[ { "title": "Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms", "authors": "Xinjie Shen et al.", "doi": "10.48550/arxiv.2609.27321", "arxiv": "2609.27321", "url": "https://arxiv.org/abs/2609.27321", "fulltext_url": "https://arxiv.org/html/2609.27321v1", "first_published": "2026-09-23", "verified_version": "v1", "version_date": "2026-09-23", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "read_first": "Appendix B; Section 5.2 Table 2; Appendix K", "id": "arxiv:2609.27321", "appearances": [ "digests/agent-environments/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments", "authors": "Jie Wu et al.", "doi": "10.48550/arxiv.2609.04148", "arxiv": "2609.04148", "url": "https://arxiv.org/abs/2609.04148", "fulltext_url": "https://arxiv.org/html/2609.04148v1", "first_published": "2026-09-03", "verified_version": "v1", "version_date": "2026-09-03", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "read_first": "Sections 3.1, 3.3, 4.2; Table 2", "id": "arxiv:2609.04148", "appearances": [ "digests/agent-environments/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ReVeal: A Reconstruction-Aware Real-to-Sim Framework for VLA Policy Evaluation", "authors": "Xinyi Wang et al.", "doi": "10.48550/arxiv.2609.23910", "arxiv": "2609.23910", "url": "https://arxiv.org/abs/2609.23910", "fulltext_url": "https://arxiv.org/html/2609.23910v1", "first_published": "2026-09-20", "verified_version": "v1", "version_date": "2026-09-20", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "read_first": "Sections V-A/B/C; Tables II/III; Conclusion", "id": "arxiv:2609.23910", "appearances": [ "digests/agent-environments/2026/2026-09-28.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Skill-based Agentic Evaluation for Real-time Data Science Tasks", "authors": "Aniruddha Tamhane; Raghavendra Addanki; Ayushi Aggarwal; Aditya Bansal; Rui Wang; Charles Menguy; Swati Jain", "team": "Adobe", "doi": "10.48550/arxiv.2609.16487", "arxiv": "2609.16487", "url": "https://arxiv.org/abs/2609.16487", "fulltext_url": "https://arxiv.org/pdf/2609.16487", "first_published": "2026-09-15T01:20:37Z", "verified_version": "v1", "version_date": "2026-09-15", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "55 cases, 53 comparable pass/fail cases; reported MCC .331 to .427 against natural-language ground truth", "limitations": "Synthetic database; expert reference; same model family for agent and judge; no cross-harness validation", "read_first": "Section 5 Table 3; Limitations; ground-truth-as-code method", "id": "arxiv:2609.16487", "appearances": [ "digests/agent-environments/2026/2026-09-29.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "OSWorld-Pro: Process-based Evaluation for Computer Use Agents", "authors": "Zhilin Wang; Shaokun Zhang; Yifan Zhang; Hao Zhang; Jin Xu; Binfeng Xu; Jian Hu; Yunheng Zou; Karan Sapra; Andrew Tao; Jan Kautz; Yi Dong", "team": "NVIDIA", "doi": "10.48550/arxiv.2609.24890", "arxiv": "2609.24890", "url": "https://arxiv.org/abs/2609.24890", "fulltext_url": "https://arxiv.org/html/2609.24890v1", "first_published": "2026-09-21T16:55:24Z", "verified_version": "v1", "version_date": "2026-09-21", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "305 tasks, 2814 subgoals, 67264 human step labels; local feasibility judging remains imperfect", "limitations": "GUI process benchmark; not service reconstruction or unseen API-combination evaluation", "read_first": "Section 2; Section 4.2 Table 1; Appendix D.6", "id": "arxiv:2609.24890", "appearances": [ "digests/agent-environments/2026/2026-09-29.