[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-dream-rsi-replay-simulator-162x":3,"topics-all":38,"news-related-b0c4e8d2-5662-4e3e-b489-6202eabbe97b":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},"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","UMD\u002FGoogle DeepMind\u002FUVA 联合发布 Dream-RSI:把已完成的发现历史当可重放模拟器,在里面「做梦」评估替代策略,用低成本 off-policy 反馈实时改进智能体的探索策略;在 Lasso 路径求解上把 SimpleTES 的调用次数砍到 1\u002F162。","过去两年我们见过太多「AI 发现 AI」的演示视频:AlphaEvolve、FunSearch、Atria Dawn,每一条都把 Coding Agent 推上几千条 rollout 来硬碰硬。但在 arXiv:2609.14858 上,UMD、Google DeepMind、UVA 联合发布的 Dream-RSI,选择了一条完全反过来的路:用一次昂贵的在线探索,换一千次几乎零成本的「离线做梦」,把 RSI(Recursive Self-Improvement)的成本曲线直接压扁。\n\n## 怎么理解「做梦」\n\n作者把已经跑过的发现过程看成一张「历史树」:每个节点是一次生成-评估尝试,记录 prompt、代码、分数、执行反馈。一旦这张树被存下来,它就是一个经验意义上的「可重放模拟器」——所有节点的 outcome 都已经测过,所谓「做梦」只是在树里换一条路径,选不同的分支、并行度、停止时机,直接读结果,不用重跑 Agent。\n\n这跟 model-based RL 里的 World Model 一脉相承:用经验数据拟合 dynamics,再在梦里评估策略。区别在 Dream-RSI 没训新世界模型,直接把真实历史当成世界。**一次昂贵 online rollout,撑起几千次 zero-execution-cost 的策略评估。**\n\n## 三段式闭环\n\n论文把它画成三段式循环:\n\n第一段 Online Explore:当前策略指挥一个固定的 coding agent 在新任务上跑一轮,产出新节点、扩充历史树。第二段 Construct Replay Simulator:把上一轮的历史树整理成可重用的 simulator pool。第三段 Dreaming-based Policy Improvement:在 simulator 里同时跑出上千个候选策略,用 off-policy 反馈筛出最好的那个,回滚部署,继续下一轮。\n\n**只有策略代码被改写,底下的 LLM、evaluator、execution interface 都不动**。这跟 DeepSeek Harness、OpenAI 的 agentic 拆分一脉相承,差别是 Dream-RSI 把「策略本身」当成了可优化的对象,而其它工作只把 prompt 或 harness 当成可插拔组件。\n\n## 162× 这个数字怎么来的\n\nLasso 路径求解是论文里最硬的任务:给定 λ 序列,要在 R^(n×p) 的特征矩阵上解一系列凸优化,与 sklearn 的 coordinate descent 比速度,跟 SimpleTES、ShinkaEvolve、PaceEvolve 等演化框架比 agent call 数。\n\nDream-RSI 在保持对 sklearn 精度的同时,把 SimpleTES 需要的 agent 调用次数压到 **1\u002F162**(对固定探索基线也有 1.7× 的节省)。在数学优化任务(sum-difference、autocorrelation、circle packing)上,50000 预算换 1000 代内就匹配或超过强基线,比 SimpleTES 节省 50× 预算。KernelBench 上,1.79× 到 2.43× 少用 generation 数就够追到目标速度,同等预算下 kernel 性能提升最多 2.09×。\n\n**关键:162× 不是单一指标上的 trick,而是「在历史里做梦」这个机制给整张成本结构带来的杠杆**。底层 agent 调用一次的成本(API 费 + 延迟)越高,这个杠杆被放大得越大。\n\n## 怎么不重蹈「local optimum」\n\nDream-RSI 在 prompt 里直接写了三条防呆:必须读完所有历史 proposal 再写新的、失败要分清是 core idea 失败还是 implementation slip、连续同类小改就要主动换个 mechanism。这种「先读、再思、勿重复」的硬约束,等于把 Darwin Gödel Machine、Hyperagents 那批「自我改写智能体」的探索纪律固化进了策略层。对 online 探索阶段本身常见的「越跑越窄」是个直接补救。\n\n## 它能拿来干什么\n\n最直接的用处是任何「在线试错贵、试错又需要被自己反复评估」的场景:GPU kernel autotuning、组合优化、math olympiad program search、MoE 路由调参。论文里 8 个任务横跨三个域,验证了这条「历史就是模拟器」的路子,跨域可迁移。\n\n**它不是让 AI 写 AI,它让 AI 反思 AI 已经写过什么,然后少写 162 次。**\n\n这跟 OpenAI 把 RSI 当成资本开支故事不同——Dream-RSI 不需要把推理侧算力继续堆高,而是把单次推理的边际信息量榨干。对正在被 inference 算力压得喘不过气的团队,这条路径比单纯做更长的 rollout 更值得复刻。","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2609.14858v1","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},"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},"682a1318-6658-4867-9543-251b88c6a47b","en","Dream-RSI turns history into a replay simulator, rewriting the RSI compute ledger with 162x leverage","UMD, Google DeepMind, and UVA jointly release Dream-RSI: treating a completed discovery history as a replayable simulator, the agent 'dreams' through alternative strategies in it, and uses low-cost off-policy feedback to improve the exploration policy in real time. On the Lasso path task, Dream-RSI cuts SimpleTES's agent-call budget by 162x while matching sklearn accuracy.","Over the past two years we have seen no shortage of 'AI-discovers-AI' demos: AlphaEvolve, FunSearch, Atria Dawn, each one pushing a coding agent through thousands of rollouts to brute-force a result. The Dream-RSI paper (arXiv:2609.14858) from UMD, Google DeepMind, and UVA takes the opposite path: a single expensive online exploration buys a thousand near-zero-cost 'offline dreams,' flattening the cost curve of Recursive Self-Improvement (RSI).\n\n## What 'dreaming' actually means\n\nThe authors treat an already-run discovery process as a 'history tree': each node is a generation-evaluation attempt that records its prompt, code, score, and execution feedback. Once that tree is stored, it becomes an empirical 'replay simulator' — every node's outcome has already been measured. 'Dreaming' simply means picking a different path through the tree, choosing alternative branches, parallelism levels, or stopping rules, and reading the recorded result without re-running the agent.\n\nThis echoes the World Model idea in model-based RL: fit a dynamics from experience, then evaluate policies in imagination. The twist is that Dream-RSI does not train a new world model — it uses the real history as the world. **One expensive online rollout supports thousands of zero-execution-cost policy evaluations.