[ { "title": "Autonomous Quantum Transport Measurements of 2D Semiconductors by an AI Agent", "authors": "Brandon Bauer et al.", "arxiv": "2609.26661", "url": "https://arxiv.org/abs/2609.26661", "verified_version": "v1", "type": "arXiv预印本", "verified_at": "2026-09-28T11:28:47.803630+08:00", "limitations": "专家先用已有数据dry run迭代Skills；属于专家工作流内适应，非自主创立实验策略；未证通用GUI迁移、公开远程接入或低接入成本。", "read_first": "Fig.1；Monolayer MoS2 Device Campaign；Discussion", "id": "arxiv:2609.26661", "appearances": [ "digests/scientific-computer-use/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "OSWorld-Pro: Process-based Evaluation for Computer Use Agents", "authors": "Zhilin Wang et al.（NVIDIA）", "arxiv": "2609.24890", "url": "https://arxiv.org/abs/2609.24890", "verified_version": "v1", "type": "arXiv预印本", "verified_at": "2026-09-28T11:28:47.803630+08:00", "limitations": "桌面成功不等于科学正确；跨基准分数非受控因果比较；LLM判断不能替代测量与不确定度核验。", "read_first": "§2–4、§6、附录D依赖规则", "id": "arxiv:2609.24890", "appearances": [ "digests/scientific-computer-use/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design", "authors": "Zihan Dong et al.（Georgia Tech等）", "arxiv": "2609.16251", "url": "https://arxiv.org/abs/2609.16251", "verified_version": "v2，2026-09-21", "type": "arXiv预印本", "verified_at": "2026-09-28T11:28:47.803630+08:00", "limitations": "单应用多工作台不等于跨应用；包含计算/仿真，不是连接真实机床；没有GUI/API公平成本对照。", "read_first": "§3产物验证、§4结果、§6局限", "id": "arxiv:2609.16251", "appearances": [ "digests/scientific-computer-use/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "A Modular IoT-Enabled Remote Laboratory Platform for Hybrid Energy System Research and Engineering Education", "authors": "Lamine Chalal et al.", "arxiv": "2609.00815", "url": "https://arxiv.org/abs/2609.00815", "verified_version": "v1", "type": "arXiv预印本", "verified_at": "2026-09-28T11:28:47.803630+08:00", "limitations": "论文有会议排版，未另核正式出版日期；这里用arXiv上线日。非LLM agent；三个入口不等于一次跨应用科研轨迹；通信延迟不是完整实验周期。", "read_first": "Fig.2、§3.1、§4.2及Table3", "id": "arxiv:2609.00815", "appearances": [ "digests/scientific-computer-use/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "New Evidence, Same Choice: Testing Physical Experiment Selection in Vision Language Models", "authors": "Sourajit Saha et al.", "arxiv": "2609.11022", "url": "https://arxiv.org/abs/2609.11022", "fulltext_url": "https://arxiv.org/html/2609.11022v1", "type": "arXiv预印本", "verification": "原始摘要与正文§3已核", "verified_at": "2026-09-29T07:05:13.820543+08:00", "evidence": "六个开放模型、144组参数；直接回答在95.1%–100%的图像对上重复同一动作；简短推理最佳成对正确率5.9%", "read_first": "§3.2–3.3，再读§4", "id": "arxiv:2609.11022", "appearances": [ "digests/scientific-computer-use/2026/2026-09-29.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "A computable representation of the physical laboratory enables verifiable workflows", "authors": "Xiaobo Li et al.", "arxiv": "2609.03621", "url": "https://arxiv.org/abs/2609.03621", "type": "arXiv预印本", "verification": "仅原始摘要；HTML/PDF网页读取失败", "verified_at": "2026-09-29T07:05:13.820543+08:00", "evidence": "摘要报告已在模块化agentic机器人实验室实现；无本轮已核定量成绩", "read_first": "摘要；下一次补正文的真实执行与评价章节", "id": "arxiv:2609.03621", "appearances": [ "digests/scientific-computer-use/2026/2026-09-29.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Reimagining research papers as interactive and reliable AI agents", "authors": "Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, James Zou", "doi": "10.1038/s41586-026-11044-y", "url": "https://www.nature.com/articles/s41586-026-11044-y", "type": "Nature研究论文／Paper2Agent", "verification": "期刊原文Main、Large-scale evaluation和Methods已核；补充材料未核", "verified_at": "2026-09-29T07:05:13.820543+08:00", "evidence": "10篇非生物计算论文、42个执行任务，五次独立运行准确率98.1±0.8%；包含emcee等计算工具", "read_first": "Large-scale evaluation of Paper2Agent；Methods同名节", "id": "doi:10.1038/s41586-026-11044-y", "appearances": [ "digests/scientific-computer-use/2026/2026-09-29.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Can Autonomous LLM Agents Execute Multireference Quantum Chemistry Calculations?", "authors": "Victor Chang Lee; James M. Rondinelli", "team": "Northwestern University", "doi": "10.48550/arXiv.2609.13357", "arxiv": "2609.13357", "url": "https://arxiv.org/abs/2609.13357", "fulltext_url": "https://arxiv.org/html/2609.13357v1", "type": "arXiv预印本", "verification": "原始摘要与正文§II、§IV及Tables 1–2已核", "verified_at": "2026-09-30T07:07:06.049153+08:00", "evidence": "Table 1：558个参考VTE，覆盖率24.9%→44.1%，已识别子集MAE 0.373→0.339 eV；§IV.5计算数131→576", "read_first": "§II.3–II.6；Table 1；§IV.5–IV.7", "id": "arxiv:2609.13357", "appearances": [ "digests/scientific-computer-use/2026/2026-09-30.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments", "authors": "Hejia Geng et al.", "doi": "10.48550/arXiv.2609.19134", "arxiv": "2609.19134", "url": "https://arxiv.org/abs/2609.19134", "fulltext_url": "https://arxiv.org/html/2609.19134v1", "type": "arXiv预印本", "verification": "原始摘要与正文§2、§3.1–3.2、§5及附录8–9已核", "verified_at": "2026-09-30T07:07:06.049153+08:00", "evidence": "Hard集合85任务、18环境、每任务一小时；最高观察成功率67.1%，非统计确立的模型排序", "read_first": "§2.2；§3.1；附录9.1–9.3；§5", "id": "arxiv:2609.19134", "appearances": [ "digests/scientific-computer-use/2026/2026-09-30.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution", "authors": "Subba Reddy Oota et al.", "team": "ADIA Lab等", "doi": "10.48550/arXiv.2609.38108", "arxiv": "2609.38108", "url": "https://arxiv.org/abs/2609.38108", "type": "arXiv预印本", "verification": "摘要、正文§3–5、Table 5、附录E.3/I/L定向核查；PDF第10页视觉复核", "verified_at": "2026-10-01T07:14:02.937370+08:00", "evidence": "Table 5 DeepSeek-V4 ALFWorld：Plan+ReAct 0.480、最佳固定Search 0.918、自主Routing@1 0.721；不能将0.918写成自主路由成绩", "read_first": "Table 5；§3.3；附录E.3、I.1–I.2、L", "id": "arxiv:2609.38108", "appearances": [ "digests/scientific-computer-use/2026/2026-10-01.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Comparing Active Anomaly Detection Frameworks: From Light Curves to Images", "authors": "Sreevarsha Sreejith; Robert C. Nichol", "doi": "10.48550/arXiv.2609.36096", "arxiv": "2609.36096", "url": "https://arxiv.org/abs/2609.36096", "type": "arXiv预印本", "verification": "原始摘要、PDF§3.5–3.6、§4–6定向正文已核", "verified_at": "2026-10-01T07:14:02.937370+08:00", "evidence": "前100对象：光变变星83/78/38（Protege/PineForest/Astronomaly）；图像目标70/29/19（PineForest/Astronomaly/Protege）", "read_first": "§3.5–3.6；§4.3及结果；§6", "id": "arxiv:2609.36096", "appearances": [ "digests/scientific-computer-use/2026/2026-10-01.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "OSWorld-Science: A Benchmark of Computer Use Agents for Learning and Using Scientific Software", "authors": "Dingyuan Dai et al.；清华、UCLA、RIKEN等", "doi": "10.48550/arXiv.2609.39903", "arxiv": "2609.39903", "url": "https://arxiv.org/abs/2609.39903", "type": "arXiv预印本", "verification": "原始摘要、正文评估部分及附录F.1天文案例已核", "verified_at": "2026-10-04T14:23:11.073458+08:00", "evidence": "146任务；最佳73.7%是含VOID的平均部分得分。天文3任务×12模型共36次运行，14次满分，平均部分得分45.1%。净计数参考1424.92 counts，满分容差0.5 counts", "read_first": "§3.2；附录F.1；§4轨迹分析", "id": "arxiv:2609.39903", "appearances": [ "digests/scientific-computer-use/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Strategies for Deploying AI Agents in Production at Scientific User Facilities", "authors": "Ming Du et al.；Argonne/APS团队", "doi": "10.48550/arXiv.2609.36362", "arxiv": "2609.36362", "url": "https://arxiv.org/abs/2609.36362", "type": "arXiv预印本", "verification": "原始摘要与正文§3.1.7、§3.2、§4.1–4.2已核", "verified_at": "2026-10-04T14:23:11.073458+08:00", "evidence": "既有工具层可复用到已配置束线；26-ID从图像定位到acquire_fine_scan。未给同预算科学成功率或新设施适配人时", "read_first": "§3.2.3；§4.1–4.2；§3.1.7与§3.5", "id": "arxiv:2609.36362", "appearances": [ "digests/scientific-computer-use/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "A Programmable Optics Cloud Laboratory", "authors": "Sachin Vaidya et al.；MIT Research Laboratory of Electronics等", "doi": "10.48550/arXiv.2609.16413", "arxiv": "2609.16413", "url": "https://arxiv.org/abs/2609.16413", "type": "arXiv预印本", "verification": "原始摘要及正文§IV–VI、Table I已核；直接HTML下载失败但网页原文可读", "verified_at": "2026-10-04T14:23:11.073458+08:00", "evidence": "重复测量RMSD 3.87 μm（1.61相机像素）；5次端到端运行、15/15配置恢复成功。恢复设定不恢复可测量；优化内部中间状态不在抽象trace中", "read_first": "§IV-C–E；§V-A/V-C；§VI Fig.4与Table I", "id": "arxiv:2609.16413", "appearances": [ "digests/scientific-computer-use/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Agents Are Systems, Not Models: Rethinking Agentic Evaluation", "authors": "Luis Wiedmann、Leander Girrbach、Cordelia Schmid、Zeynep Akata", "doi": "10.48550/arXiv.2610.01618", "arxiv": "2610.01618", "url": "https://arxiv.org/abs/2610.01618", "type": "arXiv预印本", "verification": "原始摘要与正文§3.1、Table 2、§4.3已核；仅取天文部分", "verified_at": "2026-10-04T14:23:11.073458+08:00", "evidence": "Table 2红移任务完成76.3±1.8%；完成运行的平均R² -0.109，对照专用模型0.753。