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "How to Evaluate AI Agents From Tool Calls to Task Completion", "authors": "Sophia Abbassi; Chris Alexiuk; Davide Onofrio; Gomathy Venkata Krishnan", "team": "NVIDIA", "url": "https://developer.nvidia.com/blog/how-to-evaluate-ai-agents-from-tool-calls-to-task-completion/", "fulltext_url": "https://developer.nvidia.com/blog/how-to-evaluate-ai-agents-from-tool-calls-to-task-completion/", "first_published": "2026-09-21", "verified_version": "page accessed 2026-09-29; no version history supplied", "type": "official_technical_article", "verification": "original_article_body", "evidence": "Process/outcome decomposition and public SWE-bench trace illustration", "limitations": "Methodological guidance, not controlled simulator-fidelity evidence; marketing claims not adopted", "read_first": "From scoring calls to scoring the environment; How to read an evaluation; Table 1", "id": "url:https://developer.nvidia.com/blog/how-to-evaluate-ai-agents-from-tool-calls-to-task-completion/", "appearances": [ "digests/agent-environments/2026/2026-09-29.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Breaking the Environment Wall: A Unified Framework for Preparing and Evolving Agent-Native Environments", "authors": "Yukai Wu; Yuanjing Yang; Le Zhou; Shaokun Han; Haoyu Wang; Zirui Tang; Xuzhou Zhu; Weihuang Zheng; Maxm Pan; Xuanhe Zhou; Fan Wu", "team": "Shanghai Jiao Tong University; Theseus Labs; Tencent Hunyuan", "doi": "10.48550/arxiv.2609.29773", "arxiv": "2609.29773", "url": "https://arxiv.org/abs/2609.29773", "fulltext_url": "https://arxiv.org/html/2609.29773v2", "first_published": "2026-09-24T13:17:57Z", "verified_version": "v2", "version_date": "2026-09-27T07:45:46Z", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "30 tasks / 9 models: mean rubric pass rate 57.6% to 72.7%. Evolution: 32 of 55 retained paired tasks meet >=20pp decrease in >=3 models; nominal 3 trials per arm, completed-trial denominators.", "limitations": "Jointly updates state, reference outcomes and evaluator; not fixed-target black-box fidelity evidence. Task-guided mode has privileged information. Paired-task selection and small trial counts limit interpretation.", "read_first": "Section 4.2 → Appendix C/C.1 → Section 5.4", "id": "arxiv:2609.29773", "appearances": [ "digests/agent-environments/2026/2026-09-30.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation", "authors": "Renxi Wang; Mingshan Hee; Fajri Koto; Timothy Baldwin; Haonan Li", "team": "MBZUAI", "doi": "10.48550/arxiv.2609.37539", "arxiv": "2609.37539", "url": "https://arxiv.org/abs/2609.37539", "fulltext_url": "https://arxiv.org/html/2609.37539v1", "first_published": "2026-09-29T13:21:38Z", "verified_version": "v1", "version_date": "2026-09-29T13:21:38Z", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "6,772 tasks from 3,494 skills; builder/reviewer; verifier must fail on initial state and pass reference solution; quality review of leakage and verifier strictness.", "limitations": "Offline/CPU/noninteractive selection; jointly constructs task, reference solution and verifier; executable success does not establish fidelity to a fixed hidden external service.", "read_first": "§3.2.3 → §3.2.2 → Appendix C", "id": "arxiv:2609.37539", "appearances": [ "digests/agent-environments/2026/2026-10-01.