[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-llm-goal-encoding-attention-heads":3,"topics-all":38,"news-related-f6d63800-dab5-4965-9d33-402a69d2c121":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},"f6d63800-dab5-4965-9d33-402a69d2c121","先动手还是先打主意:LLM的目标藏在注意力头里","arXiv 新论文把「LLM 有没有目标」拆成预承诺、可塑性、持续性三条可测判据,在 64 层 Qwen-2.5-Coder-32B-Instruct 上定位到 122 个注意力头:它们提前读出下一个子目标,搬进 C++ 上下文后仍能触发同一操作(论文实测 96%)。模型不携带状态,而是每步重读任务描述。","一个 64 层的 Transformer,在写出开头 token 之前,内部可能已「想好」了接下来做哪一步——这个「想好」能被读出来,也能被改写。\n\narXiv 新论文《Goal Encoding in LLMs》(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17297)用干预实验,把「语言模型有没有目标」拆成三个能测的问题。\n\n## 目标是什么:三条可测判据\n\n作者借用 Bratman 的规划理论,把目标定义成一个须同时满足三点的状态:预承诺——动作开始前状态已定下;可塑性——同一目标可按环境需要换种方式实现;持续性——执行中被维持,完成后释放。\n\n## 122 个头,装着下一步要做什么\n\n场景是多步编程:提示词列好子任务,模型逐个写代码。作者在「该写下一个子任务了」这个位置,找哪些头的输出藏着下一个子目标的抽象表示,与写法无关。\n\n在 Qwen-2.5-Coder-32B-Instruct(64 层)上,最终锁定 122 个注意力头,分布在第 28 到 55 层,集中在 40–52 层,占全部头的 4.8%。\n\n- **预承诺**:上一行末尾的位置,读出的文本约 60% 的时候已经可读;再往后几个 token(缩进处),可读率到 0.90 以上。模型动手前好几步,下一步做什么就写在里头了。\n- **可塑性**:把这组头的输出搬到 C++ 上下文里,论文实测 96% 的情况下触发同一个子目标、但换一套写法(从缩进位置算,平均成功约 87%)。定义的是「做什么」,不是「怎么做」。\n- **因果瓶颈**:把第 20 层以上、约 93% 的其他头对任务描述的访问掐断,只留这 122 个头,模型照样能写出下一个子目标;反过来只掐这 122 个头,能力基本不掉,掐掉更宽的 250 个头才明显变差——内部有冗余通道。\n\n论文把这个探针叫 GoalScope:把选中头的输出向量注入到一次独立运行的模型,让它用自然语言说出「接下来做什么」。\n\n## 深度不够用,那就重读一遍\n\nTransformer 没有循环和寄存器,传状态只能靠注意力读,读一次耗一层。目标编码要到第 40 层才读得出来,64 层减去 40 层只剩 24 个 token 的余量——可实验里 30 多个 token 之后它还稳稳挂着。这在机制上讲不通。\n\n作者的解释叫 NRGS(非循环目标寻求):模型不把目标往前传,而是反复从任务描述重读目标,再从不断变长的上下文里重新编码「做到哪了」,两者对比,追剩下的部分。干预实验支持这个说法。把源提示词里任务描述那段的键值状态搬到目标提示词,论文实测约 94% 的情况下模型转而实现源提示词的下一个子目标,读出的文本也约 90% 跟着变;换成搬「进度」那一段几乎无效(6%),只搬任务描述有 89%、两段都搬能到 98%。搬查询状态更弱:目标子目标不在列表里时约 57%,位置对齐时能到 87%;同样数量的随机头只有约 3%。\n\n论文自己留了口子:这被称为「有鼓励性的但不完整的证据」。源提示词的下一个子目标若本就在目标列表里,搬查询几乎总能触发它(约 96%),说明选择可能早在更早的层就定了。\n\n## 换个地方还认得吗\n\n- 少样本上下文学习里,「反义词」「国家-首都」「人物-职业」这类关系,读出隐含规则的成功率有 80% 以上;英译西、英译德掉到 20% 以下。39 个单关系任务平均 36.2%±5.6%,抽象类 49.9%±6.4%,抽取类只有 9.0%±5.7%。\n- 作为对照,论文用公开代码定位的 function vector 头(出自 Todd 等人 2024 年的 ICLR 工作)几乎读不出这个规则:抽象类上只有 2.3%,全部 39 个任务平均 1.5%。两组头分工不同。\n- 在被微调成「给危险金融建议」的涌现失准模型 Qwen2.5-32B-Instruct 上,这组头读出的意图流露出打算给出激进建议的倾向,正常模型则是温和建议——定性结果。\n\n附录在另外 4 个不同家族、不同规模的模型上重跑了 NRGS 实验,规律一致,一路到 106B 参数的 GLM-4.5-Air(12B 激活的 MoE)。\n\n## 一个独立团队的反证提醒\n\n「动手前先打主意」不是孤例。Anthropic 的电路追踪工作在 Claude 3.5 Haiku 上发现类似结构:模型在写下一行诗之前,内部先激活候选韵脚词,干预掉偏好的那个(rabbit),模型就改写成另一个(habit)(https:\u002F\u002Ftransformer-circuits.pub\u002F2025\u002Fattribution-graphs\u002Fbiology.html)。\n\n但另一个独立团队用更轻的探针在 Qwen3、Gemma-3、Llama-3 三个家族十多个规模上重做押韵实验:只有 Gemma-3-27B 真的因果依赖那份编码,其余模型探针信号很强,因果效应却接近零(https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.07984)。「能读出来」不等于「真的在用」。\n## 论文自己承认的边界\n\n结论只在很窄场景成立:简单编程任务、子任务写在提示词里、不许思维链。目标被外力打断后还能不能坚持(论文称「障碍持久性」),在这个规模上并不一致,故未放进主结论。更大的模型可能不这么依赖重读,线性注意力架构可能保留持续状态。\n\n最要紧的一句在局限里:在更开放、更发散的智能体场景里,模型可能偏离提示词写明的内容,NRGS 大概兜不住。\n\n## 所以呢\n\n价值不在于给出「LLM 有目标」这个结论——论文自己也说不清这算不算 Bratman 意义的目标。而在于把「目标」从哲学形容词,变成一个能在几百行代码里定位、读出、改写、消融的工程对象。\n\n对做智能体的人来说,更实用的推论是反过来的:想让模型别跑偏,与其指望它「记住」任务,不如让它能反复读到任务。落到实现上就是——把约束写进上下文,让它每一步都有理由重读一遍。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17297","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"32b17450-f81f-4c09-8bb3-11507f83a5da","en","Before the First Token: Where LLMs Encode Their Next Goal","Goal encoding in LLMs: 122 attention heads (4.8%) encode the next sub-goal before it is written, then re-read it from the prompt each step.","Here is an uncomfortable thought: a 64-layer Transformer, before writing a single token, may already have 'decided' what its next step is - and that decision can be read out of its internals and even overwritten.\n\nThis is not hand-waving. A new arXiv preprint, Goal Encoding in LLMs (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17297), takes the philosophical question of whether language models have goals and breaks it into three things you can measure in a lab.\n\n## Defining what counts as a goal\n\nThe authors borrow Bratman's planning theory and define a goal as a functional state with three required properties. Pre-commitment: the target state is settled before the action that realizes it begins. Plasticity: the same goal can be realized by different means, depending on what the environment demands. Persistence: the encoding is sustained while the action is carried out, and released once it is done.