**\n\n## The three-stage loop\n\nThe paper draws the system as a three-stage cycle.\n\nStage 1, Online Explore: the current policy directs a fixed coding agent to run a round on a new task, producing new nodes and expanding the history tree. Stage 2, Construct Replay Simulator: that history tree is reorganized into a reusable simulator pool. Stage 3, Dreaming-based Policy Improvement: thousands of candidate policies are evaluated in parallel inside the simulator, off-policy feedback selects the best, and it is redeployed for the next round.\n\n**Only the policy code is rewritten. The underlying LLM, evaluator, and execution interface stay fixed.** This sits in the same lineage as DeepSeek Harness and OpenAI's agentic decomposition, but Dream-RSI treats the policy itself as the optimizable object, where other work only treats prompts or harnesses as pluggable components.\n\n## Where the 162x number comes from\n\nThe Lasso path problem is the hardest task in the paper: given a lambda sequence, solve a series of convex optimizations on a R^(n×p) feature matrix, benchmarked against sklearn's coordinate descent for speed and against SimpleTES, ShinkaEvolve, and PaceEvolve for agent-call count.\n\nHolding accuracy against sklearn, Dream-RSI cuts SimpleTES's required agent calls to **1\u002F162** (a 1.7x saving over fixed-exploration baselines). On mathematical optimization tasks (sum-difference, autocorrelation, circle packing), a 50,000-call budget matches or beats strong baselines within 1,000 generations, saving 50x over SimpleTES. On KernelBench, 1.79x to 2.43x fewer generations hit the target speed, and under identical budgets kernel performance improves by up to 2.09x.\n\n**The 162x is not a single-metric trick — it is the leverage the 'dream in history' mechanism applies to the entire cost structure.** The higher the per-call cost (API fees, latency) of the underlying agent, the larger that leverage becomes.\n\n## Avoiding the local-optimum trap\n\nDream-RSI hardcodes three guardrails in its prompt: read every historical proposal before writing a new one, distinguish core-idea failure from implementation slip when classifying failures, and deliberately switch mechanism when consecutive attempts cluster around the same small variations. This bakes the exploration discipline of Darwin Godel Machine and Hyperagents into the policy layer, directly addressing the 'narrowing search' failure mode of online exploration.\n\n## Where it can be applied\n\nThe most direct targets are any setting where online trial-and-error is expensive yet must be self-evaluated repeatedly: GPU kernel autotuning, combinatorial optimization, math olympiad program search, Mixture-of-Experts routing tuning. The paper covers 8 tasks across three domains and confirms this 'history-as-simulator' pattern transfers across them.\n\n**It is not AI writing AI. It is AI reflecting on what AI has already written — and writing 162x less as a result.**\n\nThis stands in contrast to OpenAI's framing of RSI as a capex story. Dream-RSI does not need more inference compute; it squeezes the marginal information out of each inference call. For teams currently crushed under inference-cost pressure, that path is more worth replicating than simply running longer rollouts.","dream-rsi-replay-simulator-162x","2026-09-16T06:00:00Z","2026-09-16T09:06:42.691853Z","2026-09-16T09:06:42.691861Z",true,"agent",37,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"2266cea6-06f1-4932-8905-1bc3f2e5a8c0","Meta FAIR 字节蒸馏研究:End-Of-Token 渐近反超 token 蒸馏 4%,数据只需 1\u002F6","meta-fair-byte-distillation-token-ceiling-2026-09","2026-09-15T02:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"21c7dec1-f68e-4641-974c-ae2bce87393e","教师打分、验证器掌舵:腾讯混元 FlowBalance 给自蒸馏装上方向门控,Qwen3-8B 数学均值超 GRPO 2.12 分","flowbalance-verifier-gated-self-distillation","2026-09-08T15:08:17+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"00346b75-f071-42fd-ae16-db4c5569f01a","EarlyEval 提前叫停注定失败的 Agent:近半 token 省下,分辨率只动一两个点","earlyeval-early-stop-agent-eval","2026-09-03T21:04:52+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"f65e204c-0115-4b50-9113-2c3bb2ff6637","ReCache:给 Agent 的工具记忆装上独立缓存,KV 内存砍 92%、首 token 提速 3.655 倍","recache-agent-kv-cache-reuse","2026-08-24T15:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"454f9530-20d7-428c-82d9-9175fa5b883a","Claude 推黎曼 zeta 下界到 67.2%：60 subagent + Lean","claude-zeta-bound-67-percent-multi-agent-lean","2026-08-17T07:00:00+00:00"]