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "SWE-Serve: Benchmarking Agentic Engineering For Production Inference Serving", "authors": "Jennifer Williams; Dave Farris; Jeff Farris; Jiantao Jiao", "team": "NVIDIA official benchmark; affiliations not separately used", "doi": "10.48550/arxiv.2609.26777", "arxiv": "2609.26777", "url": "https://arxiv.org/abs/2609.26777", "fulltext_url": "https://arxiv.org/html/2609.26777v1", "first_published": "2026-09-22T17:54:59Z", "verified_version": "v1", "version_date": "2026-09-22T17:54:59Z", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "53 SGLang tasks. Fixed 627 patches on 19 E2E tasks: full verifier 45.9%, E2E omitted 69.4%. Matched control: 12 tasks/396 patches/38 tests per arm, 10,000 masks; transitions 16.1 vs 8.0; average pass-rate difference 2.1pp, 95% mask interval -0.5 to 4.3pp.", "limitations": "Feature implementation with source access, not black-box simulation. Matched-mask interval includes zero; masks are not independent task replications; GPU/serving-stack scope.", "read_first": "§5 E2E gap → Appendix A.4.1 matched control → Appendix A.1.3", "id": "arxiv:2609.26777", "appearances": [ "digests/agent-environments/2026/2026-10-01.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Locating Hidden Failures Makes Long-Horizon Agents More Reliable", "authors": "Salman Rahman; Yubin Kim; Mihir Parmar; A. Ali Heydari; Genglin Liu; Simon A. Lee; Weizhi Zhang; Arian Hosseini; Ahmed A. Metwally; Yuzhe Yang; Baharan Mirzasoleiman; Xin Liu; Pavel Izmailov; Saadia Gabriel; Mark Malhotra; Shwetak Patel; Daniel McDuff; Hamid Palangi", "team": "UCLA; NYU; Google Research/DeepMind (original HTML affiliations)", "doi": "10.48550/arxiv.2609.17930", "arxiv": "2609.17930", "url": "https://arxiv.org/abs/2609.17930", "fulltext_url": "https://arxiv.org/html/2609.17930v1", "first_published": "2026-09-15T23:40:42Z", "verified_version": "v1", "version_date": "2026-09-15T23:40:42Z", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "2,518 trajectories; Table 1 maximum exact first-error location 26.8% SWE-bench, 32.34% TerminalBench, 63.16% BixBench; human agreement kappa 0.73 for location.", "limitations": "Draft placeholders and pending recovery/cross-domain analyses remain; abstract under-one-third statement does not cover BixBench Table 1. Human first-error label may be diffuse/multicausal; not mechanical simulator state divergence.", "read_first": "Table 1 / §3.2 → §4 limitations → §2 annotation", "id": "arxiv:2609.17930", "appearances": [ "digests/agent-environments/2026/2026-10-01.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "The Agent Said It Was Done. The Database Disagreed.", "authors": "Tuhin Kundu; Microsoft and Hugging Face joint article", "doi": "10.48550/arxiv.2608.19741", "arxiv": "2608.19741", "url": "https://huggingface.co/blog/microsoft/thinkingbox", "first_published": "2026-10-03", "verified_version": "v4", "version_date": "2026-10-01T16:14:54Z", "type": "substantial_official_article_with_preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "507 synthetic tasks, 20 independent agent trials. 67.24% of 79,853 check-failing attempts satisfy weak clean-termination/mutating-tool/no-terminal-error proxy.", "limitations": "Synthetic environment, weak proxy comparator; not real-service fidelity or independent simulator-build repetition.", "read_first": "Article How it works → paper D.2 Table 16 → B.4", "id": "arxiv:2608.19741", "appearances": [ "digests/agent-environments/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Network World Models as Environments for Algorithm Design on Complex Systems", "authors": "Rishab Alagharu; Hongji Pu; Zeeshan Memon; Xinyuan Song; Yuntong Hu; Liang Zhao", "team": "Emory University; UIUC", "doi": "10.48550/arxiv.2610.01048", "arxiv": "2610.01048", "url": "https://arxiv.org/abs/2610.01048", "fulltext_url": "https://arxiv.org/html/2610.01048v1", "first_published": "2026-10-01T04:44:00Z", "verified_version": "v1", "version_date": "2026-10-01T04:44:00Z", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "LT