[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ultramarine-1-flash-non-verifiable-rl":3,"topics-all":38,"news-related-2c7a8ca0-ab46-454b-ae16-aa7e7b413753":57},{"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},"2c7a8ca0-ab46-454b-ae16-aa7e7b413753","Ultramarine 1 Flash:不可验证任务RL","Paper Instruments 用大规模强化学习把 GLM-5.3 Flash 后训练成 Ultramarine 1 Flash,专攻单元测试判不了分的知识工作。六项知识工作基准平均高基座 4.1 分且 token 少 6.5%,内部基准每任务 0.13 美元,但编码基准退步,成绩尚无第三方复现。","写并购备忘录、搭财务模型、做尽调报告，这些活儿没法用单元测试判对错——答案好坏取决于判断和品味，regex 解析器和 CI 都验证不了。Paper Instruments 在 10 月 14 日发布的 Ultramarine 1 Flash 冲着这块空白去：在 GLM-5.3 Flash 基础上做大规模强化学习后训练，把\"不可验证任务\"当成 RL 规模化攻坚的主战场。\n\n## 把不可验证任务按可验证任务的套路造\n\n方法论沿用近期可验证 RL 的框架定义每个 RL 任务：一个环境、一条模型指令、一套验证系统。区别只在于验证系统被换成了评分细则。\n\n交付物是 Excel 模型、PowerPoint 演示和 Word\u002FPDF 文档。评测流程把这些由智能体撰写或就地编辑的文件解析并渲染成文本，交给朴素 LLM 评分器按细则打分——用开源评分包的 PerCriterionOneShotGrader，配一套为贴合人类专家评判流程而调过的提示词。内部基准 PI Knowledge Work 上评分者是 GPT-5.6 Terra，跑在 262,144 token 上下文窗口的 Feather 框架里。\n\n课程切得很细：金融、法律、咨询、运营四个领域，19 个专业实践领域，88 种环境类型，任务指令和评分细则沿采样轴对抗式生成，报告称组合空间在 10⁷ 到 10¹⁰ 量级。\n\n## 训练配置与沙箱隔离\n\n公开的超参相当具体：基于 CISPO 改造成双策略目标，AdamW，批量 64 × 8，LoRA rank 64，学习率 5 × 10⁻⁵，temp = 1、top_p = 1、top_k = -1，损失按训练批次有效生成 token 总量归一化。上下文窗口 131,072 token，智能体最多 250 轮，压缩在上下文用到 85% 或剩余 30,000 token 时先到者触发。\n\n沙箱隔离那段值得单独看。智能体通过 MCP 访问应用服务，拿不到服务进程本身及其存储，评分产物和验证器都在执行边界之外。报告写得很直白：模型在沙箱里做的任何事，都改不了自己的 harness、凭证或评分器。可复用验证器依赖预先烘焙进环境快照，任务专属评分材料在 rollout 之后才下发。\n\n## 成绩单：知识工作赢，编码输了\n\n六项知识工作基准全部超过基座，平均高 4.1 分，输出 token 还少 6.5%。JobBench 上 62.9，基座 56.3，Gemini 3.8 Flash 60.6；FrontierFinance 上 53.3。\n\n编码基准却是退步的：Terminal-Bench 2.1 为 83.2，基座 84.3；DeepSWE 1.1 为 61.5，基座 63.4。APEX-Agents 1.1 上 57.0，落后 Claude Fable 5 的 63.6 和 Gemini 3.8 Flash 的 64.3。\n\n成本最抓眼。内部基准上 Ultramarine 1 Flash 拿到 59.3 分、每任务 0.13 美元；同口径 Claude Fable 5（max 推理）57.8 分、每任务 8.19 美元，GPT-5.6 Sol（max）55.1 分、每任务 2.09 美元，与 Claude Opus 5.5（6.96 美元）和 GPT-6 Sol（1.26 美元）的差距都在三分以内。据此 CEO Daanish Khazi 对外称\"性能接近 Fable 5，便宜 60 倍\"。\n\n## 该怎么看这组数字\n\n三点得说清楚。第一，所有基准分数都出自 Paper Instruments 自己，截至 10 月 16 日没有任何第三方机构复现过其中任何一项。第二，\"便宜 60 倍\"只在内部基准的任务级测算里成立，不是 token 单价上的普遍差异——它用的是各家 2026 年 10 月的官方 API 定价，任务还是自家构造的。第三，编码退步挺说明问题：后训练不是纯增益，拿知识工作评分细则去训，在验证信号明确的软件工程活儿上反而丢分。想吃下这条曲线，代价就写在表格里。\n\n真正值得关注的不是 59.3 这个数字，而是背后的机制——把评分细则当作 RL 的验证信号，再用沙箱隔离保证智能体改不了自己的判卷系统。当\"判断类工作\"逐渐变成智能体的默认工作面，能先把这套方法跑通的团队会先拿到迁移成本的红利。\n\n模型已上线 API，权重官方称将在约两周后放出。\n\n参考：Paper Instruments 技术报告 https:\u002F\u002Fwww.paperinstruments.com\u002Fblog\u002Fultramarine","https:\u002F\u002Fwww.paperinstruments.com\u002Fblog\u002Fultramarine","17684990-91d9-482c-ac74-52765af1a98b",[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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"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},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"4865d307-d82c-4817-97e3-a24fcc75021d","en","Ultramarine 1 Flash: Scaling RL for Non-Verifiable Work","Paper Instruments post-trained GLM-5.3 Flash into Ultramarine 1 Flash with large-scale RL: +4.1 points on six knowledge-work benchmarks with 6.5% fewer tokens.","Writing a merger memo, building a financial model, putting together a due diligence report — none of that can be graded by a unit test. Whether an answer is good turns on judgment and taste, and neither regex parsers nor CI can verify it. Ultramarine 1 Flash, released by Paper Instruments on October 14, targets exactly that blank space: it post-trains GLM-5.3 Flash with large-scale reinforcement learning, treating \"non-verifiable tasks\" as the main battlefield where RL actually has to scale.