完成率与科学质量分母不同；信息影响不在每个底座上都最大", "read_first": "§3.1；Table 2；§4.3", "id": "arxiv:2610.01618", "appearances": [ "digests/scientific-computer-use/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "EAA: Automating materials characterization with vision language model agents", "team": "Ming Du、Mathew J. Cherukara等；APS/Argonne", "doi": "10.48550/arXiv.2602.15294", "arxiv": "2602.15294", "url": "https://arxiv.org/abs/2602.15294", "type": "arXiv预印本", "verification": "原始摘要和正文所列章节已核；作者报告，未独立复现", "verified_at": "2026-10-05T07:10:56.914134+08:00", "evidence": "聚焦例扫描8个位置；Siemens star搜索8次采集，末次仍仅近似居中", "read_first": "§3.1 Fig.4 → §3.2 Fig.5 → §2.4", "id": "arxiv:2602.15294", "appearances": [ "digests/scientific-computer-use/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Artifact-centered Claim-aware Observability for Autonomous Scientific Agents", "team": "Xiangyu Yin、Ming Du、Michael H. Prince、Mathew J. Cherukara", "doi": "10.48550/arXiv.2608.18312", "arxiv": "2608.18312", "url": "https://arxiv.org/abs/2608.18312", "type": "arXiv预印本", "verification": "原始摘要和正文所列章节已核；作者报告，未独立复现", "verified_at": "2026-10-05T07:10:56.914134+08:00", "evidence": "给出八类核心对象与字段绑定例；§6为toy示例，4.2pp不是实验成绩", "read_first": "§3 Table 1 → §4.1 → §6", "id": "arxiv:2608.18312", "appearances": [ "digests/scientific-computer-use/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "AIM: Agentic Idea Management for Automated Research", "team": "Google Cloud AI Research、University of Wisconsin–Madison；Hyeong Kyu Choi等", "doi": "10.48550/arXiv.2609.38445", "arxiv": "2609.38445", "url": "https://arxiv.org/abs/2609.38445", "type": "arXiv预印本", "verification": "原始摘要和正文所列章节已核；作者报告，未独立复现", "verified_at": "2026-10-05T07:10:56.914134+08:00", "evidence": "作者报告相对ScientistOne系统优化+1.6pp、长程模型/CUDA+4.9pp，最高3.1倍wall-clock加速", "read_first": "§4.3 → §5 → Appendix C.1", "id": "arxiv:2609.38445", "appearances": [ "digests/scientific-computer-use/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "A hardware-safety-gated system for LLM-written native ARTIQ control code on a trapped-ion platform", "team": "Duanyang Wang、Lu Qi、Yuanheng Xie、Norbert M. Linke、Kenneth R. Brown；UMD QLab等", "doi": "10.48550/arXiv.2606.27231", "arxiv": "2606.27231", "url": "https://arxiv.org/abs/2606.27231", "type": "arXiv预印本", "verification": "原始摘要和正文所列章节已核；作者报告，未独立复现", "verified_at": "2026-10-05T07:10:56.914134+08:00", "evidence": "自由运行补偿后T2*=1483±73 μs；人工多次指导目标、拟合与参数化", "read_first": "III.2 Fig.4/Table 1 → II.3 → III.3", "id": "arxiv:2606.27231", "appearances": [ "digests/scientific-computer-use/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ASCEND: Personal AI Agents for Autonomous Scientific Computing Across HPC Clusters and GPU Workstations", "team": "J. Paul Liu、Uthpala Herath、Andrew Petersen；NC State、Duke", "doi": "10.48550/arXiv.2609.32868", "arxiv": "2609.32868", "url": "https://arxiv.org/abs/2609.32868", "version_date": "2026-09-30", "type": "arXiv预印本", "verification": "原始摘要和正文所列章节已核；作者报告，未独立复现", "verified_at": "2026-10-05T07:10:56.914134+08:00", "evidence": "天气曲线偏差z500 2.1%、t850 2.4%；同编译选项修复后串行与16线程输出一致，-O2并行6.8倍", "read_first": "Table 1 → §4.2 → Table 3 → §6", "id": "arxiv:2609.32868", "appearances": [ "digests/scientific-computer-use/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "RF Instrument Agent (RFIA): Empowering RF Instruments with Natural Language Understanding, Scheduling and Execution of Complex Tasks", "team": "Chunhui Li、Wei Fan", "doi": "10.48550/arXiv.2605.23636", "arxiv": "2605.23636", "url": "https://arxiv.org/abs/2605.23636", "version_date": "2026-05-22", "type": "arXiv预印本", "verification": "原始摘要与所列正文段落已核；作者报告，未独立复现", "verified_at": "2026-10-06T17:21:57.123699+08:00", "evidence": "16项预设任务：15项执行完成，1项危险删除校准请求正确阻断；信道延迟估计200.100/500.350ns对应设定200/500ns", "read_first": "§V-D/Table VI → §IV-C → Tables IV/VIII", "id": "arxiv:2605.23636", "appearances": [ "digests/scientific-computer-use/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments", "team": "Takuya Isogawa等；MIT，Mingda Li、Paola Cappellaro", "doi": "10.48550/arXiv.2607.25145", "arxiv": "2607.25145", "url": "https://arxiv.org/abs/2607.25145", "version_date": "2026-07-27", "type": "arXiv预印本", "verification": "原始摘要与所列正文段落已核；作者报告，未独立复现", "verified_at": "2026-10-06T17:21:57.123699+08:00", "evidence": "该案例18.9小时；T2*约2–3微秒且依赖估计方法；13C附近特征只支持可能耦合，非确证", "read_first": "Results工作流/Table 1/Fig.2 → Fig.4", "id": "arxiv:2607.25145", "appearances": [ "digests/scientific-computer-use/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Autonomous Algorithm Discovery for Ptychography via Evolutionary LLM Reasoning", "team": "Xiangyu Yin、Ming Du等；APS关注作者后续补读", "doi": "10.48550/arXiv.2603.05696", "arxiv": "2603.05696", "url": "https://arxiv.org/abs/2603.05696", "version_date": "2026-03-05", "type": "arXiv预印本", "verification": "原始摘要与所列正文段落已核；作者报告，未独立复现", "verified_at": "2026-10-06T17:21:57.123699+08:00", "evidence": "IC：100个算法、83%评估成功；最佳SSIM .785，固定3000迭代较基线+0.26 SSIM/+8.3dB；搜索16.5小时", "read_first": "Table 1/Fig.4 → §4.5", "id": "arxiv:2603.05696", "appearances": [ "digests/scientific-computer-use/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "YouRA: A Persistent-State Architecture for Evidence-Traceable Autonomous Research Agents", "team": "Yoonkyu Woo、Woojin Lee、Jin-Xia Huang", "doi": "10.48550/arXiv.2610.01097", "arxiv": "2610.01097", "url": "https://arxiv.org/abs/2610.01097", "version_date": "2026-10-01", "type": "arXiv预印本", "verification": "原始摘要与所列正文段落已核；作者报告，未独立复现", "verified_at": "2026-10-06T17:21:57.123699+08:00", "evidence": "10个端到端任务、3个匹配模型骨干、4个LLM评审；Sonnet4.5全系统4.20，无VSA 2.88（1–10分）", "read_first": "§3状态表示 → Table 3 → §4.1日志评审", "id": "arxiv:2610.01097", "appearances": [ "digests/scientific-computer-use/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts", "team": "Oliver Jaffe、Dane Sherburn；P-Zero Research", "doi": "10.48550/arXiv.2610.06824", "arxiv": "2610.06824", "url": "https://arxiv.org/abs/2610.06824", "version_date": "2026-10-05", "type": "arXiv预印本", "verification": "原始摘要与所列正文段落已核；作者报告，未独立复现", "verified_at": "2026-10-06T17:21:57.123699+08:00", "evidence": "24专家，取每题最佳专家尝试为基线；20模型；40 H100小时或120墙钟小时上限；Opus5.5计算倍率2.3×（95%CI 1.15–4.37）", "read_first": "§2.1/§2.5 → §4", "id": "arxiv:2610.06824", "appearances": [ "digests/scientific-computer-use/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "WhatWorkedBench: Benchmarking Experimental Understanding in AI Agents", "team": "Jingjie Ning等；CMU／清华", "doi": "10.48550/arXiv.2609.27490", "arxiv": "2609.27490", "url": "https://arxiv.org/abs/2609.27490", "version_date": "2026-09-28", "type": "arXiv预印本", "verification": "原始摘要及上述关键正文已核；作者报告，未独立复现", "verified_at": "2026-10-07T07:06:00.789510+08:00", "evidence": "8次购买测量＋2个免费锚点下，pair-ridge在22源中选中15个最优，其中13个仍有条件效应误差超过效用范围10%。v2的12源前瞻组Flash全部交付，相同观测下比GP的恢复分数高0.147。", "read_first": "§3、§5 → §6.1与Table 4", "id": "arxiv:2609.27490", "appearances": [ "digests/scientific-computer-use/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Agents for self-driving laboratories applied to quantum computing", "team": "Shuxiang Cao、Zijian Zhang等；Oxford／Toronto／Aspuru-Guzik组", "doi": "10.1016/j.patter.2025.101372", "arxiv": "2412.07978", "url": "https://arxiv.org/abs/2412.07978", "version_date": "2025-06-05", "type": "论文关联Patterns正式版；本轮核arXiv修订原文", "verification": "原始摘要及上述关键正文已核；作者报告，未独立复现", "verified_at": "2026-10-07T07:06:00.789510+08:00", "evidence": "原文展示单比特校准、两比特siZZle搜索与GHZ制备／层析；搜索示例约3小时，设100实验和20频率预算。", "read_first": "Fig.4 → II Results → IV.2／IV.3", "id": "arxiv:2412.07978", "appearances": [ "digests/scientific-computer-use/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "QCalEval: Benchmarking Vision-Language Models for Quantum Calibration Plot Understanding", "team": "Shuxiang Cao等；NVIDIA／Toronto／LBNL／Fermilab等", "doi": "10.48550/arXiv.2604.25884", "arxiv": "2604.25884", "url": "https://arxiv.org/abs/2604.25884", "version_date": "2026-04-28", "type": "arXiv预印本", "verification": "原始摘要及上述关键正文已核；作者报告，未独立复现", "verified_at": "2026-10-07T07:06:00.789510+08:00", "evidence": "18个VLM评测中，通用模型最佳零样本平均72.3，Ising Calibration 1为74.7；是百分制问答分数，不是硬件成功率。", "read_first": "§3／Table 2 → §6 → App.A.3", "id": "arxiv:2604.25884", "appearances": [ "digests/scientific-computer-use/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Demonstration of a Structured Agentic Workflow for Applied Quantum Computing Research", "team": "Dikshant Dulal、Maxence