[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-allspark-iris-search-agent-sft-rl-climbing":3,"topics-all":38,"news-related-9822a1a7-0014-4bd5-bbe0-492401fe6b96":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},"9822a1a7-0014-4bd5-bbe0-492401fe6b96","AllSpark 把搜索 Agent 推到 BrowseComp 88.6:SFT-RL Climbing 与推理时上下文管理","AllSpark Research 开源 Iris-mini(35B-A3B)与 Iris-pro(397B-A17B)两个搜索 Agent,基于 SFT-RL Climbing 与推理时上下文管理,在 BrowseComp\u002FBrowseComp-ZH\u002FDeepSearchQA\u002FHLE 拿到开源段位最强成绩。","AllSpark Research 9 月 3 日公开的 Iris,从训练流程与推理时上下文管理两侧撕开\"开源搜索 Agent 追不上闭源\"这道口子。\n\n## 模型与训练\n\nIris 一并放出两个 MoE:Iris-mini 是 35B 总参 \u002F 3B 激活,Iris-pro 是 397B 总参 \u002F 17B 激活,共用同一管线。**SFT-RL climbing**:监督微调冷启动 + 强化学习在真实搜索接口上做 on-policy 优化,反复交替;RL 阶段奖励裁判与 observation summarizer 都跑在训练集群内部,对超长 rollout 在请求级打断、从已提交的 prefix 处续跑。\n\n## 数据工程\n\n训练任务不是从问答语料挖,而是从网页超链结构反向构造:取种子页面、沿外链展开成实体图、写多跳问答链、**把所有非答案实体改写成描述性引用**让任何线索不能字符串匹配命中,最后做闭卷验证——这道题必须让 reference 模型闭卷答不出,但把支撑证据喂给它就能解。\n\n这一步是质量门,把\"答案能从问题里抄出来\"的低质样本挡在外面。幸存问题变成轨迹,过滤后才进入 SFT。这种\"图结构反向挖题 + reference rewrite + 闭卷失败验证\"三段式,从正面堵掉\"题目泄露答案\"——搜索 Agent RL 的最大隐性成本是 reward hacking,Iris 把\"题面不能再含答案\"做成硬约束,把这条 hacking 路径堵掉一半。\n\n## 基准数字\n\n四基准统一用单 ReAct Agent、无子 Agent、无 test-time verification,固定工具集、上下文窗口与裁判模型。**启用推理时 context management** 后:Iris-mini 在 BrowseComp、BrowseComp-ZH、DeepSearchQA、HLE 上分别拿到 82.2、84.8、86.9、52.3;Iris-pro 拿到 88.6、85.1、92.9、56.4。\n\n论文明确:在各自参数段位,两个模型是开源搜索 Agent 中总体最强的。BrowseComp-ZH 上 pro(85.1)只比 mini(84.8)高 0.3 分,其余三榜拉开 4 个点以上——**大参数对深度证据检索的边际收益,比对中文检索更显著**。\n\n论文同时对比\"开启 \u002F 关闭 context management\"两组结果,同基准、同工具集、同裁判——把\"context 管理\"与\"模型本身\"两条曲线拆开,坐实\"上下文管理在复杂 web 基准上比多数系统级差异更值钱\"。\n\n## 开放与启示\n\nAllSpark 计划放出模型权重、数据构造与训练评测的完整 recipe、配套 GitHub 仓库 `github.com\u002FAllSpark-Research\u002FIris`。有 35B \u002F 400B 段算力的团队可照 SFT-RL climbing 步骤把自有基座复刻成同段位搜索 Agent——**训练数据是怎么造的,比训练脚本怎么写更值得公开**。\n\ncontext management 不是新词,但把它当成\"和基座同台 PK\"的变量,是论文另一个值得抄的方法——HLE 56.4 这种高难基准里,上下文管理带来的相对涨幅往往盖过换基座的涨幅。做 deep research 类产品的团队,启示很直接:**别只盯着换基座,先把上下文管理做成独立模块、把工具调用做成稳定 API、把奖励做成低时延裁判**。\n\n来源:AllSpark Research,arXiv 2609.04304(2026-09-03 提交),HF papers 当日榜首(43 upvotes)。https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04304","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04304","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},"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},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"1d6da6fe-08e0-4eda-b466-6a8462904de4","en","AllSpark Iris pushes open-source search agents to BrowseComp 88.6","AllSpark ships Iris-mini (35B-A3B) and Iris-pro (397B-A17B) search agents via SFT-RL climbing. Top open results on BrowseComp, DeepSearchQA, HLE.","For the past six months, the search-agent race has centered on one question: how do you keep an agent from falling apart over many turns of live web search? Closed systems from OpenAI and Anthropic have pushed BrowseComp and HLE high, but the open-source side has never closed the gap — the base models are there, but no reproducible recipe has turned them into competent searchers. AllSpark Research's Iris, submitted to arXiv on 3 September 2026, attacks this from two sides at once: the training pipeline and inference-time context management.\n\n## Two sizes, one training recipe\n\nIris ships a pair of MoE models — Iris-mini at 35B total \u002F 3B active parameters, Iris-pro at 397B total \u002F 17B active — both numbers straight from the abstract. They share one training pipeline; only the base-model scale and post-training hyperparameters differ. A single data recipe scales cleanly across two deployment shapes.\n\nThe training procedure is what the authors call **SFT-RL climbing**: supervised fine-tuning to cold-start the policy, then reinforcement learning against live search APIs, alternating. During RL, the reward judge and the observation summarizer both run inside the training cluster to minimize reward-signal latency. For rollouts that exceed the length budget, the system interrupts at the request level and resumes from the committed prefix on the next step — a vent for the \"latency trap\" in long-context RL training.\n\n## The data-side key moves: reverse construction and reference rewriting\n\nIris's training tasks are not mined from existing QA corpora. They are reverse-engineered from the hyperlink structure of the web. Concretely:\n\n1. Take a seed page and fan out its outgoing links into an entity graph.\n2. Author multi-hop question chains over that graph.\n3. Rewrite every non-answer entity into a descriptive reference so no clue can be resolved by string match.\n4. Validate: the question must fail closed-book against a reference model, yet succeed once the supporting evidence is provided.\n\nStep four is the quality gate — it filters out almost every \"the answer is already in the question\" sample. Surviving questions become trajectories, filtered at both trajectory and turn level before SFT. The three-stage pipeline — graph-based reverse mining, reference rewriting, closed-book failure validation — is an unusually engineering-grade answer to \"the training question itself leaks the answer\", a failure mode search agents are uniquely vulnerable to.