[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lora-skill-document-recovery":3,"topics-all":38,"news-related-d30f7434-f5ad-487d-94b2-269c0dcdb382":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},"d30f7434-f5ad-487d-94b2-269c0dcdb382","LoRA反编译来了:读权重不行,得让模型跑起来","技能文档被编译成LoRA省上下文,代价是一块读不了、难审计的数字。arXiv新工作研究反编译:只读权重仅传约1比特;让适配模型自己写的Executing Writer在rank 64下恢复90%-93%文本、还原85%-91%任务增益;行为型适配器靠执行代理读差异,技能把Llama-3.1-8B从5%拉到55%。","一份几百到几千词的技能文档(skill),教模型怎么审代码、怎么分析抓包,现在越来越多被直接编译成 LoRA 适配器——省上下文、调用便宜,但从此你手里握着的是一块数字:读不了、改不了、也没法装到另一个 tokenizer 不同的模型上。文本技能\"可以在信任之前被读一遍\",权重技能没有这个待遇。arXiv 新贴出的论文 From LoRA to Language 问了一个反向问题:给定适配器和冻结的基座模型,能不能把技能文档写回来?\n\n## 读权重:认得出是哪个技能,读不出它说了什么\n\n论文先泼了一盆冷水。不同技能适配器的权重更新之间,余弦相似度高达 0.88;把共享成分扣掉,只剩 11.6% 的能量是这个技能独有的。权重并非没有痕迹:一个适配器 top-40 关键词里 73% 出现在它自己的文档里(其他文档只有 11%),检索编码器能把 37 个留出测试技能全部认对。但\"知道是哪个技能\"离\"知道它说了什么\"很远——学习出来的权重到文本通道,每千 token 文档只传得出约 1 比特信息,而光是从 11000 个候选里点名这个技能就要 13.4 比特。直接读数字这条路是死的。\n\n## 跑起来,让它自己写:Executing Writer\n\n有效的做法是让网络自己读。Executing Writer 把适配器挂上冻结基座,再在顶上训一个 43.6M 参数的小 writer 适配器,让适配后的模型把自己的说明书默写出来。在 37 个留出技能上,它恢复出 38.7%-43.7% 的关键指令(只读权重的最好方法是 22.3%);换成另一个技能的适配器,数字直接塌到 1.3%-1.5%,证明文本确实来自适配器;胡编内容占比也从其他方法的 54%-66% 降到 39%。在 41 个 SkillsBench 外部技能上,拟合到\"会背诵文档\"的适配器让它恢复出 rank 64 下 90%-93% 的正文,把恢复文档放回提示词,能还原原始技能 85%-91% 的任务增益——文档顶上,适配器的大部分效果就回来了。\n\n## 只学行为的适配器才是常态,得换执行代理\n\n论文划了一条关键界线:背诵型适配器训练时见过文档原文,恢复它是提取;行为型适配器只见过被文档塑造过的答案,恢复它是推断——后者才是实践中的常态。对行为型适配器,训练出的 writer 恢复率只有 1%-2%,甚至低于错误适配器对照组。作者于是换了武器:让一个执行代理在真实任务上分别跑基座模型和适配后模型,从两者行为不一致的地方把技能写出来。在 ALFWorld 家务游戏里,基座 Qwen2.5-7B-Instruct 胜率 25.4%,LoRA 达到 81.3%,代理读适配器后单次会话写出的技能恢复 58% 的 LoRA 增益,加上奖励循环能到 71%-75%。更扎心的是对照组:当 LoRA 的行为是\"猜也猜得到\"的常识,奖励循环不带适配器写得一样好——适配器里的信息,只在行为猜不到时才值钱。docker-compose 任务上的惯例技能,真实适配器写出 0.212 的精确匹配,错误适配器的所有会话全是 0.000。隐藏规则实验里,代理 12 次会话有 10 次把正确规则写进了技能文本,可 7B 小模型照着执行的比率仍然很低。\n\n## 文档最大的优势:它哪里都能去\n\nLoRA 离不开自己的基座模型,文档可以。从 Qwen2.5 的 LoRA 抄出来的技能,放进 Llama-3.1-8B 的提示词,把它在 ALFWorld 的胜率从 5% 拉到 55%;Qwen3.5-4B 从 2.2% 拉到 66.4%;在费用计算任务上,给 4B 模型写的技能把 Claude Haiku 从 0% 拉到 96%——后者根本装不上这个 LoRA。\n\n## 所以呢\n\n对技能市场和企业的适配器资产来说,这篇论文把\"反编译适配器\"从没人想过的问题,变成了可测量、可迭代的工程问题:接入一个来路不明的适配器前,可以先让它默写成说明书给人审一遍;技能跨模型流通,也不再必须重训。限制也写得明白:打分靠 LLM 裁判,485 个判定里两个裁判只对得上 272 个;主实验只有 37 个技能;会话间方差常常大于组间差距;小模型读得懂规则、执行不执行得了,还是另一回事。反编译的门被推开了一条缝,后面是数据量和耐心的问题。(论文: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09454)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09454","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"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},"0f74335b-0336-45ce-b1ea-8941aa2fefd6","en","LoRA decompilation: run the adapter, don't read the weights","Skills compiled into LoRA become numbers. New arXiv work writes the text back by running the adapter, not reading weights: 90-93% of prose recovered.","A skill document - a few hundred to a few thousand words teaching a model how to review code or analyse a packet capture - can be compiled into a LoRA adapter. It saves context on every call and serves cheaply, but what you hold afterwards is a block of numbers: hard to read, hard to edit, impossible to load onto a model with a different tokenizer. A text skill \"can be read before it is trusted\"; a weight skill has no such privilege. A new arXiv paper, From LoRA to Language, asks the reverse question: given the adapter and its frozen base model, can the skill document be written back?\n\n## Reading the weights: it knows which skill, not what it says\n\nThe paper starts with cold water. The cosine similarity between the weight updates of different skills' adapters is 0.88; after removing what adapters share, only 11.6% of the energy is specific to the skill. The weights do carry traces: 73% of an adapter's top-40 keywords appear in its own document (against 11% for other documents), and a retrieval encoder names the correct document for all 37 held-out test skills. But knowing which skill an adapter holds is far from knowing what it says - a learned channel from weights to text passes on about one bit per thousand-token document, while merely naming the skill among 11,000 candidates costs 13.4 bits. Reading the numbers directly is a dead end.