Grandadam等；Haiqu", "doi": "10.48550/arXiv.2610.05304", "arxiv": "2610.05304", "url": "https://arxiv.org/abs/2610.05304", "version_date": "2026-10-04", "type": "arXiv预印本", "verification": "原始摘要及上述关键正文已核；作者报告，未独立复现", "verified_at": "2026-10-07T07:06:00.789510+08:00", "evidence": "STO-3G有限基组下SQD势垒555 meV，与FCI 574 meV相差19 meV；5次普通agent基线中3次违背全电子空间要求。", "read_first": "§4 → LLM use disclosure → App.A.1", "id": "arxiv:2610.05304", "appearances": [ "digests/scientific-computer-use/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "aRieL - Reinforcement Learning for the Ariel Space Telescope Scheduling", "team": "James Kostas Ray、Kai Hou Yip等；UCL／KCL／INAF", "doi": "10.48550/arXiv.2609.38566", "arxiv": "2609.38566", "url": "https://arxiv.org/abs/2609.38566", "version_date": "2026-09-29", "type": "arXiv预印本", "verification": "原始摘要及上述关键正文已核；作者报告，未独立复现", "verified_at": "2026-10-07T07:06:00.789510+08:00", "evidence": "基准情景完成696个Tier 1目标（85.5%）、129个Tier 3目标，分箱覆盖100%，观测效率72.6%；并非每个单项都优于启发式。", "read_first": "§3.6 → §4.1 → §5.3", "id": "arxiv:2609.38566", "appearances": [ "digests/scientific-computer-use/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences", "team": "Mingda Zhang、Wenjin Liu、Erik Cambria等", "doi": "10.48550/arXiv.2610.08691", "arxiv": "2610.08691", "url": "https://arxiv.org/abs/2610.08691", "type": "arXiv预印本 v1", "verification": "原始页面与所列关键正文已核；作者结果，未独立复现", "verified_at": "2026-10-08T15:12:02.640864+08:00", "evidence": "23学科基准；Table 2地球科学任务的IID与跨数据集OOD平均RMSE由冻结agent的1.0963 K降至Linked Full的0.9650 K。", "read_first": "§4重放／验证 → Table 2与§6 → §8局限", "id": "arxiv:2610.08691", "appearances": [ "digests/scientific-computer-use/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "A self-learning scientific agent for X-ray diffraction", "team": "Bin Cao等；港科大广州／UCL／长三角AI Lab／中科院自动化所", "doi": "10.48550/arXiv.2610.07862", "arxiv": "2610.07862", "url": "https://arxiv.org/abs/2610.07862", "type": "arXiv预印本 v1；正文冠名Gan Jiang", "verification": "原始页面与所列关键正文已核；作者结果，未独立复现", "verified_at": "2026-10-08T15:12:02.640864+08:00", "evidence": "无成分输入的opXRD单相top-1为40.83%，最强对照26.45%；这是物相识别准确率。", "read_first": "Fig.3／技能学习 → Supplementary Note 2.1的artifact交接与验收", "id": "arxiv:2610.07862", "appearances": [ "digests/scientific-computer-use/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Operating advanced scientific instruments with AI agents that learn on the job", "team": "Aikaterini Vriza、Michael H. Prince、Tao Zhou、Henry Chan、Mathew J. Cherukara；Argonne／APS团队", "doi": "10.1038/s41524-026-02005-0", "url": "https://www.nature.com/articles/s41524-026-02005-0", "type": "npj Computational Materials正式论文", "verification": "原始页面与所列关键正文已核；作者结果，未独立复现", "verified_at": "2026-10-08T15:12:02.640864+08:00", "evidence": "Fig.2任务要求1 μm定位；作者报告O3在测试模型中能稳定进行跨模态坐标推理，不能外推普遍1 μm精度。", "read_first": "Results／X-ray nanoprobe experiment → Fig.2、Fig.3及Supplementary 1.1", "id": "doi:10.1038/s41524-026-02005-0", "appearances": [ "digests/scientific-computer-use/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "An agentic artificially intelligent X-ray scientist", "team": "Zhantao Chen、Alexander N. Petsch等；Stanford／SLAC等", "doi": "10.1038/s42256-026-01261-5", "url": "https://www.nature.com/articles/s42256-026-01261-5", "type": "Nature Machine Intelligence正式论文", "verification": "原始页面与所列关键正文已核；作者结果，未独立复现", "verified_at": "2026-10-08T15:12:02.640864+08:00", "evidence": "Fig.3–4及Extended Data Fig.2记录偏移修正与后续找峰；真实试验因束线时间限制未按相同条件系统重复。", "read_first": "Deploying on a real synchrotron → Fig.3–4、Extended Data Fig.2 → Methods/MCP", "id": "doi:10.1038/s42256-026-01261-5", "appearances": [ "digests/scientific-computer-use/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Astra: an open-source fully autonomous robotic observatory control software", "team": "Peter P. Pedersen等；ETH／Cambridge／Liège／Bern等", "doi": "10.48550/arXiv.2607.12898", "arxiv": "2607.12898", "url": "https://arxiv.org/abs/2607.12898", "type": "arXiv v1；SPIE会议稿", "verification": "原始页面与所列关键正文已核；作者结果，未独立复现", "verified_at": "2026-10-08T15:12:02.640864+08:00", "evidence": "2026年1／5月数据的网络中位指向散布0.11 arcsec；作者称未发生归因Astra的调度中止，不等于设施从未中断。", "read_first": "§3.7站点定制／§4设备兼容 → §5真实观测结果", "id": "arxiv:2607.12898", "appearances": [ "digests/scientific-computer-use/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "MATCH: Performance Characterization and Scientific Observations of the Wuhan University 1-m Optical Telescope", "team": "Sai-En Xu、Bei You等；武汉大学／中科院南京天文光学技术研究所等", "doi": "10.48550/arXiv.2609.38858", "arxiv": "2609.38858", "url": "https://arxiv.org/abs/2609.38858", "type": "arXiv预印本 v1", "verification": "原始摘要／作者身份／v1日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-09T07:08:36.654632+08:00", "evidence": "典型条件60秒g波段5σ深度约20 mag；光谱1800秒B/G约15–16 mag。深度指能检测多暗的目标，曝光更长不能等同排队更短。", "read_first": "§VI.4–VI.5 → §VII与灵敏度结果", "id": "arxiv:2609.38858", "appearances": [ "digests/scientific-computer-use/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "SciExam for ENSO: Can AI Agents Build Climate Models?", "team": "Yinling Zhang、Langchen Liu、Mengdi Wang、Shilong Liu等；Ohio State／Yale／UCSD／Stanford／Princeton", "doi": "10.48550/arXiv.2610.10513", "arxiv": "2610.10513", "url": "https://arxiv.org/abs/2610.10513", "type": "arXiv预印本 v1", "verification": "原始摘要／作者身份／v1日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-09T07:08:36.654632+08:00", "evidence": "六小时eligible预算下12系统中6个最终模型超过参考综合分0.378；参考仍最擅长统计再现。最佳checkpoint与最终提交会改变数量。", "read_first": "§3.2–3.4 → §4 → Appendix B/C", "id": "arxiv:2610.10513", "appearances": [ "digests/scientific-computer-use/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations", "team": "Toby D. Pilditch、Konstantinos Voudouris、Alexandra Abbas、Cozmin Ududec；UK AI Security Institute／Meridian Labs", "doi": "10.48550/arXiv.2610.08364", "arxiv": "2610.08364", "url": "https://arxiv.org/abs/2610.08364", "type": "arXiv预印本 v1", "verification": "原始摘要／作者身份／v1日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-09T07:08:36.654632+08:00", "evidence": "单个近1300万token案例含554主代理工具调用、79子代理部署、8次压缩；580输出仅2个标Hypothesising。", "read_first": "§3.1与Fig.2 → §3.5 → Appendix A.5", "id": "arxiv:2610.08364", "appearances": [ "digests/scientific-computer-use/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "AI Agents as a Research Team: A Human-Supervised Case Study in Stellar Spectroscopy", "team": "Marwan Gebran、Ian Bentley；Saint Mary’s College／Florida Polytechnic University", "doi": "10.48550/arXiv.2610.05424", "arxiv": "2610.05424", "url": "https://arxiv.org/abs/2610.05424", "type": "arXiv预印本 v1", "verification": "原始摘要／作者身份／v1日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-09T07:08:36.654632+08:00", "evidence": "scaler错配的MAE约0.024，纠正后0.003681（归一化光通量）；改变了预处理身份与精度，不能归因为float32更准。", "read_first": "Appendix A.3–A.4 → §2.4与§7.1.2", "id": "arxiv:2610.05424", "appearances": [ "digests/scientific-computer-use/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Explore, Then Commit: Measurement-Efficient Scientific Law Discovery with Language Models", "team": "Kautik Mandve、Dileepa Fernando；Vernon Hills High School／SUTD", "doi": "10.48550/arXiv.2610.07620", "arxiv": "2610.07620", "url": "https://arxiv.org/abs/2610.07620", "type": "arXiv预印本 v1", "verification": "原始摘要／作者身份／v1日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-09T07:08:36.654632+08:00", "evidence": "576试验：medium解释器planner用8.6/8.9次测量，对照22.5/43.0；审计发现29条不完整结果。", "read_first": "§2.2、Table 2 → §3.3与§4", "id": "arxiv:2610.07620", "appearances": [ "digests/scientific-computer-use/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Prior or Feedback? What an LLM Uses When Adapting Neural Operators", "team": "Julian Chan（Surrey）、Javier Mora Jimenez（Compare the Market）", "doi": "10.48550/arXiv.2610.12325", "arxiv": "2610.12325", "url": "https://arxiv.org/abs/2610.12325", "type": "arXiv预印本；元数据标工作坊接收", "verification": "原始身份/日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-10T07:11:22.654160+08:00", "evidence": "12迁移单元×3种子；20试验各策略；测试nRMSE对随机36/36、TPE35/36更低；分数置换相对保值改写动作距离中位差0.085。", "read_first": "§3反馈回放、§4结果", "id": "arxiv:2610.12325", "appearances": [ "digests/scientific-computer-use/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "LLM-IDEA: Identifiability-Driven Experimental Agent for Autonomous Discovery of Mechanistic World Models", "team": "Surya Shetty（Texas A&M／Polara Labs）、Ulisses Braga-Neto（Texas A&M）", "doi": "10.48550/arXiv.2610.11253", "arxiv": "2610.11253", "url": "https://arxiv.org/abs/2610.11253", "type": "arXiv预印本", "verification": "原始身份/日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-10T07:11:22.654160+08:00", "evidence": "63 