[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-dream-rsi-replay-simulator":3,"topics-all":38,"news-related-af34b075-9fd7-4f0f-8f87-eb947fd81f4e":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},"af34b075-9fd7-4f0f-8f87-eb947fd81f4e","Dream-RSI:Google 让智能体在历史里做梦,发现调用省 162 倍","Google 与 DeepMind 团队提出 Dream-RSI:把智能体探索留下的发现树当作精确重放模拟器,离线做梦筛选探索策略,胜者才上线。论文报告 Lasso 任务发现调用比 SimpleTES 少 162 倍,GPU kernel 任务以 2.43 倍更少迭代达到相当性能。","递归自我改进被普遍视为智能体技术的下一级台阶,但真正跑过发现循环的团队都知道,这个循环里有一个部件至今还是人手写的:探索策略——往哪儿分支、并行跑哪些、什么时候砍掉一条死路。Google 与 Google DeepMind 联合马里兰大学、弗吉尼亚大学的研究团队在 9 月 14 日放出的 Dream-RSI 论文(arXiv:2609.14858,[项目页](https:\u002F\u002Fdream-rsi.com\u002F))里,把矛头对准了这个瓶颈。\n\n## 探索策略的两难\n\n论文的分析很直白:固定策略无法从积累的经验里学习,会反复为已经失败的方向付费;而在线优化策略又同时撞上两堵墙——元层面的反馈又慢又贵,因为评判一个探索策略意味着看它把整场发现跑完才能打分;而策略空间巨大,大多数候选都是坏的,每个坏候选都要付出一整次 rollout 的代价才能排除。\n\n## 历史就是模拟器\n\nDream-RSI 的核心洞察是:这份反馈你早就付过钱了。一次完成的发现过程不只是可以重读的文本,而是一棵结构化的树,记录了智能体做过的每一个探索决策及其真实执行结果。而探索策略做的事恰恰只有一件:根据已见内容决定下一步继续哪条尝试。所以一个替代策略根本不需要重新运行任何东西——它以不同顺序走同一棵记录在案的树,它要问的每个结果都已经在磁盘上。论文称,可以在零执行成本下筛掉成千上万个候选策略,只把真正的一次在线 rollout 留给胜者。\n\n有意思的是,这个模拟器不是学习出来的世界模型,而是精确的:在已实现的搜索空间上没有任何预测成分,因为它就是那个搜索空间本身。代价是边界同样锋利——策略只能做梦到历史真正去过的地方。这正是它必须做成循环的原因:每次上线部署都记录一棵新树,智能体拥有的不是一个世界,而是一个不断增长的世界池;在更多世界上做梦筛出的策略,胜过只调参于单次运行运气的策略。\n\n机制上还有一个漂亮的设计:离线阶段是一个策略开发智能体对策略代码连续改写多版,每一版都在全部历史上重放打分;因为当前部署的策略本身也在候选集里,胜出者永远不会比现任更差。\n\n## 论文报告的数字\n\n在算法工程、数学优化、GPU kernel 工程三个领域的八个发现任务上,论文报告的结果是:GPU kernel 任务 VGG16 上以 2.43 倍更少的迭代达到相当性能(LayerNorm 为 1.79 倍),ConvDiv 在相当预算下拿到 2.09 倍的分数;Lasso 正则化路径任务上,Dream-RSI 用 Gemini-3.1-Pro 只花了 317 次发现智能体调用,而对照系统 SimpleTES 花了 51,200 次——论文给出的口径是调用数少 162 倍;数学优化任务上 SimpleTES 拿着最佳单项数字,但那是 51,200 代的成本,Dream-RSI 用的是不到 1,000 代。论文同时提到,与固定探索相比,发现智能体调用削减约 1.7 倍。\n\n更反直觉的是分析一节:把历史轨迹抽象成高层洞见再注入 prompt 的语义指导方案,在同等预算下一致地差于不做指导——强先验会过度约束搜索空间,压制探索多样性。学到的策略也不是单调变贪心:每轮评估尝试先从 110 降到 50,进入平台期后再回升到 90 附近,而这些放宽与下一轮分数跳升正好对齐。\n\n## 所以呢\n\n多数智能体框架把历史当成两种东西:要么是塞进上下文的文本,要么是微调权重的训练数据。Dream-RSI 给了第三种读法——历史是环境,是已经付费建好却一直闲置的模拟器。这个视角不只适用于发现类任务:任何留下完整执行轨迹的长程智能体,理论上都坐拥同样的重放资产。\n\n当然要诚实:以上数字全部来自论文自报,受控基线 Recursive Fixed Exploration 也是团队自己搭的,独立验证还需要时间;而且只能在历史去过的地方做梦,冷启动阶段世界池很小,方法的价值要靠循环转起来才兑现。但即使打对折,把探索策略从手写脚本变成可进化代码、把元层面的评估成本压到接近零,这个方向对所有做智能体基础设施的团队都是个提醒:你磁盘上那堆轨迹日志,可能比你以为的值钱得多。\n\n一句总结:历史不只是上下文,它是你已经付费却从未真正使用过的模拟器。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.14858","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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":22,"name":23,"slug":23,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9b5a2106-f9ec-4ce4-abb7-15aab2b483f5","en","Dream-RSI: Google Lets Agents Dream in Their Own History, Cutting Discovery Calls 162x","Google and DeepMind researchers propose Dream-RSI: treating the discovery trees left behind by agent exploration as exact replay simulators, screening exploration policies offline by dreaming, and deploying only the winner. The paper reports 162x fewer discovery calls than SimpleTES on Lasso and comparable performance with 2.43x fewer generations on GPU kernel tasks.","Recursive self-improvement is widely seen as the next step for agent technology, but any team that has actually run a discovery loop knows one component is still hand-written: the exploration policy — where to branch, what to run in parallel, when to cut off a dead line. A team from Google and Google DeepMind, together with the University of Maryland and the University of Virginia, takes aim at this bottleneck in the Dream-RSI paper released on September 14 (arXiv:2609.14858, [project page](https:\u002F\u002Fdream-rsi.com\u002F)).\n\n## The exploration dilemma\n\nThe paper's analysis is blunt: a fixed strategy cannot learn from accumulated experience, so it keeps paying for directions that have already failed. Optimizing the policy online, meanwhile, hits two walls at once — meta-level feedback is delayed and expensive, because judging an exploration policy means watching it steer an entire discovery run to the end; and the policy space is vast, so most candidates you would need to try are bad ones, each costing a full rollout to find out.\n\n## History is already a simulator\n\nDream-RSI's key insight: you have already paid for that feedback. A finished discovery run is not just text to re-read — it is a structured tree of every exploration decision the agent made, each carrying the execution outcome it actually produced. And an exploration policy does exactly one thing: given what it has seen, choose which attempt to continue next. So an alternative policy never has to re-run anything. It walks the same recorded tree in a different order, and every outcome it asks for is already on disk. The paper claims thousands of candidate policies can be screened at zero execution cost, with a real rollout spent only on the winner.