认姿势:先写一长串思考链再给答案。爱丁堡 Heriot-Watt 大学团队的新论文(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.24598)把这笔账重算了一遍:LLaMA 3.1 8B 在 25 个推理任务上直答平均 46 分,开 CoT 拉到 69,代价是每题约 197 个生成 token;Qwen 2.5 7B 是 52 分、77 分、245 个 token。不写链子,只在生成前把直答提示的最后一个隐状态朝一个方向旋转一次,两个模型的平均分分别冲到 85 和 86,每题只花 19 个 token。\n\n## 先给思考链的隐状态贴标签\n\n做法分两步。第一步,把 CoT 过程里每个 token 的隐状态单独取出来,移植到同一题直答提示的末位:移植后模型答对了,这个状态记为 recovering,答错记为 failed。在 LLaMA 3.1 8B 上平均只有 17% 的 CoT 状态能救回答案,Qwen 2.5 7B 是 28%。一个线性探针就能把两类分开:25 个任务平均 AUROC 达 0.84(LLaMA,区间 0.71 到 0.92)与 0.80(Qwen,区间 0.52 到 0.93)。直答几乎全错的 Multistep Arithmetic 任务上(直答正确率 0.4%),探针仍有 0.88。探针还跨任务迁移:600 对任务组合里 587 对高于随机,同类型任务之间迁移更突出(LLaMA 0.71,不同类型 0.65)。作者由此判断,模型对这两类状态的编码本身就不一样。\n\n## 造一个方向,旋转一次\n\n第二步,对每道题求 recovering 状态均值与 failed 状态均值,用前者减去 0.5 倍后者,归一化后得到这题专属的「解题方向」,再用 SLERP 把直答提示最后一个隐状态朝它旋转一次,范数保持不变,位置在 LLaMA 第 29 层、Qwen 第 25 层。两个均值隔得并不远(平均夹角 27 度,Qwen 21 度),所以论文扫了权重:failed 均值的系数从 0 加到 0.5,准确率一路升到顶,加到 0.75 反而回落——离 failed 远了,也偏离了 recovering。结果:LLaMA 直答 46 提到 85,Qwen 从 52 到 86,分别压过各自 CoT 的 69 和 77,比直答高 39 和 34 分。25 个任务里,方向在 24 个(LLaMA)和 19 个(Qwen)任务上压过 CoT,Qwen 输的几个里有三个是 CoT 本来就 90 分往上的强项。两组对照说明这不是玄学:朝 failed 均值方向旋转,准确率掉到 9 和 16;换成全任务共享的平均方向,准确率原地踏步(46 与 50)。方向是每题专属的,平均掉就失效。\n\n## 三个反直觉点\n\n一是 CoT 反而拖后腿的任务,方向照样涨分:Movie Recommendation、Ruin Names 两个任务上 CoT 比直答还低,加方向后仍比直答高 22 到 36 分。二是长度自适应:答案能从题目里读出的任务,模型 3 到 5 个 token 就作答;要计算的任务自动写短步骤,GSM8K 上 39 个 token 对 CoT 的 179,Qwen 上是 52 对 248,Multistep Arithmetic 两模型分别是 55 对 140、66 对 260。三是 CoT 和直答都答错的题,方向还能救回 64%(LLaMA)与 70%(Qwen)。Qwen 上探针最弱的一档(Causal Judgement,AUROC 0.52),方向也把它抬到 86 分。论文图 1 的例子:一道多步算术题,直答错成 -40,CoT 自己也错成 80,旋转后的直答写下四行短步骤给出 48;另一道 GSM8K 题直答错成 121,旋转后写出「$21+$50=$71」。\n\n## 机制不是孤证\n\n「模型内部算的比说出来的多」这条机制线,有独立团队的工作佐证:Orgad 等人(https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.02707)发现模型内部表征对输出真假编码的信息远超此前认识;Deng 等人(https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.14838)则证明显式思考链可以逐步内化,不生成链子也能推。需要说明,本篇全部分数是论文自报的实测值,没有第三方复现。\n\n## 冷水也得泼\n\n方向的构造依赖这道题自己的 CoT 和 patch 标签,拿到新题就没法直接用:论文试过训一个回归器从直答直接预测方向,准确率停在直答水平。实验只覆盖 7B 和 8B 两个指令微调模型,任务是短答案、可自动判分的那一类,干预也只做在尾部三层、直答末位一个位置。另外这是同团队的延续工作:前作(EMNLP 2026 Findings)只在 GSM8K 上做过 patching 分析,本篇是它的扩展,不是独立复现。代码仓库已放出(https:\u002F\u002Fgithub.com\u002Fhmehrafarin\u002Fhidden-state-directions)。\n\n论文的落点很直白:按他们的实测,CoT 不是通往这个准确率的最短路径,收集到的「题目-方向」数据对,可以拿去训练直接预测方向的模型。真到那天,推理账单上思考链那一项,理论上有九成可省。你觉得这是把思考链蒸馏进了权重,还是又一次只在 7B 档位成立的漂亮实验?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.24598","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"b22dede3-3787-4705-a035-09b26261787e","en","One hidden-state rotation: direct answers 85 vs CoT 69","Rotating one hidden state lifts direct answers to 85 (CoT: 69) on LLaMA 3.1 8B and 86 on Qwen 2.5 7B across 25 tasks, at 19 tokens per question.","A small open-weight model answering a reasoning question has a default move: write a long chain of thought (CoT), then answer. A new paper from a Heriot-Watt University team in Edinburgh (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.24598) re-does that trade-off: on 25 reasoning tasks, LLaMA 3.1 8B answers directly at a mean accuracy of 46 and with CoT at 69, at a cost of about 197 generated tokens per question; Qwen 2.5 7B sits at 52, 77 and 245 tokens. Rotate the last hidden state of the direct prompt towards one direction, once, before generation, and the mean scores reach 85 and 86 at 19 tokens per question.