[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-astabrief-8b-scientific-report-model":3,"topics-all":39,"news-related-52f192a9-5630-435d-a76a-237e2c94eda5":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":25,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":15,"view_count":38},"52f192a9-5630-435d-a76a-237e2c94eda5","AstaBrief 开源:8B 写科研报告快 3.5 倍","Ai2 开源 Asta 的报告引擎 AstaBrief-8B:Qwen3-8B 基座,仅用 SFT+DPO,研究问题加检索片段一次写成带引用报告。平均 51.1 秒一份,比 Claude 版 Thinking 模式快约 3.5 倍;训练用 4.7 万条真实查询报告对,Apache 2.0 权重数据全放出。","写文献综述最磨人的不是读论文,而是读完之后把几十篇的结论缝成一份有引用、有结构的报告。这个活儿现在被 Ai2(艾伦人工智能研究所)拆出来单干了:他们把自家科研助手 Asta 背后的报告生成模型 AstaBrief-8B 连同训练数据一起开源,Apache 2.0 协议,权重放在 Hugging Face 上随便下。\n\n## 51 秒出一份带引用的报告\n\nAstaBrief 的定位很窄:输入一个研究问题和一批检索到的文献片段,一次直出带引用的完整报告。官方博客给出的数字:接入 Asta 全管线后,Fast 模式平均 51.1 秒一份报告,Claude 驱动的 Thinking 模式是 178.5 秒,约 3.5 倍差距。提速的关键不是模型更大,而是流程更短——它跳过了 Thinking 模式里昂贵的片段摘要和聚类步骤,也不逐节写作,直接一遍生成终稿。\n\n## 训练配方:真实查询,双裁判把关\n\n模型基于 Qwen3-8B,没上 RL,走的是更朴素的 SFT+DPO 路线。SFT 数据从 Asta 系统的真实用户查询里来:过滤掉测试和机器人流量后留下 9 万条科研查询,再用 ScholarQA 管线生成报告作训练目标,报告生成一侧混用了 Claude 3.5 Sonnet、Claude 3.7 Sonnet、o3、o4-mini 和 GPT-4.1,筛完剩 4.7 万条。DPO 数据约 6 千对,胜负由 GPT-4.1 和 DeepSeek-R1 双裁判判定,只保留两位裁判一致的对,裁判与人类偏好的一致率标称 95%。AI2 还给出一个反直觉发现:测试过的四种统计过滤里,最有效的只是\"引用密度\"这一个简单信号——合成报告是否持续给论断挂引用,比更复杂的过滤组合都好用。\n\n## 官方自报的分数,得辩证看\n\n模型卡显示,在 ScholarQA-CS2 测试集上 Qwen3-8B 基线平均 77.3 分,SFT 后 83.7,最终 DPO 版 87;对 Asta 原有管线的胜率在测试集上到 72%。但同一张表也摆着不那么好看的数:DeepScholarBench 上 AstaBrief 是 53.50,落后 DR-Tulu 的 56.26;14 题人工小样本评测里,总体偏好 DR-Tulu 赢,不过三位研究员中有两位在引用准确性上更喜欢 AstaBrief。这些数字全部来自团队自报,评测完成于 2025 年,AI2 自己也注明没有和今天的最新模型重跑。\n\n## 所以呢\n\n对科研机构,这件事的实际意义是自托管:报告生成这一步可以挪到自己防火墙后面,研究问题里那些没发表的敏感内容不必再发给闭源 API。对行业,它是一份便宜的配方样本——不堆 RL,把功夫花在真实查询、引用密度过滤和双裁判偏好对上,8B 小模型就能在特定任务上顶住闭源管线的大头。问题只剩一个:当所有人都能下载这个模型,报告生成这门生意的护城河还剩什么?\n\n参考:https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fallenai\u002Fastabrief","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fallenai\u002Fastabrief","446fa1f9-89de-415a-a793-275018ca9899",[11,16,19,22],{"id":12,"name":13,"slug":13,"description":14,"color":15},"9112951a-2abb-4214-b63a-385ec7afb2ba","ai-for-science","AI for Science 专题：追踪 AI 在生命科学、化学材料、物理世界模型等科学方向的关键突破",null,{"id":17,"name":18,"slug":18,"description":15,"color":15},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":20,"name":21,"slug":21,"description":15,"color":15},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":23,"name":24,"slug":24,"description":15,"color":15},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[26],{"id":27,"lang":28,"title":29,"summary":30,"content":31},"35c7ed5d-ff73-405f-b98c-9b1add7e2243","en","AI2 Open-Sources AstaBrief 8B: Cited Reports 3.5x Faster","Ai2 open-sources AstaBrief-8B, Asta's report engine: Qwen3-8B base, SFT+DPO, one-pass cited reports at 51.1s vs 178.5s for Claude-powered mode.","Scientific report writing has a tedious core: after reading dozens of papers, stitching their findings into one structured, citation-backed report. Ai2 (the Allen Institute for AI) has now split that job out and open-sourced it: AstaBrief-8B, the model behind the report-generation feature of its Asta research assistant, ships under Apache 2.0 with weights and training data on Hugging Face.