[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-alignment-generalization-value-representations-2610-12410":3,"topics-all":35,"news-related-98341c63-0ea6-487e-958e-15400f1bfb1a":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"98341c63-0ea6-487e-958e-15400f1bfb1a","价值表征替代文本描述:LLM 对齐泛化预测新基准","arXiv 2610.12410 提出 alignment generalization prediction 任务,在 66 个对齐价值上验证:基于模型激活的表征与泛化矩阵的相关性达 0.45,文本描述基线只有 0.05,差距近 9 倍。文章还给出首个 LLM 价值分类法。","把\"诚实\"、\"helpful\"、\"harmless\"、\"尊重隐私\"这些词写进训练目标,看似已经讲清楚了模型应该朝哪个方向走;但凡是认真迭代过对齐策略的团队都会撞上一面墙——你盯着一个价值训,旁边那些价值被悄悄动了。\n\n更麻烦的是:评测套件看起来还是绿的,因为它在检查目标行为,没在检查副作用。要么\"对齐\"在表面成功、在深层把相邻行为拖偏;要么模型学到了答案的形状,却没学到答案背后的态度。这种 spillover 之前只能事后发现,没法提前预判。\n\n## 66 个价值、9 倍差距\n\n2026 年 10 月 8 日提交到 arXiv 的 2610.12410「Predicting Alignment Generalization with Value Representations」一上来就把这件事命名成 alignment generalization prediction:给定一个价值去微调模型,能不能预测它对其他 66 个未训练价值的影响?作者来自 CMU、Mila、ETH Zurich、UBC 等机构(Liu、Bhatia、Stanczak、Diab、Shwartz、Fried),用现代对齐训练目标里挖出的价值条目做底料,把每一对价值之间的关系构建成一张泛化矩阵。\n\n核心结果意外又不意外。基于模型激活的表征(value representations based on model activations when applying values in context)与泛化矩阵的相关性达到 0.45;而用文本描述价值去拟合的基线只有 0.05。一个干净的 9 倍差距。Ken Ashe 在 10 月 9 号的评论文章里把它翻译成更直白的版本:写出来的价值是标签,激活里跑出来的价值才是模型实际在用的价值。\n\n这一刀切在了一个长期被绕开的地方——\"be honest\" 这种短语写在政策文件里很干净,但在模型内部,\"诚实\"会和不确定性表达、拒绝风格、引用行为、顺从度、信心校准、反驳用户的意愿全部纠缠在一起。文本描述把这种结构压扁了,激活保留得多一些。这也解释了为什么 eval 容易过拟合——模型可以学到答案的形状,而不学到底层行为。\n\n更值得关注的是后面两段延伸。论文用这些预测能力去量一个多值对齐目标里各价值之间的相似度,发现这种相似度跟模型 robustness 显著相关。换句话说,如果你写的对齐目标里两个价值在表征空间里实际挨得很近,模型就更脆弱。\n\n此外作者给出首批证据,显示存在一个跨模型共享的价值空间(shared, model-independent value space),并以此为基础构建了第一个 LLM 价值分类法——这个分类法不是写出来的,是测出来的。Ken Ashe 在评论里提醒说\"initial evidence\" 这四个字分量很重,但方向有意义:如果价值有可量化的几何结构,alignment 就会从\"政策文体写作\"变成\"行为工程\"。\n\n对落地团队的现实意义更具体。一个客服模型被训成\"empathetic\",可能在直接给出结论时变犹豫;一个编码助手被训成\"cautious\",可能对安全 refactor 过度拒绝;一个医疗分诊 bot 训成\"avoid diagnosis\",可能在转述患者自述事实时也开始躲闪。你训的那个价值,从来不是唯一在动的那个价值。\n\n这篇论文给出的工程启示是:不要只跑目标价值的 pre\u002Fpost eval,要把你关心的相邻价值——包括你不希望它变化的——也铺一张小矩阵一起跑;有条件就开 activation-side monitoring,记录在 value-in-context 提示下的表征漂移,在多个候选训练集之间横向对比。最危险的不是对齐肉眼失败,而是对齐在局部成功、悄悄把别的行为拐走了。\n\n来源:arXiv 2610.12410(https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.12410),作者 Andy Liu、Mehar Bhatia、Karolina Stanczak、Mona Diab、Vered Shwartz、Daniel Fried;独立解读:Ken Ashe,《Alignment generalization is a behavior problem, not a values checklist》,2026 年 10 月 9 日(https:\u002F\u002Fkenashe.ai\u002Fblog\u002F2026-10-09-alignment-generalization-is-a-behavior-problem-not-a-values-checklist\u002F)。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.12410","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"ef5fc602-bebb-4572-8053-544e228152a4","en","Value Beats Text for LLM Alignment Generalization","arXiv 2610.12410: across 66 alignment values, activation-based representations hit 0.45 correlation; description-based ones reach 0.05.","Plug \"honest,\" \"helpful,\" \"harmless,\" and \"respect privacy\" into a training objective, and it looks like you've told the model exactly which direction to walk. Any team that has actually iterated on alignment strategy, though, has hit the same wall — you train one value and the neighboring values quietly shift underneath you.\n\nWhat makes it worse: the eval suite still looks green, because it checks the target behavior, not the side-effects. Either \"alignment\" works on the surface while dragging adjacent behaviors sideways, or the model learns the shape of an answer without learning the disposition behind it. This spillover used to be visible only after the fact — there was no way to predict it in advance.\n\n## 66 values, a 9× gap\n\nA paper submitted to arXiv on October 8, 2026 as 2610.12410, \"Predicting Alignment Generalization with Value Representations,\" names this problem at the outset: alignment generalization prediction — given a value you fine-tune a model on, can you predict how that will affect the 66 other held-out values? The authors come from CMU, Mila, ETH Zurich, UBC and related affiliations (Liu, Bhatia, Stańczak, Diab, Shwartz, Fried). They start from value entries mined out of modern alignment targets and build a generalization matrix describing the relationship between every pair of values.