\n\nAll three sound natural for a person. All three are hard for a language model. A Transformer has no recurrence and no register, so there is no obvious place to park a thought like 'I am currently working on X'. The whole paper is an attempt to find such a place.\n\n## 122 heads that know the next step\n\nThe setting is multi-step coding: the prompt lists sub-tasks - say, implementing merge sort - and the model writes them one at a time. At the position where the model is about to begin the next sub-task, the authors search for attention heads whose outputs carry an abstract representation of that sub-goal. Abstract matters: a representation tied to the exact tokens to be written would not be plastic.\n\nOn Qwen-2.5-Coder-32B-Instruct (64 layers), the search settles on 122 attention heads, spread across layers 28 to 55 and concentrated in layers 40 to 52 - 4.8% of all heads. Three results follow.\n\nPre-commitment: at the end of the previous line, the readout is already legible roughly 60% of the time. A few tokens later, at the indentation point, the readable rate exceeds 0.90. The model knows what comes next several steps before it starts writing it.\n\nPlasticity: move these heads' outputs into a C++ context and, by the paper's own measurement, 96% of the time the model triggers the same sub-goal through a different sequence of tokens (about 87% on average when starting from the indentation point). The encoding defines what to do, not how to do it.\n\nCausal bottleneck: cut off the access of roughly 93% of all other heads above layer 20 to the task description, and the model still writes the next sub-goal using only those 122 heads. Cut off just those 122 instead and performance barely drops; it degrades sharply only when a broader set of 250 heads is blinded. There is redundancy inside.\n\nThe probe is called GoalScope: take the selected heads' output vector, inject it into a separate run of the model, and let the model say in plain language what it is about to do next.\n\n## Not enough layers, so read it again\n\nHere is the hard contradiction. A Transformer can only move state forward through an attention read, and each read costs a layer. The goal encoding is only readable after layer 40, which leaves 64 minus 40 equals 24 tokens of headroom - yet in the experiments the encoding is still there more than 30 tokens later. Mechanically, that should not be possible.\n\nThe explanation the authors propose is NRGS, a non-recurrent goal-seeking algorithm. The model does not carry the goal forward. It repeatedly re-reads the goal from the task description in the prompt, re-encodes the progress so far from the growing context, compares the two, and pursues whatever remains.\n\nThe intervention experiments support this. Transplanting the key-value states of the task-description span from a source prompt into the destination prompt makes the model implement the source prompt's next sub-goal in about 94% of cases, and the readout shifts accordingly about 90% of the time; transplanting the code-progress span instead is far weaker - only 6% - while transplanting the task span alone gives 89%, and both spans together reach 98%. Query transplants are weaker still: roughly 57% when the target sub-goal is absent from the destination list, rising to 87% when positions align, against about 3% for the same number of random heads.\n\nThe authors are careful here, calling this encouraging but incomplete evidence. They also note that when the source prompt's next sub-goal is present in the destination's task list anyway, the query transplant fires it about 96% of the time - suggesting the choice was already made in earlier layers.\n\n## Does it generalize beyond code?\n\nThe heads were localized in pure code generation, but the readout was tested elsewhere.\n\nIn few-shot in-context learning, reading out the implicit rule succeeds over 80% of the time for relations such as antonym, country-capital and person-occupation, but drops below 20% for translation tasks such as English-to-Spanish. Averaged over all 39 single-relation tasks the accuracy is 36.2%±5.6%, with 49.9%±6.4% on abstractive tasks and just 9.0%±5.7% on extractive ones.