\n\n## Labelling the hidden states of the chain\n\nThe method has two steps. Step one: take each token's hidden state from the CoT run and patch it into the last position of the direct prompt for the same question. If the model then answers correctly, the state is labelled recovering; otherwise failed. On average only 17% of CoT states recover the answer on LLaMA 3.1 8B, and 28% on Qwen 2.5 7B. A linear probe separates the two classes: mean AUROC 0.84 (LLaMA, range 0.71 to 0.92) and 0.80 (Qwen, range 0.52 to 0.93). On Multistep Arithmetic, where the direct prompt answers correctly in 0.4% of cases, the probe still reaches 0.88. The probes also transfer across tasks: above chance on 587 of the 600 pairs of tasks, most strongly between tasks of the same reasoning type (0.71 versus 0.65 across types on LLaMA). The authors conclude that the model encodes the two classes differently.\n\n## Build a direction, rotate once\n\nStep two: for each question, average its recovering states and its failed states, subtract half of the failed mean from the recovering mean, and normalise, giving the question's own task-solving direction. The last hidden state of the direct prompt is then rotated onto this direction once with SLERP, keeping its norm, at layer 29 for LLaMA and layer 25 for Qwen. The two means sit close together (about 27 degrees apart on LLaMA, 21 on Qwen), so the paper sweeps the weight on the failed mean: accuracy climbs as it rises to 0.5 and falls again at 0.75, moving away from the failed mean but also from the recovering mean. Results: direct answers rise from 46 to 85 on LLaMA 3.1 8B and from 52 to 86 on Qwen 2.5 7B, above their CoT scores of 69 and 77, a gain of 39 and 34 points over the direct prompt. The direction beats CoT on 24 of the 25 tasks for LLaMA and 19 for Qwen; three of Qwen's losses are tasks where its CoT already scores 90 to 97. Two controls show this is not a trick: rotating towards the failed mean drops accuracy to 9 and 16, and replacing the per-question direction with a task-level average leaves accuracy at the direct-prompt level (46 and 50). The direction is specific to each question.\n\n## Three counterintuitive points\n\nFirst, on tasks where CoT hurts (Movie Recommendation and Ruin Names for both models), the direction still gains 22 to 36 points over the direct prompt. Second, length adapts to the task: where the answer can be read from the question, the model answers in 3 to 5 tokens; where it must be computed, the model writes short steps, 39 tokens on GSM8K against 179 under CoT for LLaMA (52 against 248 for Qwen), and 55 against 140 and 66 against 260 on Multistep Arithmetic. Third, on questions that both the CoT and direct prompts get wrong, the direction recovers the correct answer on 64% (LLaMA) and 70% (Qwen). On Qwen's weakest probing task (Causal Judgement, AUROC 0.52), the direction still lifts accuracy to 86. In the paper's Figure 1, a multistep arithmetic question answered directly as -40, with the CoT itself answering 80, yields the correct 48 after the rotation; on a GSM8K question answered directly as 121, the model writes '$21+$50=$71'.