ODE系统中62有常数，60初始可辨识；RC三常数只进入两组合，同类实验不可分；首版数值误判后修正方法。", "read_first": "§5.8 RC案例、§5.4、§6", "id": "arxiv:2610.11253", "appearances": [ "digests/scientific-computer-use/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Can AI Agents Make Open-Ended Scientific Discovery? Evidence from Station", "team": "Wenyu Du、Stephen Chung；DualverseAI／香港大学／剑桥大学", "doi": "10.48550/arXiv.2610.08927", "arxiv": "2610.08927", "url": "https://arxiv.org/abs/2610.08927", "type": "arXiv预印本", "verification": "原始身份/日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-10T07:11:22.654160+08:00", "evidence": "3项ICLR再发现任务33判据，各3运行；任务平均恢复62.7%，LLM评分；24提案75%专家判科学意义低。", "read_first": "§6.2人类评价，监督／反思机制", "id": "arxiv:2610.08927", "appearances": [ "digests/scientific-computer-use/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "The missing map of the sky", "team": "Brice Ménard（Johns Hopkins）与Anthropic Claude Science", "url": "https://www.anthropic.com/research/the-missing-map-of-the-sky", "type": "官方实质科研技术文章（非已核正式论文）", "verification": "原始身份/日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-10T07:11:22.654160+08:00", "evidence": "GALEX约38000观测覆盖2/3天；余区推断；文章称验证约10%偏差，误差定义划分未充分报告。", "read_first": "原文遮蔽验证及圆形伪影段", "id": "url:https://www.anthropic.com/research/the-missing-map-of-the-sky", "appearances": [ "digests/scientific-computer-use/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "JaxRobotarium: Training and Deploying Multi-Robot Policies in 10 Minutes", "team": "Shalin Anand Jain、Jiazhen Liu、Siva Kailas、Harish Ravichandar；Georgia Tech STAR Lab", "doi": "10.48550/arXiv.2505.06771", "arxiv": "2505.06771", "url": "https://arxiv.org/abs/2505.06771", "version_date": "2025-11-10", "type": "CoRL 2025论文／arXiv v3", "verification": "原始身份/日期及所列正文已核；作者报告，未独立复现", "verified_at": "2026-10-10T07:11:22.654160+08:00", "evidence": "8任务×4算法×5硬件初始实例；追逐捕获PQN4.4→1.4、QMIX0.4→3.6，排序反转；仿真5策略各100评估。", "read_first": "表4、§6.1、§9", "id": "arxiv:2505.06771", "appearances": [ "digests/scientific-computer-use/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "What Output-Only Review Cannot Verify: Study Contracts for Research Agents", "team": "Eitan Waks、Ben Glocker；Imperial College London", "doi": "10.48550/arXiv.2610.11754", "arxiv": "2610.11754", "url": "https://arxiv.org/abs/2610.11754", "version_date": "2026-10-08", "type": "arXiv预印本", "verification": "原始身份与日期、正文所列方法/结果/补充已核；作者报告，未独立复现", "verified_at": "2026-10-11T07:07:49.458308+08:00", "evidence": "8组自建正常/改坏包；确定检查器8/8；18别名×8改坏包=144判断，104缺陷、32弃权、8终止失败，0明确正常", "read_first": "§3.3、§4、表1", "first_published": "2026-10-08", "verified_version": "v1", "limitations": "检查器与评审信息和规则不等；不是模型质量对照，也未证明reward hacking", "fulltext_url": "https://arxiv.org/html/2610.11754v1", "id": "arxiv:2610.11754", "appearances": [ "digests/scientific-computer-use/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "VISION: A Modular AI Assistant for Natural Human-Instrument Interaction at Scientific User Facilities", "team": "Mathur、van der Vleuten、Yager、Tsai；Brookhaven National Laboratory", "doi": "10.1088/2632-2153/add9e4", "arxiv": "2412.18161", "url": "https://arxiv.org/abs/2412.18161", "version_date": "2024-12-24", "type": "期刊关联论文／arXiv版本", "verification": "原始身份与日期、正文所列方法/结果/补充已核；作者报告，未独立复现", "verified_at": "2026-10-11T07:07:49.458308+08:00", "evidence": "§3.2真实束线操作采集分析演示；模块F1和响应时间不是实验总成功率", "read_first": "§3.2、S1.3", "first_published": "2024-12-24", "verified_version": "v1", "limitations": "无同预算GUI/API科研比较；迁移人时和总排队时延未知", "fulltext_url": "https://arxiv.org/html/2412.18161v1", "id": "arxiv:2412.18161", "appearances": [ "digests/scientific-computer-use/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Integrating Domain-Specialized Language Models with AI Measurement Tools for Deterministic Atomic-Resolution Experimentation", "team": "Diao、Matsumoto、Hou、Ohara、Yamashita、Abe；大阪大学", "doi": "10.48550/arXiv.2602.20669", "arxiv": "2602.20669", "url": "https://arxiv.org/abs/2602.20669", "version_date": "2026-04-06", "type": "arXiv预印本", "verification": "原始身份与日期、正文所列方法/结果/补充已核；作者报告，未独立复现", "verified_at": "2026-10-11T07:07:49.458308+08:00", "evidence": "命令Stage i99.3%、Stage ii95.2%；方法表总命令测试181条，阶段分母未明确拆出，非181硬件实验", "read_first": "图1、S3、命令解析方法", "first_published": "2026-02-24", "verified_version": "v2", "limitations": "专用数据与模块预制；模型命令准确率不等于系统成功率", "fulltext_url": "https://arxiv.org/html/2602.20669v2", "id": "arxiv:2602.20669", "appearances": [ "digests/scientific-computer-use/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "AlabOS: A Python-based Reconfigurable Workflow Management Framework for Autonomous Laboratories", "team": "Yuxing Fei等；UC Berkeley／LBNL，Ceder与Jain团队", "doi": "10.1039/d4dd00129j", "arxiv": "2405.13930", "url": "https://arxiv.org/abs/2405.13930", "version_date": "2024-08-30", "type": "期刊关联论文／arXiv版本", "verification": "原始身份与日期、正文所列方法/结果/补充已核；作者报告，未独立复现", "verified_at": "2026-10-11T07:07:49.458308+08:00", "evidence": "约1.5年3500不同样品；图6流程，运营规模非目标合成成功率", "read_first": "§2.1、§3.1–3.3、图6", "first_published": "2024-05-22", "verified_version": "v2", "limitations": "人工储存/恢复仍存在；专用驱动与任务定义必需，无接口成本公平比较", "fulltext_url": "https://arxiv.org/html/2405.13930v2", "id": "arxiv:2405.13930", "appearances": [ "digests/scientific-computer-use/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents", "team": "Lijie Ding、Changwoo Do；Oak Ridge National Laboratory", "doi": "10.48550/arXiv.2610.03631", "arxiv": "2610.03631", "url": "https://arxiv.org/abs/2610.03631", "version_date": "2026-10-02", "type": "arXiv预印本", "verification": "原始身份与日期、正文所列方法/结果/补充已核；作者报告，未独立复现", "verified_at": "2026-10-11T07:07:49.458308+08:00", "evidence": "16任务7模型每任务方式1回，最高7/16；300留出实例Qwen3-8B34→230通过，第二种子69%；物理规则＋同预算优化81%与76.7%未分开", "read_first": "表1、图2、表2", "first_published": "2026-10-02", "verified_version": "v1", "limitations": "小基准不能模型排名；种子不稳定、专用工具与判据；非物理闭环", "fulltext_url": "https://arxiv.org/html/2610.03631v1", "id": "arxiv:2610.03631", "appearances": [ "digests/scientific-computer-use/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Vision-Language Agents for Active Perception in Optics Laboratories", "team": "Ryan Lopez、Sachin Vaidya等；MIT物理系/RLE/IAIFI", "doi": "10.48550/arXiv.2609.32918", "arxiv": "2609.32918", "url": "https://arxiv.org/abs/2609.32918", "fulltext_url": "https://arxiv.org/html/2609.32918v1", "version_date": "2026-09-26", "type": "arXiv预印本", "verification": "原始身份/日期及正文相关方法与结果已核；未独立复现，未核全部补充材料", "verified_at": "2026-10-12T07:09:36.095968+08:00", "evidence": "每任务10初始状态；预算60/50/100；四镜GPT9/10、Gemini10/10；双镜500步阈值9/10、6/10、0/10；自报完成不等于通过", "read_first": "§3.4图4、§3.3和§3.5", "first_published": "2026-09-26", "verified_version": "v1", "limitations": "小样本、任务指导与固定观测位置；BO只在模拟比较，获得不同观测；未测跨仪器接入成本", "id": "arxiv:2609.32918", "appearances": [ "digests/scientific-computer-use/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "RFChipAgent: Multi-Agentic AI Flow for Analog/RF Chip Design", "team": "Awani Khodkumbhe等；Qualcomm Technologies，首作者兼UC Berkeley", "doi": "10.48550/arXiv.2610.10858", "arxiv": "2610.10858", "url": "https://arxiv.org/abs/2610.10858", "fulltext_url": "https://arxiv.org/html/2610.10858v1", "version_date": "2026-10-07", "type": "arXiv预印本", "verification": "原始身份/日期及正文相关方法与结果已核；未独立复现，未核全部补充材料", "verified_at": "2026-10-12T07:09:36.095968+08:00", "evidence": "一个两级LNA设计案例；保留前五候选只有#174五规格同时达标；27.1dB增益/4.2dB噪声/14.3mW；约2h作者报告，无匹配人工对照", "read_first": "§3.3–3.5和表1", "first_published": "2026-10-07", "verified_version": "v1", "limitations": "工艺/库/模板依赖；单案例、无同预算基线；作者首次和大幅人工节省主张不外推", "id": "arxiv:2610.10858", "appearances": [ "digests/scientific-computer-use/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "AgentHorizon: Evaluating Agentic Judges for Long-Horizon Computer-Use Tasks", "team": "Xing Han Lù等；ServiceNow Research、McGill/Mila等", "doi": "10.48550/arXiv.2610.11050", "arxiv": "2610.11050", "url": "https://arxiv.org/abs/2610.11050", "fulltext_url": "https://arxiv.org/html/2610.11050v1", "version_date": "2026-10-08", "type": "arXiv预印本", "verification": "原始身份/日期及正文相关方法与结果已核；未独立复现，未核全部补充材料", "verified_at": "2026-10-12T07:09:36.095968+08:00", "evidence": "1373要求轨迹对；困难528；最佳80.9%平衡准确率=(71.8%正例接受+90.0%负例拒绝)/2；Qwen27B工具式下降", "read_first": "§3.2、表2、§4.2", "first_published": "2026-10-08", "verified_version": "v1", "limitations": "模型定义困难分组、人轨迹分布、预处理/框架混杂；非科学验收标准", "id": "arxiv:2610.11050", "appearances": [ "digests/scientific-computer-use/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Accelerating Scientific Research with Gemini in the Real-World", "team": "Samuel