\n\n## Run it and let it write: the Executing Writer\n\nWhat works is letting the network do the reading. The Executing Writer attaches the adapter to its frozen base model and trains one small writer adapter - 43.6M parameters - on top, so that the adapted model writes out its own manual. On 37 held-out skills it recovers 38.7%-43.7% of the key instructions (the best weights-only method reaches 22.3%); swap in another skill's adapter and the number collapses to 1.3%-1.5%, evidence that the text really comes from the adapter. The unsupported share of what it writes falls to 39%, against 54%-66% for the other methods. On 41 external SkillsBench skills, adapters fitted to recite their document let it recover 90%-93% of the prose at rank 64, and putting the recovered documents back into the prompt restores 85%-91% of the original skill's task gain - with the document in place, most of the adapter's effect comes back.\n\n## Behaviour-only adapters are the norm - and need an executing agent\n\nThe paper draws a line that matters: a recitation adapter saw its document text in training, so recovering it is extraction; a behaviour-only adapter saw only answers the document shaped, so recovering it is inference - and the second kind is the common case in practice. For behaviour-only adapters the trained writer recovers just 1%-2%, below the wrong-adapter control. So the authors switch tools: an executing agent runs the base and adapted models on real tasks and writes a skill from where they differ. On ALFWorld, the base Qwen2.5-7B-Instruct wins 25.4% of games and the LoRA 81.3%; a skill written in one session by the agent reading the adapter recovers 58% of the LoRA's gain, rising to 71%-75% with a reward loop. The controls bite: where the LoRA's behaviour is guessable common sense, the reward loop without any adapter writes just as well - the adapter matters only when its behaviour cannot be guessed. On a docker-compose task the skill from the real adapter reaches 0.212 exact match against 0.000 in every wrong-adapter session. In hidden-rule experiments the agent stated the correct rule in 10 of 12 sessions, yet the 7B actor still rarely carried it out.\n\n## The document's real advantage: it travels\n\nThe LoRA cannot leave its base model; the document can. A skill written from a Qwen2.5 LoRA, dropped into the prompt of Llama-3.1-8B, lifts its ALFWorld win rate from 5% to 55%; it lifts Qwen3.5-4B from 2.2% to 66.4%; on a fee-calculation task, a skill written for a 4B model raises Claude Haiku from 0% to 96% - a model that cannot load the LoRA at all.\n\n## So what\n\nFor skill marketplaces and enterprise adapter inventories, this paper turns \"decompile the adapter\" from a question nobody asked into a measurable, iterable engineering problem: before adopting an unknown adapter you can transcribe it into a manual for human review, and moving a skill across models no longer demands retraining. The limits are stated plainly: scoring relies on LLM judges, and the two judges agree on only 272 of 485 verdicts; the main experiments cover 37 skills; session spread often exceeds the gap between arms; and whether a small actor executes a written rule is another matter. The door to decompilation has been pushed open a crack; what follows is a matter of data volume and patience. (Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09454)","lora-skill-document-recovery","2026-11-09T19:11:39Z","2026-11-09T19:11:39.510795Z","2026-11-09T19:11:39.510803Z",true,"agent",262,[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},"c4623b85-37f5-435e-ae98-e67eb72eb1dc","别再逐篇问大模型:AI写程序当评委,每主题0.17美元起","execjudges-llm-written-program-judges-retrieval-eval","2026-11-20T01:40:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"4544ed60-b5fb-41b6-9fee-dd2f59012547","隐状态转个方向:直答46到85,超CoT省九成token","cot-hidden-state-directions-direct-answers","2026-11-19T13:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"648ee35f-321e-4cc7-8e88-9a9d3ebae033","小模型每晚自学:一年后从 52 分掉到 5 分","on-device-llm-self-training-judge-failure","2026-11-19T03:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"43903d4a-7891-411b-9ea6-80708f7f8309","让大模型改权重:42道题测出能力边界","weightbench-llm-weight-space-programming","2026-11-16T07:20:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"0a4281e9-69b1-43f0-9e7b-78f0cd7ea4c2","医疗小模型回答砍72.8%,准确率反升3.2点","peca-conciseness-qlora-medical","2026-11-15T15:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"b602ac1b-9734-4b80-bb3c-16f19ae493a0","LoRA 的秩在虚标:配了 896,只用到 361","lora-effective-rank-capacity-utilization","2026-11-14T23:30:00+00:00"]