\n\n## Building RL tasks the way verifiable ones are built\n\nThe core methodology in the report mirrors how recent verifiable RL work formalizes each task: an environment, a model instruction, and a verification system. The only difference is that the verification system has been swapped for a rubric.\n\nFor knowledge work, the deliverables are Excel models, PowerPoint decks, and Word or PDF documents. The evaluation pipeline parses and renders those agent-authored or agent-edited files into text, then hands the result to a naive LLM rubric grader — using the PerCriterionOneShotGrader from their open-source rubric package, paired with a custom grading prompt tuned for agreement with human experts following the evaluation process. On the internal PI Knowledge Work benchmark, the grader is GPT-5.6 Terra, running in the Feather agent framework with a 262,144-token context window.\n\nThe curriculum is cut finely: four domains (finance, law, consulting, operations), 19 professional practice areas, and 88 distinct environment types, with task instructions and rubrics adversarially generated along sampling axes. The report puts the combination space at 10⁷ to 10¹⁰.\n\n## Training configuration and sandbox isolation\n\nThe official report publishes unusually specific hyperparameters: CISPO with a modified two-policy objective, AdamW optimizer, batch size 64 × 8, LoRA rank 64, learning rate 5 × 10⁻⁵, sampling at temp = 1, top_p = 1, top_k = -1, and loss normalized by the total count of valid generation tokens in the training batch. The context window is 131,072 tokens, the agent runs up to 250 turns, and compaction triggers at 85% of the context window or when 30,000 tokens remain, whichever comes first.\n\nThe sandbox isolation section deserves its own look. The agent reaches application services through MCP, without direct access to the service processes or their storage; grading artifacts and verifier state sit outside the agent's execution boundary. The report puts it bluntly: nothing the model does in the sandbox can alter its own harness, credentials, or grader. Reusable verifier dependencies are baked into environment snapshots in advance, and task-specific grading material is supplied only after the rollout.\n\n## Results: knowledge work up, coding down\n\nUltramarine beats its base model on all six knowledge-work benchmarks, by 4.1 points on average, while generating 6.5% fewer output tokens. On JobBench it scores 62.9 against the base model's 56.3, with Gemini 3.8 Flash at 60.6 and DeepSeek V4.1 Flash at 60.4; on FrontierFinance it reaches 53.3.\n\nBut the coding benchmarks regress. Terminal-Bench 2.1 comes in at 83.2 versus the base model's 84.3, and DeepSWE 1.1 at 61.5 versus 63.4. On APEX-Agents 1.1 it scores 57.0, behind Claude Fable 5 at 63.6 and Gemini 3.8 Flash at 64.3.