[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-system-prompt-role-vectors-llm-steering":3,"topics-all":38,"news-related-67340a61-0b6f-4fac-b8a1-e430d206634e":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},"67340a61-0b6f-4fac-b8a1-e430d206634e","9.3%涨到32.0%:角色向量读得出,替不了提示词","arXiv 论文把「系统提示词到底有没有给模型派上角色」做成了可测量的问题:在 Qwen2.5-7B 与 Llama-3.1-8B 上提取职业角色向量,30 次独立抽取的同角色余弦相似度达 0.898 以上,连并购律师和公司法教授都分得开;但注入向量后角色命中率最多从 9.3% 拉到 32.0%,仍不及一句提示词。","系统提示词几乎成了每个 AI 产品的标配。一项被论文引用的分析统计了 1309 条系统提示词,超过 90% 里面写着角色或身份设定——「你是一位资深律师」「你是一个严谨的数据库管理员」。但提示词写了,模型就真的进入角色了吗?这篇论文([arXiv:2611.15989](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.15989))把它拆成了一个可测量的问题。\n\n## 角色确实存在,而且能读出来\n\n作者挑了六个职业角色:行政助理、运动员、并购律师、法学教授、通用教授、收银员。选法很讲究,法务与学术这几组是故意留下的难例——并购律师和公司法教授同属公司法领域但立场不同,通用教授和法学教授立场相近但专精不同。\n\n提取方法很直接:同一个问题,分别让模型在「带角色提示词」和「默认提示词」下回答,拿两组激活做差,差值就是这个角色在模型内部的方向。每个角色独立抽取 30 次,每次 90 个问题(15 个本行问题、75 个跨行问题,两类等权)。评分交给 gpt-4.1-mini 裁判,角色贴合度 0 到 3 分,连贯性 0 到 100 分。\n\n读出来的结果很干净。同角色两次抽取的余弦相似度,Qwen 上是 0.898 到 0.984,Llama 上 0.907 到 0.969;随机对半抽取再合并,余弦在 0.993 以上。难分的那几组也有真实间隔:并购律师与法学教授的夹角在 Qwen 上是 48.8 到 53.1 度,Llama 上 37.1 度;通用教授与法学教授是 54.4 到 58.1 度和 41.9 度。每一次抽取与自己角色中心点的余弦都高于与其他角色的余弦,比例是 100%。同一角色用本行问题和跨行问题分别建向量,同角色平均余弦 0.79(Llama 0.77),不同角色只有 0.37(Llama 0.55)。\n\n也就是说,角色在模型内部确实有独立表示,不只是一句提示词在输出端做戏。\n\n## 但注入之后,并没有更好用\n\n作者接着把方向加进残差流,看模型会不会更像那个角色。Qwen 上,温度 1.0、注入强度 0.3,角色表达成功率从 9.3% 升到 32.0%,提升 22.8 个百分点(每组 540 对答案,六个角色汇总)。角色特异性是 24.7 个百分点;让裁判在六个角色里盲猜哪个被 steer 了,猜中率 36.9%,未 steer 时 17.0%,随机基线 16.7%。同时替换成同尺寸随机方向,效果与未 steer 没有可辨差异——起作用的是这个具体方向,不是「往里加点东西」。\n\n问题是,和一句角色提示词比,向量就没那么能打了。Qwen 上 steer 减 prompt 的差是整体 −2.6 个百分点,本行问题上 −24.4 个百分点(55.6% 对 80.0%)。提示词在本行占优,论文给了一个没去验证的解释:角色提示词会明写这个角色怎么思考本领域。Llama 上差距更大,各个设置下提示词都领先大约 19 到 20 个百分点;Llama 的向量只在温度 0.0 时有效(+7.0 个百分点),温度 0.5 和 1.0 就与零不可区分了。\n\n## 30 轮之后,角色照样会掉\n\n多轮场景才是提示词的痛点。论文引用的另一项研究指出,系统提示里的指令大约在八轮对话内就会漂移。论文用 30 轮对话测角色的持久性(每个条件 18 组),Qwen 上所有条件都在衰减:纯 steer 的衰减速率与提示词不可区分(p=0.071),提示词加向量起步最高但掉得更快(p=0.006),最后落回提示词的水平。第 26 到 30 轮纯 steer 反而领先提示词 13.3 个百分点(p=0.044),但论文自己说这个窗口是看到首批结果之后才选的,只能算提示性证据。Llama 上纯 steer 全程低于提示词 8.5 个百分点,衰减更慢(p=0.049),可它第一轮就只有 11.1% 对 44.4%,属于「没那么多可掉」。\n\n## 真正有用的组合,是提示词加低强度向量\n\n把向量当补充而不是替代,结论就变了。低强度 0.15 叠加在提示词上,在五个保持连贯的设置里全部把成功率抬高 5.0 到 7.8 个百分点,提升主要来自跨行问题,角色特异性也涨 5.9 到 11.6 个百分点。代价是连贯性:Qwen 上从 95.3 掉到 91.7。\n\n还有一个反直觉的拆解。角色向量长度的 36% 到 84%(平方长度)躺在六个角色共享的一个方向上——那部分更像「助理感」,而不是具体职业。只保留角色特有的余量,steer 增益是 16.9 个百分点,共享部分只贡献 4.4 个百分点。\n\n机制上这条线不是孤例。Anthropic 的 [persona vectors](https:\u002F\u002Fwww.anthropic.com\u002Fresearch\u002Fpersona-vectors) 工作用的是同两个模型,同样靠对比带特质和不带特质的激活来提取方向;EMNLP 2025 Findings [另一篇独立工作](https:\u002F\u002Faclanthology.org\u002F2025.findings-emnlp.963\u002F)构造了 29 个角色向量,用激活相加和方向消融两种干预,发现能提升本行任务表现,还带出意外的跨行收益。后者衡量的是基准成绩而不是角色贴合度,两篇并不冲突,但共同说明「角色能压成一个方向」不是单篇论文的孤例。\n\n## 所以呢\n\n这论文真正扎人的地方不是「向量没用」,而是一个更别扭的结论:角色在模型内部可读,却不因此可控。你能把它测出来,不代表你能用它改。\n\n落到实践上两句话:多轮产品别指望用向量替代提示词,提示词的漂移是结构性的;想用向量,就用低强度当加成项,并且单独测它的连贯性成本。至于角色到底有没有被「装进去」,别只盯着输出像不像——论文自己的做法是同一条答案跑两个镜头,激活层和输出层都过一遍,两边对不上才算数。\n\n如果你在给产品设计「专家人设」,你会选更长的系统提示词,还是更硬的激活干预?\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.15989","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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"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},"2796bd28-f98b-4baf-98f0-d8978f267796","en","Role Vectors Are Real, but Steering Them Loses to Prompting","Role vectors from Qwen2.5-7B and Llama-3.1-8B are stable and separable, but steering them lifts role adherence from 9.3% to 32.0%, below prompting.","System prompts have become a fixture of nearly every AI product. One analysis cited by this paper counted 1,309 system prompts and found that more than 90% of them carry a role or identity instruction, something like 'you are a senior litigator' or 'you are a meticulous database administrator'. But writing the instruction down and having the model actually inhabit the role are two different things. This paper turns that fuzzy gap into something measurable.