deltaF1 .9926 but horizon count bias -6.0%; WS preference preservation86.0%, selection regret1.18% versus random10.07%.", "limitations": "Known mechanisms/exact interventions; trusted reference is Monte Carlo model, code upon acceptance; rollout speed excludes build cost.", "read_first": "4.4 Table2/Table3/cost → 3 assumptions", "id": "arxiv:2610.01048", "appearances": [ "digests/agent-environments/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "When Better Gets Worse: Improvement Fidelity for Self-Improving Agents in Adaptive Worlds", "authors": "Ke Wang; Zijie Zhao; Zhiyi Yuan; Changlun Li", "doi": "10.48550/arxiv.2609.32677", "arxiv": "2609.32677", "url": "https://arxiv.org/abs/2609.32677", "fulltext_url": "https://arxiv.org/html/2609.32677v1", "first_published": "2026-09-26T14:34:02Z", "verified_version": "v1", "version_date": "2026-09-26T14:34:02Z", "type": "preprint", "verification": "original_abstract_and_targeted_fulltext", "evidence": "51/90 heldout roots have disjoint proxy/deployment optimum sets. MeltingPot Table1 comparison interval includes zero.", "limitations": "Adaptive response differs from fixed service. Appendix C driving comparison changes complete selection/calibration rules; Table1 shares rules and varies allocation only. Earlier cohorts null.", "read_first": "3 → 6.4 Table1 → Appendix C / 7", "id": "arxiv:2609.32677", "appearances": [ "digests/agent-environments/2026/2026-10-04.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Agent-Diff: Benchmarking LLM Agents on Enterprise API Tasks via Code Execution with State-Diff-Based Evaluation", "authors": "Hubert M. Pysklo; Artem Zhuravel; Patrick D. Watson", "team": "Minerva University", "doi": "10.48550/arXiv.2602.11224", "arxiv": "2602.11224", "url": "https://arxiv.org/abs/2602.11224", "fulltext_url": "https://arxiv.org/html/2602.11224v3", "first_published": "2026-02-11", "verified_version": "v3", "version_date": "2026-04-28", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-05T07:10:56.875679+08:00", "evidence": "四类企业API副本、数据库状态差分契约；9模型224任务。附录G明确描述生产API与副本同请求结构对照，以及辅助agent对非法参数等边界探测。", "limitations": "验证覆盖指定请求集合，结构等价不等于全部值与跨序列行为等价；作者将通知等内部副作用视为评测无关，适用范围需检验。", "read_first": "§3服务建模与状态差分；附录G及I", "id": "arxiv:2602.11224", "appearances": [ "digests/agent-environments/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Simulate to Generalize: Scaling Stateful Supervision for API-calling Agents using LLM World Models", "authors": "Seanie Lee; Sanjoy Chowdhury; Chao Jiang; Cheng-Yu Hsieh; Ting-Yao Hu; Alexander Toshev; Oncel Tuzel; Raviteja Vemulapalli", "team": "Apple（原文通信邮箱及作者元数据）", "doi": "10.48550/arXiv.2607.16900", "arxiv": "2607.16900", "url": "https://arxiv.org/abs/2607.16900", "fulltext_url": "https://arxiv.org/html/2607.16900v3", "first_published": "2026-07-18", "verified_version": "v3", "version_date": "2026-09-22", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-05T07:10:56.875679+08:00", "evidence": "自动合成API、任务、依赖历史的模拟回复及轨迹过滤。作者用LLM judge检查1000条成功轨迹26716调用，总体96.0%通过；469个写后读调用88.7%通过。200条轨迹人工与过滤judge一致95%。", "limitations": "这些是规格/历史一致性与过滤一致性，不是真实服务配对保真。成功轨迹抽样和LLM judge不能证明失败操作语义；v3相对旧版变化未比较。", "read_first": "§3流程；§4.4；§4.6–4.7，先看write–read子集", "id": "arxiv:2607.16900", "appearances": [ "digests/agent-environments/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents", "authors": "Minki Kang et al.", "team": "NVIDIA研究论文入口（其页面日期异常，采用arXiv日期）", "doi": "10.48550/arXiv.2609.39982", "arxiv": "2609.39982", "url": "https://arxiv.org/abs/2609.39982", "fulltext_url": "https://arxiv.org/html/2609.39982v1", "first_published": "2026-09-30", "verified_version": "v1", "version_date": "2026-09-30", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-05T07:10:56.875679+08:00", "evidence": "固定生成器与harness，同历史采样多个动作，执行前以listwise/pointwise/pairwise选一个。