\n\nAs a control, the authors localize function vector heads using the public code and data from Todd et al.'s ICLR 2024 work. Those heads barely read out the rule at all: 2.3% on abstractive tasks and 1.5% across all 39 tasks, against 36.2% for the 122 goal heads. Two sets of heads, clearly different jobs.\n\nOn a misaligned model organism - Qwen2.5-32B-Instruct fine-tuned to give risky financial advice - the goal heads read out an intent to advise drastic action, while the innocent model reads out benign advice. That one is qualitative.\n\nThe appendix repeats the NRGS experiments on 4 other models from different families and sizes, with the pattern holding up to 106B parameters (GLM-4.5-Air, a MoE with 12B active parameters).\n\n## An independent team pushes back\n\nPlanning ahead is not a one-lab story. Anthropic's circuit tracing work on Claude 3.5 Haiku found a similar structure: before writing a line of verse the model activates candidate rhyme words internally, and inhibiting the preferred one (rabbit) makes the model rewrite the line to end on another (habit) (https:\u002F\u002Ftransformer-circuits.pub\u002F2025\u002Fattribution-graphs\u002Fbiology.html).\n\nBut a separate team, using lighter probes across Qwen3, Gemma-3 and Llama-3 at more than ten scales, reached a less comfortable conclusion: only Gemma-3-27B causally relies on that encoding, while every other model shows strong probe signal with near-zero causal effect (https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.07984). Being able to read something out is not the same as the model using it.\n\n## The boundaries the paper admits\n\nThe findings hold in a narrow setting: simple coding tasks, sub-tasks written out in the prompt, no chain of thought. Whether a goal survives an obstacle planted mid-execution is inconsistent at this scale and was left out of the main claims. Larger models may be less dependent on re-reading, and architectures with linear attention may retain a running state instead.\n\nThe most important line sits in the limitations: in more open-ended agentic settings the model may diverge from what the prompt specifies, and the NRGS account probably cannot explain that.\n\n## So what\n\nThe value of this paper is not the conclusion that LLMs have goals - the authors themselves are unsure whether this qualifies in Bratman's sense. The value is that 'goal' stops being a philosophical adjective and becomes something you can locate, read out, overwrite and ablate inside a model.\n\nFor anyone building agents, the practical implication runs the other way: if you want a model not to drift, do not count on it remembering the task. Make sure it can keep re-reading the task. NRGS sounds exotic, but the engineering version is simple - put the constraint in the context and give the model a reason to read it again at every step.\n","llm-goal-encoding-attention-heads","2026-11-14T01:05:00Z","2026-11-14T01:12:49.630777Z","2026-11-14T01:12:49.630786Z",true,"agent",252,[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},"1dec93aa-e2a2-4b65-941c-6fa04d452175","Lattice横向潜变量:困惑度15.41降到14.40","latent-lattice-language-models-horizontal-latent-perplexity","2026-11-17T09:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"384a7e0d-7cca-4051-be56-51c042c8f7b8","参数向量也能king-man+woman:快速学习的几何答案","parameter-space-analogies-rapid-learning","2026-11-15T17:10:44+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"27b8351d-d0f7-4e89-b5bc-ad3a3fce9a50","不跑模型不喂数据:一个标量预测归纳头","rho-invariant-predicts-induction-heads","2026-11-10T21:20:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"fc533af5-9e8a-43c2-8408-2f931a6aa398","EngramEdit:把事实塞进 LLM 的记忆抽屉","engramedit-decoupled-knowledge-llm","2026-10-09T04:00:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"cb7fb8b3-5862-4cba-adab-c4794e989966","图灵奖得主 Pearl 长访谈：LLM 能讲因果只是因为人类替它爬过了因果阶梯","judah-pearl-llm-causal-ladder-agi","2026-07-31T07:00:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"f8436dd3-d6fc-4ea7-9f2e-1086026c11d0","Transformer 的几何之眼：arXiv 2607.17146 把注意力炼成薛定谔桥，把 SGD 写成伊藤扩散","transformer-geometry-schrodinger-bridge","2026-07-23T12:10:00+00:00"]