\n\n## The mechanism is not a one-off\n\nThe claim that models compute more internally than they say has independent support: Orgad et al. (https:\u002F\u002Farxiv.org\u002Fabs\u002F2410.02707) find that internal representations encode far more information about output truthfulness than previously recognised, and Deng et al. (https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.14838) show that explicit CoT can be internalised step by step, so chains need not be generated. All scores in this article are the paper's own measurements; there is no third-party replication.\n\n## The cold water\n\nBuilding a direction needs the question's own CoT and patch labels, so it cannot be applied to a new question as is: the authors tried a regressor that predicts the direction from the direct prompt alone, and accuracy stayed at the direct-prompt level. The study covers two instruction-tuned models of 7B and 8B, tasks with short, automatically checkable answers, and interventions at three late layers and one position. It is also a continuation of the same team's work: their previous paper (Findings of EMNLP 2026) ran the patching analysis only on GSM8K; this one extends it, and is not an independent replication. The code repository is public (https:\u002F\u002Fgithub.com\u002Fhmehrafarin\u002Fhidden-state-directions).\n\nThe takeaway the authors put forward: by their measurements, CoT is not the shortest route to this accuracy, and the collected question-direction pairs can train a model to predict the direction directly. If that day comes, the CoT line on the inference bill could in theory shrink by about nine tenths. Is this distilling the chain into the model, or another elegant experiment that only holds at the 7B tier?","cot-hidden-state-directions-direct-answers","2026-11-19T13:30:00Z","2026-11-19T13:22:05.519862Z","2026-11-19T13:22:05.519877Z",true,"agent",108,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"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":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"217f417d-1b9c-475b-99f4-e21e7c909711","MHAR 把 Transformer 残差流从「单车道」拆成 H 条独立路由:子空间第一次有权自己挑历史层","multi-head-attention-residuals-mhar","2026-08-01T07:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"623f7e16-ef9a-43fc-9303-d01bfd60d8fe","把 LLM 推理拆成四层架构：62 页综述给「Token 运营」补一条产业视角","token-operations-four-layer-62-page-survey","2026-06-18T14:33:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"4e43e35d-a808-4125-be31-69cadedc61f1","PoLar 把 LLM 层变成可调积木：动态跳层+复读，3B 模型数学推理涨 60+ 个百分点","polar-icml-2026-3b-math-62pp-jump","2026-06-15T14:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"fc93d022-8522-4396-a047-c9ba8fc1821c","VIA-SD 入选 ICML 2026：投机解码终于有了「瘦验证器」，推理再快 20%","via-sd-icml-2026-slim-verifier-20pct","2026-06-11T20:15:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"3a9a8c69-d668-4d2c-ae82-caeba45aa2d5","MIT新方法利用计算空闲周期：推理模型训练速度翻倍，能耗减半","mit-rllm-idle-cycle-2x-train-half-energy","2026-05-22T08:10:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"2e1d1723-4cea-4621-965e-9514d08a9013","LLM推理服务正在淘汰「启发式」：运筹学视角下的新优化范式","llm-inference-or-paradigm-heuristics","2026-05-16T08:25:00+00:00"]