TerminalBench-Lite的TMAX-9B在强验证器下Pass@1由50.00%至68.03%；弱自验证扩宽仅小幅收益。", "limitations": "没有候选动作gold标签；教师一致率不是动作真值。终端环境结果不能直接迁移到服务API；8候选全pairwise需28次验证调用，成本需同预算比较。", "read_first": "§2机制；§4强弱验证对照；§5与附录A局限", "id": "arxiv:2609.39982", "appearances": [ "digests/agent-environments/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale", "authors": "Yash Pandya et al.", "team": "Microsoft Research", "doi": "10.48550/arXiv.2607.28074", "arxiv": "2607.28074", "url": "https://arxiv.org/abs/2607.28074", "fulltext_url": "https://arxiv.org/html/2607.28074v1", "first_published": "2026-07-30", "verified_version": "v1", "version_date": "2026-07-30", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-05T07:10:56.875679+08:00", "evidence": "种子编译应用、修复前后端与数据库、任务在活数据上落地；数据库差分作为写任务验证依据，仍由GPT-4.1比较。§6.2两域消融：Allrecipes浅世界训练80→75，深世界80→85；Hugging Face深世界48→65。", "limitations": "两域、小任务集、训练数据规模相近但非逐轨迹匹配；数据库落地不意味着判分完全无LLM。自建世界与任务共同修复不等于复现指定真实服务。", "read_first": "§3.2验证；§4 factory/co-evolution；§6.2；附录C.2", "id": "arxiv:2607.28074", "appearances": [ "digests/agent-environments/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Verify Claims, Not Scores: Evidence-Based Verification of Modular Agents", "authors": "Ali Atiah Alzahrani", "team": "原文作者；不以机构名气准入", "doi": "10.48550/arXiv.2610.01348", "arxiv": "2610.01348", "url": "https://arxiv.org/abs/2610.01348", "fulltext_url": "https://arxiv.org/html/2610.01348v1", "first_published": "2026-10-01", "verified_version": "v1", "version_date": "2026-10-01", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-05T07:10:56.875679+08:00", "evidence": "在合成投资决策系统中用已知隐藏机制、oracle替换、共同随机路径和组件干预逐项检验主张；明确没有LLM。示例显示局部保真改进未检测到决策收益，部分交互结论仍未解决。", "read_first": "§3设计；§4.3组件替换；§5.3；§7局限", "id": "arxiv:2610.01348", "appearances": [ "digests/agent-environments/2026/2026-10-05.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "StableToolBench-MirrorAPI: Modeling Tool Environments as Mirrors of 7,000+ Real-World APIs", "authors": "Zhicheng Guo; Sijie Cheng; Yuchen Niu; Hao Wang; Sicheng Zhou; Wenbing Huang; Yang Liu", "team": "清华大学等；机构来自论文署名", "doi": "10.18653/v1/2025.findings-acl.273", "url": "https://aclanthology.org/2025.findings-acl.273/", "fulltext_url": "https://aclanthology.org/2025.findings-acl.273.pdf", "first_published": "2025-07", "verified_version": "ACL Findings 2025", "version_date": "2025-07", "type": "peer_reviewed_paper", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-06T17:22:32.421632+08:00", "evidence": "用真实API请求/响应训练工具模拟器，并在MirrorAPI-Bench及StableToolBench中比较文档/指令遵循、响应与真实环境相似性及下游评测表现。覆盖7,000+ API；作者报告比提示基线更接近真实API。", "read_first": "§3–4的构建与评测；PDF第9页（印刷页5255）Limitations", "id": "doi:10.18653/v1/2025.findings-acl.273", "appearances": [ "digests/agent-environments/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "MCPGen: Benchmarking LLMs on Executable MCP Workflow Development", "authors": "Yingxuan Yang; Jiaqi Liu; Lirui Guan; Jiaye Gao; Weiwen Liu; Weinan Zhang; Ying Wen", "team": "上海交通大学", "doi": "10.48550/arXiv.2609.23925", "arxiv": "2609.23925", "url": "https://arxiv.org/abs/2609.23925", "fulltext_url": "https://arxiv.org/html/2609.23925v1", "first_published": "2026-09-20", "verified_version": "v1", "version_date": "2026-09-20", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-06T17:22:32.421632+08:00", "evidence": "100个可执行MCP项目、16个领域、11个模型，分别测工作流重建、工具实现和扩展；固定依赖与确定性外部mock。