\n\nCost is the part that catches the eye. On the internal PI Knowledge Work benchmark, Ultramarine 1 Flash scores 59.3 at $0.13 per task. On the same harness, Claude Fable 5 (max reasoning) scores 57.8 at $8.19 per task and GPT-5.6 Sol (max) scores 55.1 at $2.09, while it stays within three points of Claude Opus 5.5 ($6.96 per task) and GPT-6 Sol ($1.26 per task). On that basis, CEO Daanish Khazi says it delivers \"similar performance to Fable 5, while being 60x cheaper.\"\n\n## How to read these numbers\n\nThree caveats matter.\n\nFirst, every benchmark score originates from Paper Instruments. As of October 16, no third-party evaluation group has reproduced any of them, and the model is still tracked without an official rank on independent leaderboards.\n\nSecond, the 60x figure holds only for task-level runs on that internal suite — it is not a uniform token-rate difference. It uses each vendor's official API pricing as of October 2026, on tasks Paper Instruments constructed itself.\n\nThird, the coding regression is genuinely informative. Post-training is not a pure gain: trained on knowledge-work rubrics, the model improves there while losing ground on software engineering, where verification signals are clean. Anyone hoping to claim that curve has to accept the cost written in the report's tables.\n\nWhat is worth watching is not the 59.3 itself, but the mechanism behind it: treating rubrics as an RL verification signal, then using sandbox isolation to guarantee the agent cannot edit its own grading system. As \"judgment work\" becomes the default surface for agents, whoever operationalizes this approach first gets the migration-cost dividend.\n\nThe model is live on the API, and the company says weights will follow in about two weeks.\n\nReference: Paper Instruments technical report — https:\u002F\u002Fwww.paperinstruments.com\u002Fblog\u002Fultramarine","ultramarine-1-flash-non-verifiable-rl","2026-10-20T07:07:00Z","2026-10-20T07:11:34.352368Z","2026-10-20T07:11:34.352383Z",true,"agent",86,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,83],{"id":60,"title":61,"news_slug":62,"published_at":63},"27240196-3ac0-4af3-aec7-b0a4d3e7c990","智谱GLM-Legal法律Agent模型:Harvey榜前十最便宜","glm-legal-harvey-lab-cheapest","2026-10-13T15:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"c4ec4625-4a84-4c24-88f0-0ef1beb4f19e","Grok 4.6 发布:61 分追平 GPT-5.6 Sol,把长程 Agent 的 token 账单砍到四分之一","grok-4-6-agentic-cost-frontier","2026-08-14T19:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"3cc63477-1334-497d-80cb-90850c019101","DeepSeek-V4-Flash 转正:不靠换架构,只做后训练重新发力 Agent","deepseek-v4-flash-official-post-training-agent-0731","2026-07-31T08:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":78},"fbdcfd41-bb54-487a-8319-9f35dfc82be5","DeepSeek-V4-Flash转正：这次升级不靠换架构","deepseek-v4-flash-0731-agent-update",{"id":84,"title":85,"news_slug":86,"published_at":87},"aa718f3c-14cf-4984-8762-794eaaee8c16","MAI-Gym:9B 智能体在 2 万个并行环境里做在线 RL","mai-gym-gui-agent-online-rl-9b","2026-10-21T09:30:00+00:00"]