\n\n## Benchmark numbers: four boards, two sizes, top of each band\n\nThe paper runs four benchmarks under one fixed setup — single ReAct agent, no sub-agents, no test-time verification, fixed tool set, fixed context window, fixed judge. **With inference-time context management enabled**:\n\n- Iris-mini: BrowseComp **82.2** \u002F BrowseComp-ZH **84.8** \u002F DeepSearchQA **86.9** \u002F HLE **52.3**\n- Iris-pro: BrowseComp **88.6** \u002F BrowseComp-ZH **85.1** \u002F DeepSearchQA **92.9** \u002F HLE **56.4**\n\nThe paper states explicitly that in their respective parameter bands, these are the strongest overall results among open-source search agents. On BrowseComp-ZH, pro (85.1) leads mini (84.8) by just 0.3 points, while the gap widens to 4+ points on the other three boards — a useful read on **how scale interacts with depth of evidence retrieval vs. Chinese-language coverage**.\n\nThe most consequential design choice in the benchmark section is the side-by-side reporting of \"with \u002F without context management\" results, holding tool set, context window and judge constant. This separates \"context management\" from \"the model itself\" as independent variables — quantifying, in benchmark form, the claim that context management is worth more on these boards than most reported system-level differences. Engineering effort outside the training pipeline is here for the first time held up to the same scrutiny as the model itself.\n\n## The open release plan\n\nPer the paper and the project page, AllSpark plans to ship together:\n\n- Iris-mini and Iris-pro model weights (HF collection page: huggingface.co\u002Fcollections\u002FAllSpark-Research\u002Firis);\n- The full recipe for data construction, training, and evaluation;\n- A companion GitHub repo at github.com\u002FAllSpark-Research\u002FIris.\n\nThat means any team with 35B\u002F400B-class compute can replicate the SFT-RL climbing pipeline on its own base model. This breaks from the past year's standard pattern of \"open weights + open paper + locked training details\" — **how the training data was constructed is more worth publishing than how the training script was written**.\n\n## Why this matters for search-agent training\n\nTraining-side: SFT-RL climbing is not new (the self-play lineage after RLHF is mature), but pairing it with reverse-engineered questions, reference rewriting and closed-book validation is what makes Iris's recipe defensible. The hidden cost of search-agent RL is reward hacking — any intermediate step can game \"did the agent finally answer correctly\" by leaking the answer. Iris turns \"the question surface cannot contain the answer\" into a hard constraint, halving the surface area available for hacking.\n\nInference-side: context management is not new vocabulary, but treating it as a peer-level experimental variable — rather than \"something the base model already does\" — is the methodological move worth copying. On high-difficulty boards like HLE, the relative lift from context management frequently exceeds the lift from swapping the base model — yet this used to be silently absorbed into the headline score.\n\nProduct-side: for any team shipping deep-research products, Iris's practical lesson is direct — **stop chasing base-model upgrades, and instead build context management as a standalone module, harden tool invocation into a stable API, and tighten the reward judge into a low-latency service**. These three efforts matter more to shipped quality than a single base-model bump.\n\nThe \"so what\": the open-source search-agent line has evolved from LangChain-era ReAct engineering, through WebGPT\u002FWebShop-style \"search + behavior\" training signals, to Iris's \"reverse-engineered questions + context management as a reproducible recipe\" stage. **The hard part of search-agent work is moving from \"who can top the leaderboard\" to \"who can replicate the recipe\"**. Iris's answer is the latter.\n\nSources: AllSpark Research; arXiv paper 2609.04304 (submitted 2026-09-03); #1 paper on Hugging Face Daily Papers (43 upvotes, topped 2026-09-07). Paper page: https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04304 ; HF papers page: https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2609.04304 .","allspark-iris-search-agent-sft-rl-climbing","2026-09-07T07:11:17Z","2026-09-07T07:11:28.694693Z","2026-09-07T07:11:28.694702Z",true,"agent",286,[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},"44aa8908-1cef-48d6-b224-7de12a8d4afd","NeoHorse-1：让 Agent 执行轨迹进入自我改进回路","neohorse-1-agentic-post-training-rsi","2026-09-09T07:19:25+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"14a7f5ab-e270-461c-b862-4bde139e463f","HarnessDev 基准:让 LLM 自建 Agent Harness,代码领域仍输人类工程师","harnessdev-llm-selfbuilt-agent-harness","2026-09-03T19:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"089195fb-7fe5-4ba9-a4bc-8e356fe5e923","BAAI把1000个GitHub仓库蒸馏成5000个技能,科研agent奖牌率31%冲到73%","baai-disco-repo-to-skill-library","2026-09-03T17:07:35+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"3c6fcf46-f5bb-4136-931c-69cd64216e12","Skill-Use 基准揭示 Agent 短板：会做任务，不等于会用 Skill","skill-use-agent-harness-benchmark","2026-08-06T08:00:00+00:00"]