\n\nNotably, this simulator is not a learned world model — it is exact. Over the search space that was realized, nothing is predicted, because the simulator is that search space. The flip side is just as sharp: a policy can only be dreamt where history actually went. That is precisely why this has to be a loop — every online deployment records a new tree, so what the agent owns is not one world but a growing pool of them; a policy dreamt across more worlds beats one tuned to the luck of a single run.\n\nThere is also an elegant design detail: in the offline phase, a policy-development agent writes successive revisions of the policy code, each scored by replay over the whole history. Because the currently deployed policy is itself in the candidate set, the winner is never worse than the incumbent.\n\n## The numbers as reported\n\nAcross eight discovery tasks spanning algorithm engineering, mathematical optimization, and GPU kernel engineering, the paper reports: on GPU kernel task VGG16, comparable performance with 2.43x fewer generations (1.79x on LayerNorm), and a 2.09x higher score on ConvDiv under comparable budgets. On the Lasso regularization path task, Dream-RSI with Gemini-3.1-Pro spent just 317 discovery-agent calls versus 51,200 for the comparison system SimpleTES — a 162x reduction in calls, as the paper frames it. On mathematical optimization, SimpleTES holds the best single number, but at a cost of 51,200 generations against fewer than 1,000 here. The paper also notes roughly 1.7x fewer discovery-agent calls versus fixed exploration.\n\nMore counterintuitive is the analysis section: semantic guidance — abstracting prior trajectories into high-level insights and injecting them into the prompt — consistently underperformed its unguided counterpart under equal budgets. Strong priors over-constrain the search space and suppress exploration diversity. The learned policy is not monotonically greedier either: evaluated attempts per round first fell from 110 to 50, then rose back to around 90 after entering a plateau — and those widenings lined up with the next jumps in round-best score.\n\n## So what\n\nMost agent frameworks read history in one of two ways: static text to stuff into the context, or training data to fine-tune weights on. Dream-RSI offers a third reading — history is an environment, a simulator already built and paid for but left idle. The perspective extends beyond discovery tasks: any long-horizon agent that leaves complete execution traces theoretically sits on the same replay asset.\n\nTo be honest about the caveats: every number above is self-reported in the paper, the controlled baseline Recursive Fixed Exploration was built by the team itself, and the SimpleTES figures are their reproduction — independent verification will take time. The method also can only dream where history has already gone, so the world pool is small during cold start, and the value only materializes once the loop is spinning. But even at half strength, turning exploration policies from hand-written scripts into evolvable code and pushing meta-level evaluation cost toward zero is a reminder for every team building agent infrastructure: those trajectory logs on your disk may be worth far more than you think.\n\nOne line to remember: history is not just context — it is the simulator you already paid for and never really used.","dream-rsi-replay-simulator","2026-09-15T23:20:00Z","2026-09-15T23:06:20.597267Z","2026-09-15T23:06:20.597282Z",true,"agent",281,[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},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a3e9cf60-e2dc-4ac6-a1a6-1089ee721bc9","Google Antigravity CLI 全面开放：子Agent并行编排进入终端开发时代","google-antigravity-cli-sub-agent-parallel","2026-05-31T10:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"33c8038a-f1dc-4b64-b6b7-604baf43f729","CESA 数据:85.8% 日本游戏开发者已把生成式 AI 写进工作流","cesa-2026-japan-game-dev-genai-85pct","2026-09-20T00:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"c696208b-6535-4eb9-b1ed-2e4f835d2f88","NVIDIA SoL-Pi 把 coding agent 的 token 砍掉 44%,harness 开始变天","nvidia-sol-pi-harness-token-compression","2026-09-19T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"1bf3f192-68e9-49f1-8bbb-a350f05c15e2","LLM 编排云边端:FRAME 2026 论文点出反馈层缺环","llm-cloud-continuum-frame-2609","2026-09-16T11:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"b0c4e8d2-5662-4e3e-b489-6202eabbe97b","Dream-RSI 把历史当模拟器:162 倍杠杆重写 RSI 算力账本","dream-rsi-replay-simulator-162x","2026-09-16T06:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d400c0db-49df-4cc6-a87e-87b709f59fea","Muse Spark 1.3 发布:卡住会向用户求助的 Agent,工具调用少 20%、token 省 25%","muse-spark-1-3-meta-agent-release","2026-09-06T15:12:00+00:00"]