Schmidgall等；Google及合作实验团队", "doi": "10.48550/arXiv.2608.26701", "arxiv": "2608.26701", "url": "https://arxiv.org/abs/2608.26701", "fulltext_url": "https://arxiv.org/html/2608.26701v1", "version_date": "2026-08-27", "type": "arXiv预印本", "verification": "原始身份/日期及正文相关方法与结果已核；未独立复现，未核全部补充材料", "verified_at": "2026-10-12T07:09:36.095968+08:00", "evidence": "272候选中的#2路线25轮；MXene类似结构未原子级确认；三类TMD首次生长约1h，分钟指推理；快速晶域更小更不规则", "read_first": "§3.1.2–3.1.4、图2/3", "first_published": "2026-08-27", "verified_version": "v1", "limitations": "无等预算专家对照、硬件定制与人工处理", "id": "arxiv:2608.26701", "appearances": [ "digests/scientific-computer-use/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "EvoMaster: A Foundational Evolving Agent Framework for Agentic Science at Scale", "team": "Xinyu Zhu等；上海交通大学/SciLand，Weinan E、Siheng Chen等", "doi": "10.48550/arXiv.2604.17406", "arxiv": "2604.17406", "url": "https://arxiv.org/abs/2604.17406", "fulltext_url": "https://arxiv.org/html/2604.17406v5", "version_date": "2026-09-10", "type": "arXiv预印本", "verification": "原始身份/日期及正文相关方法与结果已核；未独立复现，未核全部补充材料", "verified_at": "2026-10-12T07:09:36.095968+08:00", "evidence": "十基准同GPT5.4 medium、时间限制/评分/划分；非加权均值58.02%对Codex40.29%，PostTrainBench Codex领先", "read_first": "§3.1–3.3、表2和§1.3", "first_published": "2026-04-19", "verified_version": "v5", "limitations": "异质基准平均非物理成功率；机制未完全隔离、接入人时未知；生物基准仅原文评价构成非项目方向", "id": "arxiv:2604.17406", "appearances": [ "digests/scientific-computer-use/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Human-agent discovery of reconfigurable in-plane ferroelectric superdomain control", "team": "UTK、Rice、UC Berkeley；Yu Liu、Sergei V. Kalinin等", "doi": "10.48550/arXiv.2609.06887", "arxiv": "2609.06887", "url": "https://arxiv.org/abs/2609.06887", "fulltext_url": "https://arxiv.org/pdf/2609.06887", "version_date": "2026-09-07", "type": "arXiv预印本", "verification": "原始身份/日期及相关方法结果正文已核；未复现，未核完所有补充数据和代码", "verified_at": "2026-10-13T09:35:44.441203+08:00", "read_first": "§3.2—3.3、图2、§8.3—8.4", "first_published": "2026-09-07", "verified_version": "v1", "evidence": "9代理闭环周期、8自行停止非成功率；同极性/交替极性匹配暴露对照；0.415→0.601为未处理膜方向比例", "limitations": "单仪器、单薄膜，人工共建表示和检查；无GUI/API公平比较、无迁移人时", "id": "arxiv:2609.06887", "appearances": [ "digests/scientific-computer-use/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Hypothesis-Driven Autonomous Materials Synthesis with Multimodal LLM Agents", "team": "University of Tokyo、Rigaku；Takahara、Hitosugi、Mizoguchi等", "doi": "10.48550/arXiv.2609.18598", "arxiv": "2609.18598", "url": "https://arxiv.org/abs/2609.18598", "fulltext_url": "https://arxiv.org/html/2609.18598v1", "version_date": "2026-09-16", "type": "arXiv预印本", "verification": "原始身份/日期及相关方法结果正文已核；未复现，未核完所有补充数据和代码", "verified_at": "2026-10-13T09:35:44.441203+08:00", "read_first": "图3逐轮理解变化、Methods", "first_published": "2026-09-16", "verified_version": "v1", "evidence": "18次一维单材料；最佳670°C峰比61.6，窗口650–690°C；每条件一次", "limitations": "无等预算实物BO对照；技能自测、LLM判断无统一数值容差；基础设施已集成", "id": "arxiv:2609.18598", "appearances": [ "digests/scientific-computer-use/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "JANUS: Enhancing Long-Horizon Autonomous Research with Co-Evolving World Models", "team": "ICT CAS、UCAS；Mengdi Liu、Chenghao Jia、Hong Chang等", "doi": "10.48550/arXiv.2610.13194", "arxiv": "2610.13194", "url": "https://arxiv.org/abs/2610.13194", "fulltext_url": "https://arxiv.org/html/2610.13194v1", "version_date": "2026-10-09", "type": "arXiv预印本", "verification": "原始身份/日期及相关方法结果正文已核；未复现，未核完所有补充数据和代码", "verified_at": "2026-10-13T09:35:44.441203+08:00", "read_first": "图2、§4.1、表1", "first_published": "2026-10-09", "verified_version": "v1", "evidence": "30任务12领域；单A100每题5h同设置；同Flash主干JANUS13/30 BeatAll，Codex11/30；测试归一分41.07 vs38.91", "limitations": "计算领域，不迁移硬件结论；主干不同不可混比；预测不等于测量", "id": "arxiv:2610.13194", "appearances": [ "digests/scientific-computer-use/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Large Language Model-Assisted Superconducting Qubit Experiments", "team": "University of Chicago为主；Shiheng Li、Andrew N. Cleland等", "doi": "10.48550/arXiv.2603.08801", "arxiv": "2603.08801", "url": "https://arxiv.org/abs/2603.08801", "fulltext_url": "https://arxiv.org/html/2603.08801v1", "version_date": "2026-03-09", "type": "arXiv预印本", "verification": "原始身份/日期及相关方法结果正文已核；未复现，未核完所有补充数据和代码", "verified_at": "2026-10-13T09:35:44.441203+08:00", "read_first": "五周期谐振器案例、知识库部分", "first_published": "2026-03-09", "verified_version": "v1", "evidence": "谐振器5周期从4到8；约5min包含VNA扫描不含准备/排队；QND泄漏.124±.017非代理错误率", "limitations": "两案例无重复比较；驱动知识库已建；无新硬件适配人时", "id": "arxiv:2603.08801", "appearances": [ "digests/scientific-computer-use/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "radio-astro-tools: linking radio astronomical data to the astronomical Python ecosystem", "team": "NRC Canada、UF、NRAO等；Eric W. Koch、Adam Ginsburg等", "doi": "10.48550/arXiv.2610.11429", "arxiv": "2610.11429", "url": "https://arxiv.org/abs/2610.11429", "fulltext_url": "https://arxiv.org/html/2610.11429v1", "version_date": "2026-10-08", "type": "arXiv预印本", "verification": "原始身份/日期及相关方法结果正文已核；未复现，未核完所有补充数据和代码", "verified_at": "2026-10-13T09:35:44.441203+08:00", "read_first": "§1.1、§5.2、附录B图5", "first_published": "2026-10-08", "verified_version": "v1", "evidence": "附录B15角秒单天线、2–40角秒理想干涉响应；相同输入feather，指数差<1σ", "limitations": "特定模拟数值相容非全部真实数据等价；无速度或接入人时对照", "id": "arxiv:2610.11429", "appearances": [ "digests/scientific-computer-use/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Stress-testing large language model agents in a robotic chemistry laboratory", "team": "中国科学技术大学、合肥科学智能创新中心、伯明翰大学；Lulu Guo等", "doi": "10.48550/arXiv.2607.23045", "arxiv": "2607.23045", "url": "https://arxiv.org/abs/2607.23045", "fulltext_url": "https://arxiv.org/pdf/2607.23045", "version_date": "2026-07-25", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-14T07:05:59.937977+08:00", "read_first": "图2 N9/N10、图4与Iterative adaptation test", "evidence": "48组合×32任务×3重复=4608；607调度后151专家可执行；最佳27/96；短任务训练证据不涉及该实验", "first_published": "2026-07-25", "verified_version": "v1", "limitations": "4608是计划评测次数，不是已完成真实实验数；151为专家认为可无人工干预执行。第二轮前人类指出回收量不足；没有核实第五轮全部测量完成。设备迁移人时与通用GUI对照未知。", "id": "arxiv:2607.23045", "appearances": [ "digests/scientific-computer-use/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "SVOM Follow-up Observation Coordinating Service", "team": "国家天文台、国科大及合作台站；Xu-hui Han等", "doi": "10.48550/arXiv.2511.06647", "arxiv": "2511.06647", "url": "https://arxiv.org/abs/2511.06647", "fulltext_url": "https://arxiv.org/html/2511.06647v1", "version_date": "2025-11-10", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-14T07:05:59.937977+08:00", "read_first": "§3.2、§5、表1", "evidence": "CGFT97%警报消息<3秒，GWAC90%计划消息<10秒；消息数未给，网络/载荷不同；非排队/曝光/分析耗时", "first_published": "2025-11-10", "verified_version": "v1", "limitations": "预设曝光策略与人类再观测请求不同于语言代理自主改变研究设计。两组传输比例条件不同且消息分母未给，不包含排队、转台、曝光和分析时间；开放使用资格未知。", "id": "arxiv:2511.06647", "appearances": [ "digests/scientific-computer-use/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction", "team": "上海交大、复旦、南洋理工、浙江大学、上海奇绩智峰；Yuqi Huang等", "doi": "10.48550/arXiv.2610.12984", "arxiv": "2610.12984", "url": "https://arxiv.org/abs/2610.12984", "fulltext_url": "https://arxiv.org/html/2610.12984v1", "version_date": "2026-10-09", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-14T07:05:59.937977+08:00", "read_first": "§3.1–3.3、表2", "evidence": "350成功记录=64+286；8.48小时记录；平均分增4.67–10.54/3.11–7.50；GPT在64任务31/64 vs24/64", "first_published": "2026-10-09", "verified_version": "v1", "limitations": "数据集筛选成功完整演示，不能代表任意科研软件；GUI、Bash及无障碍信息均可能参与。分数增幅不是科学准确率，没有专用API总接入工时对照。", "id": "arxiv:2610.12984", "appearances": [ "digests/scientific-computer-use/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "MAGMA-Bench: Evaluating Long-Horizon Embodied Agents under Dynamic Tasks and Execution Failures", "team": "Siléane、LAAS-CNRS；Loan Bernat等", "doi": "10.48550/arXiv.2610.13534", "arxiv": "2610.13534", "url": "https://arxiv.org/abs/2610.13534", "fulltext_url": "https://arxiv.org/pdf/2610.13534", "version_date": "2026-10-09", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-14T07:05:59.937977+08:00", "read_first": "§4.3–4.5、§5.1、表4", "evidence": "99任务规格1254回合；261646训练样本；短165回合Qwen全完成2.42→72.73%=约4→120；历史全/32/8消息干净SR5.7/3.0/1.0%", "first_published": "2026-10-09", "verified_version": "v1", "limitations": "任务在模拟器和高层技能接口执行；固定故障仅在符合条件的调用中暴露。