\n\n## The role really is there, and you can read it\n\nThe authors pick six occupational roles: an administrative assistant, an athlete, a corporate M&A lawyer, a law professor, a general professor, and a cashier. The selection is deliberate, because the legal and academic pairings are designed to be hard. An M&A lawyer and a corporations law professor share the corporate-law domain but differ in professional stance, while a general professor and the law professor share an academic stance but differ in specialization.\n\nThe extraction method is direct. For the same question, they let the model answer under a role system prompt and again under its default prompt, then take the difference between the two activation sets. That difference is the direction along which the role lives inside the model. Each role is extracted 30 times independently, with 90 questions per extraction, 15 in-domain and 75 out-of-domain, weighted equally. Scoring goes to a gpt-4.1-mini judge, with role adherence on a 0 to 3 scale and coherence on a 0 to 100 scale.\n\nThe readings are remarkably clean. Pairwise cosine between two extractions of the same role runs 0.898 to 0.984 on Qwen and 0.907 to 0.969 on Llama, and merging two random halves of the extractions agrees at cosine 0.993 or better. The deliberately hard pairs still separate: the angle between the M&A lawyer and the law professor is 48.8 to 53.1 degrees on Qwen and 37.1 degrees on Llama, while the general professor and the law professor sit at 54.4 to 58.1 degrees and 41.9 degrees. Every extraction is closer to its own role centroid than to any other, in 100% of cases. Building a vector from in-domain questions alone and one from out-of-domain questions alone, same-role pairs reach a mean cosine of 0.79 on Qwen and 0.77 on Llama, against 0.37 and 0.55 for different-role pairs.\n\nIn other words, the role has a genuinely separable internal representation. It is not a prompt doing theatre at the output layer.\n\n## But injecting it does not make the model better at the role\n\nThe authors then add the direction into the residual stream and see whether the model actually sounds more like the role. On Qwen, at temperature 1.0 and injection strength 0.3, role-expression success climbs from 9.3% to 32.0%, a gain of 22.8 percentage points over 540 answer pairs pooled across the six roles. Role specificity reaches 24.7 points. A judge asked to guess, blind, which of the six roles was steered identifies the right one in 36.9% of answers against 17.0% unsteered, with chance at 16.7%. Swapping in a random direction of the same size leaves success indistinguishable from unsteered, which is the useful control: it is the specific direction doing the work, not just adding a vector of some size.\n\nHere is where it gets awkward. Compared with a plain role prompt, steering loses. On Qwen the steered-minus-prompted difference is −2.6 points overall and −24.4 points on in-domain questions, 55.6% against 80.0%. The paper offers an untested explanation for the in-domain gap: a role prompt spells out how that role reasons about its own domain. On Llama the gap is wider, with prompting leading by roughly 19 to 20 points across settings, and Llama's vectors only work at temperature 0.0, a gain of 7.0 points, becoming indistinguishable from zero at 0.5 and 1.0.\n\n## Over 30 turns, the role fades anyway\n\nMulti-turn is where prompt fragility bites hardest. Another study cited by the paper reports that system-prompt instructions drift within about eight rounds of conversation. The authors measure persistence over 30-turn conversations, 18 per condition. On Qwen every condition declines. Steering alone declines at a rate not distinguishable from prompting (p=0.071), while prompt plus vector starts highest and falls faster (p=0.006), ending back at the prompted level. Over turns 26 to 30, steering alone leads prompting by 13.3 points (p=0.044), but the paper is explicit that this window was chosen after seeing the initial results and should be treated as suggestive. On Llama, steering alone stays below prompting across all 30 turns by 8.5 points, and its decline is slower (p=0.049), but it starts from 11.1% against 44.4% success at turn 1, so there is simply less room to fall.