\n\n## A cited report in 51 seconds\n\nAstaBrief has a narrow job: take a research question plus retrieved literature excerpts and emit a complete, cited report in one pass. Per the official blog, across the full Asta pipeline the Fast mode averages 51.1 seconds per report, versus 178.5 seconds for the Claude-powered Thinking mode — roughly 3.5x. The speedup comes from a shorter pipeline, not a bigger model: it skips the expensive snippet-summarization and clustering stages that Thinking mode uses, and it refuses to write section by section, generating the final draft directly.\n\n## The recipe: real queries, two judges\n\nThe model starts from Qwen3-8B and skips RL in favor of a plain SFT+DPO recipe. SFT data comes from real user queries in Asta: after filtering out beta-tester and bot traffic, 90K research queries remained; the ScholarQA pipeline then generated target reports using a mix of Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini and GPT-4.1, leaving 47K usable examples after quality filtering. DPO added about 6K pairs, with winners picked by two judges, GPT-4.1 and DeepSeek-R1; only pairs where both judges agreed were kept, and the judges' agreement with human preferences is reported at 95%. AI2's counter-intuitive finding: of the four statistical filters tested, the single most useful signal was simply whether synthetic reports consistently cite their claims — more aggressive combinations added little.\n\n## Read the self-reported numbers with care\n\nOn the ScholarQA-CS2 test set, the model card shows Qwen3-8B averaging 77.3, the SFT checkpoint 83.7, and the final DPO model 87; the win rate against Asta's original pipeline reaches 72% on test. The same table carries less flattering numbers too: on DeepScholarBench, AstaBrief scores 53.50, behind DR-Tulu's 56.26; and in a 14-question human study DR-Tulu won overall preference, though two of the three researchers preferred AstaBrief on citation accuracy. All of these are team-reported numbers, the evaluation was completed in 2025, and AI2 itself notes it has not rerun the comparison against today's frontier models.\n\n## So what\n\nFor research institutions, the practical point is self-hosting: the report-generation step can move behind their own firewall, so sensitive questions about unpublished work no longer need to reach a closed API. For the industry, this is a cheap-recipe sample — no heavy RL stack, just real queries, citation-density filtering and dual-judge preference pairs, and an 8B model holds most of the line against a proprietary pipeline on one specific task. The remaining question: once anyone can download the model, what moat is left in report generation?\n\nReference: https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fallenai\u002Fastabrief","astabrief-8b-scientific-report-model","2026-10-04T15:06:06Z","2026-10-04T15:08:59.884099Z","2026-10-04T15:08:59.884113Z",true,"agent",509,[40,48],{"slug":13,"tag_slug":13,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":36,"created_at":46,"modified_at":47},"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":36,"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},"5736961c-9502-40d0-a39b-27f8ceb4380a","AI造AI:Recursive开源笔记本编程模型Flint","recursive-flint-laptop-coding-model","2026-10-22T19:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"2ddb354a-06f7-4781-b2d5-bcfd1faf38a5","Dory 1.0:英伟达4B循环模型,指令遵循反超9B","nvidia-dory-1-0-4b-looped-llm","2026-10-20T21:09:38+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"faf5c85e-ed83-432a-8b55-80c9356862c5","腾讯优图开源5B全模态解析:文档榜96.96超GPT-5.2","youtu-parsing-omni-5b","2026-10-19T19:09:27+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00"]