Claude-Sonnet工具单测63.8%、项目整合41.0%，两者统计单位不同；最佳整合45.0%。", "limitations": "单轮产物构建，没有检索/执行反馈/修复预算；可执行项目与确定性mock不等于从真实服务黑盒重建。单测按工具目标、整合按项目，不能把差值当同一批样本的序列保真下降率。", "read_first": "§3.2–3.3；§4.2 Table 3；附录A", "id": "arxiv:2609.23925", "appearances": [ "digests/agent-environments/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Predicting LLM Safety Before Release by Simulating Deployment", "authors": "Marcus Williams; Hannah Sheahan; Cameron Raymond; Tomek Korbak; Deng Pan; Peilin Yang; Leon Maksin; Ningyi Xie; Phillip Guo; Ian Kivlichan; Micah Carroll", "team": "OpenAI", "doi": "10.48550/arXiv.2607.07184", "arxiv": "2607.07184", "url": "https://arxiv.org/abs/2607.07184", "fulltext_url": "https://arxiv.org/html/2607.07184v1", "first_published": "2026-07-08", "verified_version": "v1", "version_date": "2026-07-08", "type": "preprint_with_official_technical_article", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-06T17:22:32.421632+08:00", "evidence": "§3.3使用约12万条内部Codex轨迹；工具模拟器加入原轨迹、对应代码快照、历史调用响应库、只读网络连接器。判别器把模拟轨迹判为真实的胜率由11.6%升至49.5%，接近50%随机辨别点。", "limitations": "判别器难分真假不证明状态等价。原轨迹对新模型偏离路径帮助减弱；执行代码与访问只读连接器的权限/信息比纯黑盒不同。安全发生率预测是另一组证据，不当作工具状态保真结果。", "read_first": "§3.3 Tool simulator affordances与Figure 8；§4局限", "id": "arxiv:2607.07184", "appearances": [ "digests/agent-environments/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "WorldModelGym: a decision-based fidelity benchmark for world models", "authors": "Reka Labs（机构官方技术文章）", "team": "Reka Labs", "url": "https://reka.ai/labs/research/worldmodelgym", "fulltext_url": "https://reka.ai/labs/research/worldmodelgym", "first_published": "2026-07-02", "verified_version": "官方正文；7月28日更新标记", "version_date": "2026-07-28", "type": "substantive_official_technical_article", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-06T17:22:32.421632+08:00", "evidence": "固定观测与短历史，给5条候选开环动作序列（4条相似变体加随机基线），全部在参考环境预执行；按模拟累计奖励择优，计算真实选择遗憾，并辅以奖励预测误差。", "limitations": "物理/控制环境不能直接代表服务API；固定选项开环测试不覆盖长时闭环纠错。公式变量worst_return实际被正文定义为随机基线，并非保证最差选项；零分母/退化任务处理未核。零遗憾也可能是预测误差抵消。", "read_first": "The protocol: scoring decisions；normalized regret；Limitations", "id": "url:https://reka.ai/labs/research/worldmodelgym", "appearances": [ "digests/agent-environments/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents", "authors": "Yu Li; Zheng Zhang; Xin Liu; Shengtian Yang; Guangfeng Cai; Lei Feng", "team": "原文作者；不以机构名气准入", "doi": "10.48550/arXiv.2610.02330", "arxiv": "2610.02330", "url": "https://arxiv.org/abs/2610.02330", "fulltext_url": "https://arxiv.org/html/2610.02330v1", "first_published": "2026-10-01", "verified_version": "v1", "version_date": "2026-10-01", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-06T17:22:32.421632+08:00", "evidence": "CITA以日志上下文后续成功、贝叶斯工具转移图模拟和LLM语义比较估计候选动作相对价值，用于训练/推理；在三个工具任务数据集、两个小模型上报告收益。Tool F1与参考工具对照，任务成功率由GPT-4o裁判。", "limitations": "贝叶斯sandbox采样图上成功/可靠性先验，不等于真实服务副本；任务成功裁判读取截断观测而非机械最终状态；结果是agent方法收益，不能证明环境保真。", "read_first": "§3 comparative estimator；§4.1；附录G/H裁判协议", "id": "arxiv:2610.02330", "appearances": [ "digests/agent-environments/2026/2026-10-06.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation", "authors": "Quanyu Long; Xiao Chen; Jianda Chen; Haozhen Zhang; Qisheng Hu; Jianzhu Bao; Wenya Wang", "team": "作者名单已核；团队不作未核归属", "doi": "10.48550/arxiv.2610.06100", "arxiv": "2610.06100", "url": "https://arxiv.org/abs/2610.06100", "fulltext_url": "https://arxiv.org/html/2610.06100v1", "first_published": "2026-10-05", "verified_version": "v1", "version_date": "2026-10-05", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-07T07:06:57.343516+08:00", "evidence": "历史轨迹重建 worldbook；运行时维护持久状态，先提出状态效果，再由 harness 验证并提交。