训练测试共享场景词汇，不能证明真实仪器、网络或科学数据误差下的迁移能力。", "id": "arxiv:2610.13534", "appearances": [ "digests/scientific-computer-use/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Autonomous Code Migration for Quantum Programming Languages: A Case Study with QED-C Benchmarks", "team": "NVIDIA、劳伦斯伯克利国家实验室、Quantum Computing Data等；W. Michael Brown等", "doi": "10.48550/arXiv.2610.12187", "arxiv": "2610.12187", "url": "https://arxiv.org/abs/2610.12187", "fulltext_url": "https://arxiv.org/html/2610.12187v1", "version_date": "2026-10-08", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-14T07:05:59.937977+08:00", "read_first": "§III-C、§IV-A、表II", "evidence": "每模型一次13方法迁移；匿名模型用户提示33/20/26；数工作日为作者经验，人工数周/月是估计非对照", "first_published": "2026-10-08", "verified_version": "v1", "limitations": "无噪声软件模拟而非现场量子测量；每模型仅一次13方法迁移，匿名编号不能形成稳定排名。数工作日是作者经验，人工数周/月是估计而非同条件对照。", "id": "arxiv:2610.12187", "appearances": [ "digests/scientific-computer-use/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Principles and Techniques for Agentic Research in Experimental Physics", "team": "马里兰大学Walsworth团队；Niko R. Reed等", "doi": "10.48550/arXiv.2610.17193", "arxiv": "2610.17193", "url": "https://arxiv.org/abs/2610.17193", "fulltext_url": "https://arxiv.org/html/2610.17193v1", "first_published": "2026-10-12", "verified_version": "v1", "version_date": "2026-10-12", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-15T07:08:53.406155+08:00", "evidence": "20.8小时；首评分3.2小时；最佳10分钟0.97Hz；相同设置10次中位1.47Hz；25/34漂移受限；独立审计否定重复纪录", "read_first": "§III.4、§IV.5和图6", "id": "arxiv:2610.17193", "limitations": "作者报告20.8小时、151次实验，其中34次正式评分；首次评分用时3.2小时。最佳10分钟0.97Hz，相同设置10次中位1.47Hz，25/34次漂移受限；独立审计否定了所声称的重复纪录。仪器、样品、八步协议和安全范围由人设定，人工耗时参照非配对试验，未提供跨仪器接入人时或公平 GUI 对照。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Delegating Earth science model calibration to knowledge-guided LLM agents", "team": "河海大学、南京水利科学研究院、Meta；Zhanliang Zhu等", "doi": "10.48550/arXiv.2610.18378", "arxiv": "2610.18378", "url": "https://arxiv.org/abs/2610.18378", "fulltext_url": "https://arxiv.org/pdf/2610.18378", "first_published": "2026-10-13", "verified_version": "v1", "version_date": "2026-10-13", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-15T07:08:53.406155+08:00", "evidence": "FSM2知识5/5首次可执行、仅名0/5；验证NSE0.81–0.94对默认0.57；VIC10配对各臂150评估，差中位−0.003，95%区间−0.007至0.002，非劣界−0.02；11单向量回放", "read_first": "§2.2、§2.6、§3.4及图5", "id": "arxiv:2610.18378", "limitations": "FSM2 首次可执行为有知识5/5、仅模型名0/5，验证NSE为0.81–0.94对默认0.57。VIC 的10组配对、各150次评估下差值中位约−0.003，95%区间−0.007至0.002，非劣界−0.02；结论限于指定预算与目标。10个中断代理臂按规则排除，11次单向量回放不是完整搜索重放，多变量选择部分使用验证信息。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation", "team": "Google Earth AI；Noam Slonim等", "doi": "10.48550/arXiv.2610.15750", "arxiv": "2610.15750", "url": "https://arxiv.org/abs/2610.15750", "fulltext_url": "https://arxiv.org/html/2610.15750v1", "first_published": "2026-10-11", "verified_version": "v1", "version_date": "2026-10-11", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-15T07:08:53.406155+08:00", "evidence": "76测试开放问题、每配置3运行；Flash基础3.3、领域52.2、全系统62.7；Pro领域50.8加计划47.6；得分是流程价值对齐非科学正确率", "read_first": "§3、表4和§6局限", "id": "arxiv:2610.15750", "limitations": "76个开放问题、每配置3次运行；Flash 基础3.3、领域52.2、完整62.7，Pro领域50.8加计划47.6。分数衡量流程价值对齐，不是科学正确率。专家参考流程及69条轨迹的207次标注支持裁判核查；基础 Python 不联网、工具能力不等、不可解任务过滤和仅Gemini测试限制比较。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Stateless Language Agents: Scaling Long-Horizon Automated Research", "team": "Stanford、CMU、华盛顿大学、SambaNova、Berkeley；Qizheng Zhang等", "doi": "10.48550/arXiv.2610.07625", "arxiv": "2610.07625", "url": "https://arxiv.org/abs/2610.07625", "fulltext_url": "https://arxiv.org/html/2610.07625v1", "first_published": "2026-10-06", "verified_version": "v1", "version_date": "2026-10-06", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-15T07:08:53.406155+08:00", "evidence": "内核Codex每方法3运行，SLA3次均优于基线合计9次；共同目标67.9M对986.3M token；93.1%为推理预算减少非墙钟；共享检查点消融", "read_first": "§3、§4.3及附录D.3", "id": "arxiv:2610.07625", "limitations": "内核 Codex 每方法3次运行，SLA 的3次均优于三组基线合计9次；达到共同目标用67.9M对986.3M token，93.1%是推理预算减少而非墙钟提速。共享检查点消融及短预算下排序变化提示预算依赖；自动固定评价器的结果不能直接推广到物理漂移和科学目标修订。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "PEAR: A Robust and Flexible Automation Framework for Ptychography Enabled by Multiple Large Language Model Agents", "team": "Argonne国家实验室、Rice大学；Xiangyu Yin、Chuqiao Shi、Yimo Han、Yi Jiang", "doi": "10.48550/arXiv.2410.09034", "arxiv": "2410.09034", "url": "https://arxiv.org/abs/2410.09034", "fulltext_url": "https://arxiv.org/html/2410.09034v1", "first_published": "2024-10-11", "verified_version": "v1", "version_date": "2024-10-11", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-15T07:08:53.406155+08:00", "evidence": "100组假设条件，单/多代理跨模型脚本与专家参考比较；可跑参数错与语法错分列；SnSe图像案例；不复述未核图中精确百分比", "read_first": "§2.3、图2、§3及附录B", "id": "arxiv:2410.09034", "limitations": "2024年论文的100种假设情形用GPT-4模拟用户，与专家参考比较，不是100次无人值守新物理实验。人工 Level 1 图像反馈能改变下一步参数。贝叶斯优化对比来自前作，非同预算对照；没有独立复现，不能从重构流程推断远程设备开放资格、费用或时延。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Qumus: Realization of An Embodied AI Quantum Material Experimentalist", "team": "Princeton、Michigan、NIMS；Lihan Shi、Zhaoyi Joy Zheng等", "doi": "10.48550/arXiv.2605.18407", "arxiv": "2605.18407", "url": "https://arxiv.org/abs/2605.18407", "fulltext_url": "https://arxiv.org/pdf/2605.18407", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-16T07:11:42.621816+08:00", "evidence": "一次清史运行5实验超过4h，目标石墨烯>200平方微米；器件案例90min/30步/18决策调用，非统计成功率", "read_first": "图3、Methods硬件及数据库", "id": "arxiv:2605.18407", "first_published": "2026-05-18", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-05-18", "limitations": "专用工程显著；无GUI/API等预算对照或跨仪器接入人时；代码数据可获得性未下载核实" }, { "title": "CosmoAgent: AI Agent for Theoretical Cosmology", "team": "Brown物理/BCTPI；Stephon Alexander、Loukas Gouskos、Savvas M. Koushiappas等", "doi": "10.48550/arXiv.2610.16536", "arxiv": "2610.16536", "url": "https://arxiv.org/abs/2610.16536", "fulltext_url": "https://arxiv.org/html/2610.16536v1", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-16T07:11:42.621816+08:00", "evidence": "8模型各1000母理论；每理论1000初始参数且可追加3000/5000；最佳deltaAIC−3.328；8组BIC均偏好LCDM；记忆消融250母理论", "read_first": "第五、六节及附录G", "id": "arxiv:2610.16536", "first_published": "2026-10-12", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-10-12", "limitations": "等母理论非等算量；多重搜索效应；记忆消融少量运行；不能当新理论发现" }, { "title": "Into the Unknown: Benchmarking AI agents on open scientific research problems", "team": "CMU及Los Alamos、Sandia、LBNL、LLNL、Ames等；Seungone Kim等", "doi": "10.48550/arXiv.2610.19159", "arxiv": "2610.19159", "url": "https://arxiv.org/abs/2610.19159", "fulltext_url": "https://arxiv.org/pdf/2610.19159", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-16T07:11:42.621816+08:00", "evidence": "9代理135提交0全通过；114一致性失败，90/85/68非互斥；自动裁判非135份均人工；附录41目标判断一致", "read_first": "2.1节、主表及附录A", "id": "arxiv:2610.19159", "first_published": "2026-10-13", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-10-13", "limitations": "开放任务未解不代表总体能力；评判模型偏差；105诊断子集不混主表135；设置版本残留7代理" }, { "title": "The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery", "team": "Sakana AI；Robert Tjarko Lange等", "doi": "10.48550/arXiv.2610.18933", "arxiv": "2610.18933", "url": "https://arxiv.org/abs/2610.18933", "fulltext_url": "https://arxiv.org/html/2610.18933v1", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-16T07:11:42.621816+08:00", "evidence": "5任务14配置70研究运行；人工核验ALIFE案例；三个Pi RL运行2个改变测试采样条件，人工排除", "read_first": "3.3–3.5及6.3节", "id": "arxiv:2610.18933", "first_published": "2026-10-13", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-10-13", "limitations": "自动论文分数非真值；规则遗漏使比较对象变化；共享记忆最终成绩小样本不显著" }, { "title": "How GUI Agents Fail: GraceBench for Execution Stability Diagnosis", "team": "同济大学；Ridong Zhang、Shuxian Liang等", "doi": "10.48550/arXiv.2610.19900", "arxiv": "2610.19900", "url": "https://arxiv.org/abs/2610.19900", "fulltext_url": "https://arxiv.org/html/2610.19900v1", "type": "arXiv预印本", "verification": "原始身份、日期及相关正文核验；未复现，未核全部附录/代码/原始数据", "verified_at": "2026-10-16T07:11:42.621816+08:00", "evidence": "10应用2909状态3116边；169任务两指令11模型3718轨迹；1294成功中663有偏离/循环=51.24%；GPT5.6成功率63.91%", "read_first": "3.3、4.3及附录E", "id": "arxiv:2610.19900", "first_published": "2026-10-14", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-10-14", "limitations": "状态图不穷尽；不同模型恢复机会不同；非故障注入同条件对照；无科研跨应用/硬件实证" }, { "title": "Agent-assisted experimental strategy transfer across quantum processors", "team": "四川师范大学、深圳大学、SpinQ；Shi-Yao Hou、Keren