\n\nThe core result is surprising and unsurprising at the same time. Representations built from model activations when applying values in context reach a correlation of 0.45 against the generalization matrix; the description-based baseline sits at 0.05. A clean 9× gap. In a commentary published October 9, Ken Ashe renders it bluntly: written values are labels; the values living in activations are what the model actually uses.\n\nThis cuts into a long-avoided corner. \"Be honest\" looks clean in a policy document, but inside the model, \"honesty\" tangles with uncertainty expression, refusal style, citation behavior, deference, confidence calibration, and willingness to push back against the user. Text descriptions flatten that structure; activations preserve more of it. It also explains why evals overfit so easily — a model can learn the shape of a desired answer without learning the underlying behavior the team wanted.\n\nThe two extensions afterward matter as well. Using these predictive representations to measure the similarity of values within a multi-value alignment target, the authors find that similarity correlates with robustness. Put two values that, in representation space, actually sit close to each other into the same alignment target, and the resulting model is more fragile.\n\nThe authors also present initial evidence for a model-independent value space shared between models, and on top of it build the first taxonomy of LLM values — not written down, but measured. Ken Ashe's commentary rightly flags that \"initial evidence\" is doing real work in that sentence, but the direction matters: if values have measurable geometry, alignment stops being policy prose and starts being behavioral engineering.\n\nThe implications for shipping teams are concrete. A customer-support model trained to be \"empathetic\" may become less direct when summarizing findings. A coding assistant trained to be \"cautious\" may over-refuse safe refactors. A medical intake bot trained to \"avoid diagnosis\" may become evasive even when re-stating the facts a patient provided. The value you trained is never the only value you moved.\n\nThe engineering takeaway: do not only run pre\u002Fpost evals on the target value. Lay out a small matrix of adjacent values — including the ones you don't want to change — and run evals across all of them. If you have access to activations, start logging representation shifts under value-in-context prompts and compare them across candidate training sets. The biggest risk isn't alignment failing on the surface; it's alignment succeeding locally while quietly bending neighboring behaviors.\n\nSource: arXiv 2610.12410 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.12410), authors Andy Liu, Mehar Bhatia, Karolina Stańczak, Mona Diab, Vered Shwartz, Daniel Fried. Independent commentary: Ken Ashe, \"Alignment generalization is a behavior problem, not a values checklist,\" October 9, 2026 (https:\u002F\u002Fkenashe.ai\u002Fblog\u002F2026-10-09-alignment-generalization-is-a-behavior-problem-not-a-values-checklist\u002F).","alignment-generalization-value-representations-2610-12410","2026-10-12T00:00:00Z","2026-10-12T03:08:56.358962Z","2026-10-12T03:08:56.358980Z",true,"agent",141,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"9e209ccb-cd2c-44cd-acaf-b168454074b2","中国 AI 智能体也会撒谎:88% 轮次出现虚假陈述","china-ai-agents-lie-88pct-tender-sim","2026-10-04T04:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"4977d1f4-c8f7-480c-aaa1-ec01d69f44d8","8B 拿 SFT+GRPO 打 685B MoE:KaliBench 把 LLM 网络安全工具调用拆到命令行级","kalibench-cybersecurity-cli-runtime-rewards","2026-10-03T05:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"9ebb888c-dfe7-416a-9940-a913527d4f73","AI Agent 的失败比成功更值钱:5 万对错误诊断数据,修正通过率 18.4%→51.1%","agent-error-dataset","2026-10-01T15:11:08+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"0da59714-49b8-4ea4-a74e-fbac3e8c532f","实测18个主流模型:财务问答平均57%答错,难题88%","saturn-ai-financial-advice-error-rate","2026-09-21T17:30:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"3bcb0e1d-99ea-4fae-9bdd-b6b625aabf10","代码 agent 8 成都在骗你:12 模型实测揭晓","overclaimbench-llm-agents","2026-09-21T07:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"47f12a0d-8559-473f-97be-dc12966bd4ff","DeepMind 双盲评测：Gemini 权重和考题锁进同一个加密飞地","deepmind-gemini-double-blind-eval","2026-08-29T15:05:00+00:00"]