\n\n## The combination that does work: prompt plus a low-strength vector\n\nTreat the vector as an addition rather than a replacement and the conclusion flips. A low strength of 0.15 layered on top of the role prompt raises success rate by 5.0 to 7.8 points in all five settings that stayed coherent, with the gain concentrated in out-of-domain questions, where role specificity also rises by 5.9 to 11.6 points. The cost is coherence: on Qwen, mean coherence drops from 95.3 to 91.7.\n\nThere is also a counter-intuitive decomposition. Between 36% and 84% of each role vector's squared length lies along a single direction shared by all six roles, which reads more like a generic assistant register than a specific profession. Keeping only the role-specific remainder accounts for 16.9 points of steering gain, against 4.4 points from the shared part.\n\nThe mechanism is not a one-paper curiosity. Anthropic's persona vectors work uses the same two models, Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, and extracts directions the same way, by contrasting trait-exhibiting activations against trait-free ones. A separate EMNLP 2025 Findings paper builds 29 role vectors from activations and applies both activation addition and directional ablation, finding improvements in in-domain task performance along with unexpected cross-domain gains. That second study measures benchmark accuracy rather than role adherence, so the two are not in conflict, but together they suggest that compressing a role into a direction is a reproducible move rather than one lab's quirk.\n\n## So what\n\nThe uncomfortable conclusion is not that activation steering is useless. It is that a role being readable inside the model does not make it controllable. Being able to measure a thing does not mean you can steer it, and the paper is careful to separate the two lenses rather than let one stand in for the other.\n\nFor practice this means two things. Do not expect a vector to replace the system prompt in a multi-turn product, because prompt drift is structural rather than a tuning problem. If you do use vectors, treat a low strength as an additive bonus and measure the coherence cost on its own. And if you need to know whether a role actually landed, do not watch only whether the output sounds right: the paper's own method is to put every answer through two lenses, activation level and output level, and treat a mismatch as the finding.\n\nIf you were designing an expert persona for a product, would you reach for a longer system prompt or a harder activation intervention?\n","system-prompt-role-vectors-llm-steering","2026-11-13T11:30:00Z","2026-11-13T11:20:08.284038Z","2026-11-13T11:20:08.284059Z",true,"agent",170,[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},"42f5ea83-365c-4a4e-a318-17d7c4c11c15","论文实测:自洽性证书在AIME上24%的题上会误触发","self-consistency-certificates-misfire-correlated-reasoning","2026-11-21T09:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"78b52428-2dbb-4819-aa2a-eee5bccf7d5a","用大模型筛基准:基准反而变简单了","llm-validation-minimal-pair-benchmarks-easier","2026-11-11T11:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"e5c12c0b-5f79-47a4-becb-47d4e0c5d157","126万次盲投:模型没变,是人在变挑剔","chatbot-arena-evaluation-drift-pickier-judges","2026-11-11T07:05:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"409b5265-425f-4a26-a4a4-4a232f1cc1e0","自动指标看不见的失败:4个模型翻车实录","romanized-indic-generation-metric-blindness","2026-11-09T09:05:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"26602719-1a86-4264-812f-a57a67b966b7","1911大英百科对儿童语料:小模型排名反转","small-model-corpus-choice-skill-profile-reversal","2026-11-09T01:12:32+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"1b51020b-9a0c-46f8-98ec-e4e0c8c051a6","基准分其实是难度的加权平均:25 个主流基准的体检报告","benchmark-difficulty-measure-coverage-gaps","2026-11-01T01:16:42+00:00"]