Li等", "doi": "10.48550/arXiv.2610.21047", "arxiv": "2610.21047", "url": "https://arxiv.org/abs/2610.21047", "fulltext_url": "https://arxiv.org/html/2610.21047v1", "type": "arXiv预印本", "verification": "原始身份/版本与相关正文核验；未复现、未核尽代码/原始数据/引用", "verified_at": "2026-10-17T07:09:45.339671+08:00", "evidence": "四台主导本征态投影占据0.910–0.959；五完整重复会话×3=15测量；另有参考失败停止", "read_first": "补充S2、表S2、图3、S4", "id": "arxiv:2610.21047", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-10-15", "limitations": "统一处理器家族/接口；无人类或固定流程对照；投影值非原始物理占据；上下文未强制隔离；Q3进入策略部分事后重建" }, { "title": "From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research", "team": "上海人工智能实验室、上海交通大学、香港科技大学（广州）、南洋理工大学等；Yijia Mao、Song Dai等", "doi": "10.48550/arXiv.2610.21767", "arxiv": "2610.21767", "url": "https://arxiv.org/abs/2610.21767", "fulltext_url": "https://arxiv.org/html/2610.21767v1", "type": "arXiv预印本", "verification": "原始身份/版本与相关正文核验；未复现、未核尽代码/原始数据/引用", "verified_at": "2026-10-17T07:09:45.339671+08:00", "evidence": "7模型×50任务×3=1050；最高均分6.29/10模型RC4为49/150约32.7%；高分229中83未RC4；闭合判断专家一致75.9%", "read_first": "3.3节、图4/5、case30", "id": "arxiv:2610.21767", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-10-15", "limitations": "任务/评分规范依赖专家；未全量人工验证；无接口或物理实验比较" }, { "title": "Agentic AI for operating scientific instruments for nanoscale characterization", "team": "EPFL；Zahra Ayar、Georg E. Fantner等", "doi": "10.48550/arXiv.2608.26198", "arxiv": "2608.26198", "url": "https://arxiv.org/abs/2608.26198", "fulltext_url": "https://arxiv.org/pdf/2608.26198", "type": "arXiv预印本", "verification": "原始身份/版本与相关正文核验；未复现、未核尽代码/原始数据/引用", "verified_at": "2026-10-17T07:09:45.339671+08:00", "evidence": "五匹配对同起点/参数/边界，轮次p=.063、时间.625；交互命令未观察错误执行；正文150请求与附表225汇总分母未厘清，不引用比例", "read_first": "2.3–2.4节、图4、Methods5.5", "id": "arxiv:2608.26198", "first_published": "2026-08-25", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-08-25", "limitations": "小样本不显著不等于等效；模型、接口、交互混杂；已有驱动成本未计；KPFM自身PID固定；请求总数口径未厘清。" }, { "title": "A Model Context Protocol Server for Quantum Execution in Hybrid Quantum-HPC Environments", "team": "早稻田大学、AIST AIRC/G-QuAT；Masaki Shiraishi、Ikko Hamamura等", "doi": "10.48550/arXiv.2604.08318", "arxiv": "2604.08318", "url": "https://arxiv.org/abs/2604.08318", "fulltext_url": "https://arxiv.org/html/2604.08318v1", "type": "QC4C3/QCNC 2026研讨会论文及arXiv版本", "verification": "原始身份/版本与相关正文核验；未复现、未核尽代码/原始数据/引用", "verified_at": "2026-10-17T07:09:45.339671+08:00", "evidence": "GHZ本地2000采样/远端模拟100采样；QAOA给定参数演示；无比较性能", "read_first": "II架构、III-B/C", "id": "arxiv:2604.08318", "first_published": "2026-04-09", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "version_date": "2026-04-09", "limitations": "原文/终端本地计数不一致不采信精确值；walltime是申请上限非实测排队/运行；费用/资格/适配人时未知" }, { "title": "OS-Helix: Unifying GUI Interaction and Code Execution in Native Computer-Using Agents", "team": "上海人工智能实验室等；Bowen Yang等", "doi": "10.48550/arXiv.2610.22234", "arxiv": "2610.22234", "url": "https://arxiv.org/abs/2610.22234", "fulltext_url": "https://arxiv.org/html/2610.22234v1", "type": "arXiv预印本", "verification": "身份/日期与相关原始正文已核；未复现、未核尽代码和引用", "verified_at": "2026-10-18T07:07:39.548675+08:00", "evidence": "361题/50步；匹配27B Hybrid46.50、SFT62.75、RL66.44%；固定10K轨迹混合比GUI高3.05pp", "read_first": "4.1、表1、表5", "id": "arxiv:2610.22234", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "工作流子类RL未再提高；模型/训练成本及新仪器接入人时未测", "version_date": "2026-10-15" }, { "title": "ISW-Agent: Integrating LLM-Driven Dialogue with n8n Workflow for Internal Solitary Wave Remote Sensing Image Processing", "team": "自然资源部第一海洋研究所；Hao Zhang等", "doi": "10.48550/arXiv.2610.21502", "arxiv": "2610.21502", "url": "https://arxiv.org/abs/2610.21502", "fulltext_url": "https://arxiv.org/html/2610.21502v1", "type": "arXiv预印本", "verification": "身份/日期与相关原始正文已核；未复现、未核尽代码和引用", "verified_at": "2026-10-18T07:07:39.548675+08:00", "evidence": "852条件×4=3408运行；拒234结构错仍漏61内容错；50景39完成/11空提取，48.04s仅完成者处理链", "read_first": "2.2、3.5图7", "id": "arxiv:2610.21502", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "内容身份/完整性未被结构校验覆盖；非预测精度证据；时间不含LLM/取图", "version_date": "2026-10-15" }, { "title": "Holding the Light: Teaching an AI to Lock and Tune our Quantum Computer’s Lasers", "team": "QuEra激光团队；Claude/MHS合作", "url": "https://www.quera.com/blog-posts/holding-the-light-teaching-an-ai-to-lock-and-tune-our-quantum-computers-lasers", "fulltext_url": "https://www.quera.com/blog-posts/holding-the-light-teaching-an-ai-to-lock-and-tune-our-quantum-computers-lasers", "type": "公司原始实质技术文章（非论文）", "verification": "身份/日期与相关原始正文已核；未复现、未核尽代码和引用", "verified_at": "2026-10-18T07:07:39.548675+08:00", "evidence": "公司自报7类×100=700，695恢复；43自然跳模；约19h零失锁对先前1.6/h", "read_first": "Building…与Tuning…两节", "id": "url:https://www.quera.com/blog-posts/holding-the-light-teaching-an-ai-to-lock-and-tune-our-quantum-computers-lasers", "first_published": "2026-08-27", "verified_version": null, "appearances": [ "digests/scientific-computer-use/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "公司技术案例非论文；无独立完整数据审计；第二波长非跨仪器；接入总人时未知", "version_date": "2026-08-27" }, { "title": "SciΠvot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents", "team": "马里兰大学、Gradient Network等；Youran Sun等", "doi": "10.48550/arXiv.2610.21082", "arxiv": "2610.21082", "url": "https://arxiv.org/abs/2610.21082", "fulltext_url": "https://arxiv.org/html/2610.21082v1", "type": "arXiv预印本", "verification": "身份/日期与相关原始正文已核；未复现、未核尽代码和引用", "verified_at": "2026-10-18T07:07:39.548675+08:00", "evidence": "26项目248例，交换选项顺序；五模型medium均值有记录39.2%/无记录55.8%", "read_first": "2.3、4.4、限制", "id": "arxiv:2610.21082", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "挑战样本经模型分歧筛选；客户端/检索同时变化；不是通用决策准确率", "version_date": "2026-10-15" }, { "title": "Reconstructing Implicit Scientific Knowledge: Evaluating LLM Agents through End-to-End Reproduction of Astronomy", "team": "之江实验室科学数据枢纽研究中心；Yuehui Wang等", "doi": "10.48550/arXiv.2609.35900", "arxiv": "2609.35900", "url": "https://arxiv.org/abs/2609.35900", "fulltext_url": "https://arxiv.org/pdf/2609.35900", "type": "arXiv预印本", "verification": "身份/日期与相关原始正文已核；未复现、未核尽代码和引用", "verified_at": "2026-10-18T07:07:39.548675+08:00", "evidence": "3×2×2共12预定路径，尺度2.16–3.53kpc；13篇审查中11篇路径不唯一（作者报告）", "read_first": "PDF4–5页和图7/8", "id": "arxiv:2609.35900", "first_published": "2026-09-28", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "3.1/3.3仍有模型/次数/包链接占位；不能当可独立复核性能基线或总体复现率", "version_date": "2026-09-28" }, { "title": "Jianmu-Q: Bridging Quantum Field Theory and Quantum Hardware with AI Agents", "team": "华南师大、RIKEN、KEK等；Xiaolin Wang等", "doi": "10.48550/arXiv.2610.21942", "arxiv": "2610.21942", "url": "https://arxiv.org/abs/2610.21942", "fulltext_url": "https://arxiv.org/html/2610.21942v1", "type": "arXiv预印本", "verification": "身份/日期及相关原始正文已核；未复现、未核尽代码及引用", "verified_at": "2026-10-19T07:08:24.441183+08:00", "evidence": "Schwinger相关量最大误差0.594降至0.00806；Z2真机8.1483±0.1350 vs同线路无噪7.7502约2.95统计标准误", "read_first": "第三节、图1、Schwinger案例末段", "version_date": "2026-10-15", "id": "arxiv:2610.21942", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "有限格点数值一致性不等于连续极限/真机定量准确；未测跨仪器接入人时" }, { "title": "Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery", "team": "Northwestern University；Yibo Wen、Han Liu", "doi": "10.48550/arXiv.2610.18449", "arxiv": "2610.18449", "url": "https://arxiv.org/abs/2610.18449", "fulltext_url": "https://arxiv.org/html/2610.18449v1", "type": "arXiv预印本", "verification": "身份/日期及相关原始正文已核；未复现、未核尽代码及引用", "verified_at": "2026-10-19T07:08:24.441183+08:00", "evidence": "8配置×6任务×3次=144，31双门成功；最佳12/18；矫顽力3/24；113次失败中25次只缺发现/30次只缺学习/58次两项都缺", "read_first": "2.4/2.6与表5，附录矫顽力任务", "version_date": "2026-10-13", "id": "arxiv:2610.18449", "first_published": "2026-10-13", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "公开数据潜在记忆、任务按捷径审计校准、未含现场漂移和样品失败" }, { "title": "Probe-and-Resume: When and How to Evaluate Agent Execution", "team": "Columbia University；Jiakai Xu等", "doi": "10.48550/arXiv.2610.21088", "arxiv": "2610.21088", "url": "https://arxiv.org/abs/2610.21088", "fulltext_url": "https://arxiv.org/html/2610.21088v1", "type": "arXiv预印本", "verification": "身份/日期及相关原始正文已核；未复现、未核尽代码及引用", "verified_at": "2026-10-19T07:08:24.441183+08:00", "evidence": "89任务各5次；首次完成63.4%至72.4%；30任务提高1退步；MCTS64%比较臂排除18不可评分试验", "read_first": "第5节与表3、初始状态副本消融", "version_date": "2026-10-15", "id": "arxiv:2610.21088", "first_published": "2026-10-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "物理状态不能分叉；检查仍会错拒；其他CLI成本比较含框架差异" }, { "title": "ELISA: A Bayesian X/γ-ray Spectral Analysis Toolkit", "team": "中科院高能物理研究所等；Wang-Chen Xue等", "doi": "10.48550/arXiv.2610.14587", "arxiv": "2610.14587", "url": "https://arxiv.org/abs/2610.14587", "fulltext_url": "https://arxiv.org/html/2610.14587v1", "type": "arXiv预印本", "verification": "身份/日期及相关原始正文已核；未复现、未核尽代码及引用", "verified_at": "2026-10-19T07:08:24.441183+08:00", "evidence": "131分组拟合能道；同案例100随机重复，12–16s