7类环境测单步；另2类文本环境测多轮。Table 2 的 ALFWorld 真实环境成功率93%，直接提示/Trace2Env在模拟内成功97%/88%，将模拟产生的动作重放到真实环境后仅3%/85%；CR=W2R/Real，分别0.032/0.914。", "limitations": "多轮结果只来自ALFWorld和SciWorld，不能外推全部9环境或生产API。单步主要用LLM裁判；W2R是任务成功率，不是逐步或最终状态严格等价率。轨迹、worldbook与运行时调用也须纳入信息和成本预算。", "read_first": "§5.3 Table 2 → §4.2–4.3 → 附录A/B.3/C.3", "id": "arxiv:2610.06100", "appearances": [ "digests/agent-environments/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "BackendForge: Benchmarking Agentic End-to-End Code Generation with Backend Services", "authors": "Yuzhe Guo; Mengzhou Wu; Yuan Cao; Jialei Wei; Dezhi Ran; Wei Yang; Tao Xie", "team": "作者名单已核；不据搜索目录推断团队", "doi": "10.48550/arxiv.2607.11042", "arxiv": "2607.11042", "url": "https://arxiv.org/abs/2607.11042", "fulltext_url": "https://arxiv.org/html/2607.11042v1", "first_published": "2026-07-13", "verified_version": "v1", "version_date": "2026-07-13", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-07T07:06:57.343516+08:00", "evidence": "56个spec/OpenAPI定义的服务任务，生成Docker服务后仅通过HTTP测试评分。测试agent提出新测试，review要求可追溯到契约，code agent仅修参考实现并全套回归。最终oracle含7,250基础项和640新增项；作者报告GPT-5.5全任务通过率由55.4%降至28.6%。", "limitations": "完整可见契约不同于黑盒行为发现。增强oracle同时覆盖多类问题，差值不能归因于某一种序列错误；测试与参考实现共同演化，仍需警惕共享遗漏。", "read_first": "§3.4 → §4.2 → §5.1–5.3", "id": "arxiv:2607.11042", "appearances": [ "digests/agent-environments/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Does Learning to Predict the World Help Agents Act? Auditing World-Model Post-Training", "authors": "Xinyu Che; Hang Yan; Yanchen Liu; Haochen Liu; Ruifeng Li; Anran Shi; Heng Wang; Jun Liu", "team": "西安交通大学、USC、UCAS、华东师范大学；按正文署名", "doi": "10.48550/arxiv.2609.33335", "arxiv": "2609.33335", "url": "https://arxiv.org/abs/2609.33335", "fulltext_url": "https://arxiv.org/html/2609.33335v1", "first_published": "2026-09-27", "verified_version": "v1", "version_date": "2026-09-27", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-07T07:06:57.343516+08:00", "evidence": "比较正确下一观测GT、分布内错配MIS和独立随机奖励COIN。ALFWorld有候选动作设置下GT/MIS预测准确率57.00%/27.25%，下降29.75个百分点；pass@1为37.30%/34.55%，pass@64为83.94%/83.21%。预测质量大变，64次尝试覆盖却接近。", "limitations": "审计的是agent后训练效应，不是固定agent的服务副本保真。正确目标仍提高单次成功；不能据此说准确性无用。多次采样与单次部署含义不同，预测评估也依赖裁判。", "read_first": "§2.2–2.3 → §3.1 Table 2 → §3.2–3.3", "id": "arxiv:2609.33335", "appearances": [ "digests/agent-environments/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Long-Horizon Textual World Modeling through Structured Reasoning", "authors": "Fangxin Wang; Xiang Gao; Yuguang Yao; Kaiwen Dong; Nikash Walia; Kamalika Das", "team": "Intuit AI Research、University of Illinois Chicago；正文署名", "doi": "10.48550/arxiv.2610.06637", "arxiv": "2610.06637", "url": "https://arxiv.org/abs/2610.06637", "fulltext_url": "https://arxiv.org/html/2610.06637v1", "first_published": "2026-10-05", "verified_version": "v1", "version_date": "2026-10-05", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-07T07:06:57.343516+08:00", "evidence": "把整段未来动作变成一次解码中的稀疏状态编辑，以确定性Apply重建中间状态；蒸馏、预测增益和中间奖励训练。ScienceWorld、Jericho、CEO-Bench对照递推与直接预测。§5.4固定动作注入虚假状态事实，平均SE-score损失递推11.1、蒸馏SR5.4、完整SR1.7；后两者说明训练也影响结果。", "limitations": "文本模拟环境预测，不是可交互真实服务副本；完整方法同时改变训练及表示，须看匹配消融。