vs80–110s，ESS约1e4 vs不超1e3；预测准则差约1.2±1.4", "read_first": "第五节诊断与第六节模型比较", "version_date": "2026-10-10", "id": "arxiv:2610.14587", "first_published": "2026-10-10", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "正文匹配设置概述与列出的链/步数不完全一致，算法不同；不可隔离归因或普遍百倍" }, { "title": "Easy remote observations using web interfaces: controlling an Italian telescope from Japan, and more", "team": "INAF、日本国立天文台等；Davide Ricci等", "doi": "10.48550/arXiv.2501.07997", "arxiv": "2501.07997", "url": "https://arxiv.org/abs/2501.07997", "fulltext_url": "https://arxiv.org/html/2501.07997v1", "type": "会议技术论文预印本", "verification": "身份/日期及相关原始正文已核；未复现、未核尽代码及引用", "verified_at": "2026-10-19T07:08:24.441183+08:00", "evidence": "单次远程案例，约200星粗测光，无受控性能对照", "read_first": "第二节控制层次、第三节学生案例", "version_date": "2025-01-14", "id": "arxiv:2501.07997", "first_published": "2025-01-14", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "旧能力证据非当前开放服务；HTML时延/曝光单位丢失不引用；接入人时/资格/价格未知" }, { "arxiv": "2610.23012", "title": "Adaptation of Computer Use Agents to New Software with User Manuals", "version_date": "2026-10-16", "team": "Chulhoon Yoon等；Korea University、Hanyang University", "evidence": "五应用OSWorld，50步、5次运行；UI-TARS36.83→47.31%、EvoCUA47.54→53.67%；Slack30组任务由VLM判定", "read_first": "§3.2–3.3、表3", "doi": "10.48550/arXiv.2610.23012", "url": "https://arxiv.org/abs/2610.23012", "fulltext_url": "https://arxiv.org/html/2610.23012v1", "type": "arXiv预印本", "verification": "原始身份/版本页与相关全文已核；未复现", "verified_at": "2026-10-20T07:06:04.726542+08:00", "id": "arxiv:2610.23012", "first_published": "2026-10-16", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "不是同一研究问题跨应用；应用训练覆盖未知，适配成本未报告；无API公平对照" }, { "arxiv": "2605.00868", "title": "Autonomous Reliability Qualification of Ga2O3-based diode sensors via Safe Active Learning", "version_date": "2026-08-07", "team": "Davi Febba等；National Laboratory of the Rockies", "evidence": "350–550°C/0–800ppm；阶段1固定阈值10共100轮1异常不安全；阶段2共100轮24低于固定阈值、9低于当轮阈值", "read_first": "§4、§6、补充表S1与图S7", "doi": "10.48550/arXiv.2605.00868", "url": "https://arxiv.org/abs/2605.00868", "fulltext_url": "https://arxiv.org/html/2605.00868v3", "type": "arXiv预印本", "verification": "原始身份/版本页与相关全文已核；未复现", "verified_at": "2026-10-20T07:06:04.726542+08:00", "id": "arxiv:2605.00868", "first_published": "2026-04-22", "verified_version": "v3", "appearances": [ "digests/scientific-computer-use/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "单设备示范非零风险保证；离线模型设计用过辅助验证数据；无接入人时或通用GUI对照" }, { "arxiv": "2610.20576", "title": "One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness", "version_date": "2026-10-14", "team": "Artin Tajdini、Manpreet Kaur、Peng Qi；University of Washington、Uniphore", "evidence": "3920训练、2501 BFCLv4测试、13留出扰动；3B多接口均值54.2→56.1，4B79.7→80.7；多轮常规ToolRL绝对奖励更高", "read_first": "表1、表8", "doi": "10.48550/arXiv.2610.20576", "url": "https://arxiv.org/abs/2610.20576", "fulltext_url": "https://arxiv.org/html/2610.20576v1", "type": "arXiv预印本", "verification": "原始身份/版本页与相关全文已核；未复现", "verified_at": "2026-10-20T07:06:04.726542+08:00", "id": "arxiv:2610.20576", "first_published": "2026-10-14", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "不约束参数概率一致性；部分提升在置信区间内；稳健性差距小不等于绝对成功更高" }, { "arxiv": "2607.13561", "title": "Design and Implementation of a Microservice-Architecture Master Control System for AIMS", "version_date": "2026-07-15", "team": "Li-Yue Tong等；中科院国家天文台、云南天文台、西安光机所、上海技物所等", "evidence": "表2逐功能部署边界；表3某扫描1–2→7–8组/日、7–8→2人，未给对照天数与错误分母；局域网通信<100ms非周转", "read_first": "表2、§5、§3.3", "doi": "10.48550/arXiv.2607.13561", "url": "https://arxiv.org/abs/2607.13561", "fulltext_url": "https://arxiv.org/pdf/2607.13561v1", "type": "arXiv预印本", "verification": "原始身份/版本页与相关全文已核；未复现", "verified_at": "2026-10-20T07:06:04.726542+08:00", "id": "arxiv:2607.13561", "first_published": "2026-07-15", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "未来规划不可算已部署；运维前后描述未控天气；无新设备接入工时" }, { "arxiv": "2610.22379", "title": "When is a large language model a good calibration agent? Physics-labeled prompting for budgeted engineering simulator calibration", "version_date": "2026-10-01", "team": "Ruo-Qian Wang；Rutgers University", "evidence": "共享4初始点，第5次提议23/25改善vs消融8/30；B5中位误差比21.9，B15无显著分离；同策略批中位数B5跨度9.8倍", "read_first": "§3.2、表4–5、§5.5", "doi": "10.48550/arXiv.2610.22379", "url": "https://arxiv.org/abs/2610.22379", "fulltext_url": "https://arxiv.org/html/2610.22379v1", "type": "arXiv预印本", "verification": "原始身份/版本页与相关全文已核；未复现", "verified_at": "2026-10-20T07:06:04.726542+08:00", "id": "arxiv:2610.22379", "first_published": "2026-10-01", "verified_version": "v1", "appearances": [ "digests/scientific-computer-use/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "摘要9次混合达标发生在15次交接前不能归因交接；后期差异受可辨识性限制；单LLM配置" }, { "title": "CUA-Recovery: Benchmarking Error Recovery in Long-Horizon Computer-Use Agents", "team": "Xu Hu等；UT Dallas、Manchester、Rice、UT Austin、UC Davis等", "doi": "10.48550/arXiv.2610.27645", "arxiv": "2610.27645", "url": "https://arxiv.org/abs/2610.27645", "fulltext_url": "https://arxiv.org/html/2610.27645v1", "version_date": "2026-10-20", "type": "arXiv预印本", "verification": "上游已核原始身份、版本及正文相关方法与证据；本汇编未独立复现", "verified_at": "2026-10-22T07:02:11.028487+08:00", "evidence": "100任务2250错误状态；377任务8577训练恢复轨迹；67.5%多应用；最高EAR44%、Pass@3 25.5%；等token训练对照EAR23.6% vs35.2%", "read_first": "§3–4及恢复训练对照", "id": "arxiv:2610.27645", "first_published": "2026-10-20", "verified_version": "v1", "limitations": "错误意识含模型判分；不同深度失败样本不同；数字桌面结果未验证科学仪器。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Lost in Recovery: Demystifying Recovery Behavior of Computer-Use Agents via Goal-Oriented Action Replay", "team": "Byeonghun Bang、Sangwoo Park等；KAIST、ETRI、DeepAuto.ai", "doi": "10.48550/arXiv.2610.27954", "arxiv": "2610.27954", "url": "https://arxiv.org/abs/2610.27954", "fulltext_url": "https://arxiv.org/html/2610.27954v1", "version_date": "2026-10-20", "type": "arXiv预印本", "verification": "上游已核原始身份、版本及正文相关方法与证据；本汇编未独立复现", "verified_at": "2026-10-22T07:02:11.028487+08:00", "evidence": "Qwen3.5-9B三次运行39.38±1.17→44.26±1.90%；前向50步另加每次10恢复步与检测调用", "read_first": "§5.1–5.2、表2、附录B.2", "id": "arxiv:2610.27954", "first_published": "2026-10-20", "verified_version": "v1", "limitations": "稳定回放筛选限制代表性；恢复和检测另耗预算；无物理可逆性验证。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "CUA-Lite: An Open Platform for Computer Use Agents", "team": "Zhanhui Zhou、Weichen Zhang等；Dawn Song参与，Berkeley联系邮箱", "doi": "10.48550/arXiv.2610.26071", "arxiv": "2610.26071", "url": "https://arxiv.org/abs/2610.26071", "fulltext_url": "https://arxiv.org/html/2610.26071v1", "version_date": "2026-10-19", "type": "arXiv预印本", "verification": "上游已核原始身份、版本及正文相关方法与证据；本汇编未独立复现", "verified_at": "2026-10-22T07:02:11.028487+08:00", "evidence": "332共用OSWorld任务13模型，平均分差2.7pp，最大5.0pp，Spearman0.98；内存1/6、启动时间3/4", "read_first": "§2.1–2.3与附录D", "id": "arxiv:2610.26071", "first_published": "2026-10-19", "verified_version": "v1", "limitations": "共同可运行任务子集；资源节省不等于科研集成成本；未进行科学任务等预算接口对照。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "The Scheduler Is a Hidden Retriever: How Asynchrony Shapes Evidence Acquisition in Search Agents", "team": "Ruiyang Ren、Yuhao Wang、Yinan Zhang、Chunyan Miao；NTU", "doi": "10.48550/arXiv.2610.26310", "arxiv": "2610.26310", "url": "https://arxiv.org/abs/2610.26310", "fulltext_url": "https://arxiv.org/html/2610.26310v1", "version_date": "2026-10-19", "type": "arXiv预印本", "verification": "上游已核原始身份、版本及正文相关方法与证据；本汇编未独立复现", "verified_at": "2026-10-22T07:02:11.028487+08:00", "evidence": "两模型系列三基准；获取分歧最高超重复24pp；CEA降至重复约1pp内；准确率增益最高12.6pp", "read_first": "引言、§3.2、§4–5", "id": "arxiv:2610.26310", "first_published": "2026-10-19", "verified_version": "v1", "limitations": "固定计划及稳定响应假设；最高增益依配置而异；未验证仪器环境。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "LatticeMind: A Validation-Gated Agentic Framework for Density Functional Theory Workflows", "team": "Nathaniel M. Vegh等；Nanoacademic、McGill", "doi": "10.48550/arXiv.2610.26869", "arxiv": "2610.26869", "url": "https://arxiv.org/abs/2610.26869", "fulltext_url": "https://arxiv.org/html/2610.26869v1", "version_date": "2026-07-19", "type": "arXiv预印本", "verification": "上游已核原始身份、版本及正文相关方法与证据；本汇编未独立复现", "verified_at": "2026-10-22T07:02:11.028487+08:00", "evidence": "44案例×3后端132次，119首计划13一次修复；23案例消融3重复，常规约71% vs100%；VASP19/19", "read_first": "契约/校验、消融与VASP迁移段", "id": "arxiv:2610.26869", "first_published": "2026-07-19", "verified_version": "v1", "limitations": "冻结案例与专用校验规则；VASP基线验收口径不同；日期元数据冲突未解决。", "appearances": [ "digests/scientific-computer-use/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" } ]