作者明确闭环规划收益仍未解决；状态可机械重建不保证真实服务状态正确。", "read_first": "§3.2与§4.1 → §5.3–5.5 → 附录E.2/E.3", "id": "arxiv:2610.06637", "appearances": [ "digests/agent-environments/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning", "authors": "Zhaoyang Wang; Canwen Xu; Boyi Liu; Yite Wang; Siwei Han; Zhewei Yao; Huaxiu Yao; Yuxiong He", "team": "Snowflake、UNC Chapel Hill；正文署名", "doi": "10.48550/arxiv.2602.10090", "arxiv": "2602.10090", "url": "https://arxiv.org/abs/2602.10090", "fulltext_url": "https://arxiv.org/html/2602.10090v3", "first_published": "2026-02-10", "verified_version": "v3", "version_date": "2026-05-22", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-07T07:06:57.343516+08:00", "evidence": "生成1,000个场景：任务→数据库schema/样例→代码工具环境，并读取数据库设计奖励。Table 6对100环境用GPT-5.1/Claude-4.5裁判：AWM含bug环境比例74%/83%，阻断任务比例14.0%/11.5%；运行可靠性不能从代码可执行直接推定。", "limitations": "上述bug比例是LLM裁判分析，非全部独立机械验证；作者另报告人工错误分类。训练跨基准泛化不证明固定真实服务行为等价。arXiv备注ICML2026接收；本次核的是v3作者版，未核正式出版关联及版本差异，不算更新。", "read_first": "§3生成与奖励 → §6.1 Table 6 → §5.2泛化", "id": "arxiv:2602.10090", "appearances": [ "digests/agent-environments/2026/2026-10-07.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "MCP-Persona: Benchmarking LLM Agents on Real-World Personal Applications via Environment Simulation", "authors": "Wenhao Wang; Peizhi Niu; Gongyi Zou; Xiyuan Yang; Jingxing Wang; Haoting Shi; Yaxin Du; Jingyi Chai; Xianghe Pang; Shuo Tang; Yanfeng Wang; Siheng Chen", "team": "作者名单已核；arXiv备注ICML 2026 Camera Ready，正式出版关联未核", "doi": "10.48550/arxiv.2606.02470", "arxiv": "2606.02470", "url": "https://arxiv.org/abs/2606.02470", "fulltext_url": "https://arxiv.org/html/2606.02470v1", "first_published": "2026-06-01", "verified_version": "v1", "version_date": "2026-06-01", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-08T14:59:37.853140+08:00", "evidence": "从真实成功调用及对抗失败调用生成有状态Python仿真器，拆分持久上下文与工具内核。§4.3以Lark 14工具的50条真实调用评估；作者报告ToolTraverse准确率94%，仅文档版本58%。", "limitations": "评估通过LLM裁判，且为每条调用逆向重建并注入前置上下文；单服务小样本不能证明未见序列或失败副作用等价，也未证明盲探测预算公平。", "read_first": "§3.1 → §4.3/Table 4 → 附录A.2/B.3", "id": "arxiv:2606.02470", "appearances": [ "digests/agent-environments/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "BaxBench: Can LLMs Generate Correct and Secure Backends?", "authors": "Mark Vero; Niels Mündler; Victor Chibotaru; Veselin Raychev; Maximilian Baader; Nikola Jovanović; Jingxuan He; Martin Vechev", "team": "ETH Zurich、LogicStar、UC Berkeley、INSAIT；正文署名", "doi": "10.48550/arxiv.2502.11844", "arxiv": "2502.11844", "url": "https://arxiv.org/abs/2502.11844", "fulltext_url": "https://arxiv.org/html/2502.11844v3", "first_published": "2025-02-17", "verified_version": "v3", "version_date": "2025-05-30", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-08T14:59:37.853140+08:00", "evidence": "由OpenAPI与文字规范生成可运行后端；28场景×14框架构成392任务。容器经HTTP接受框架无关功能测试，另有安全测试；参考正确实现经人工复核以检查测试。", "read_first": "§2.1–2.2 → 附录B/D.1", "id": "arxiv:2502.11844", "appearances": [ "digests/agent-environments/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "EnvScaler: Scaling Tool-Interactive Environments for LLM Agent via Programmatic Synthesis", "authors": "Xiaoshuai Song; Haofei Chang; Guanting Dong; Yutao Zhu; Ji-Rong Wen; Zhicheng Dou", "team": "中国人民大学高瓴人工智能学院/RUC-NLPIR；正文署名", "doi": "10.48550/arxiv.2601.05808", "arxiv": "2601.05808", "url": "https://arxiv.org/abs/2601.05808", "fulltext_url": "https://arxiv.org/html/2601.05808v2", "first_published": "2026-01-09", "verified_version": "v2", "version_date": "2026-04-17", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-08T14:59:37.853140+08:00", "evidence": "主题→规则/状态→Python环境；前端代理提正负调用，后端代理检查代码、返回及状态变化。每环境100轮、质量阈值0.85；共191环境约7K场景，任务由末态布尔函数评分。", "limitations": "质量门槛包含LLM判断，环境与验证器同源生成可能共犯错误。训练收益不能替代固定外部服务的配对忠实度；本次未核版本差异，不标实质更新。", "read_first": "§3.2–3.4 → §4.2 → §5.7", "id": "arxiv:2601.05808", "appearances": [ "digests/agent-environments/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Simulating Environments with Reasoning Models for Agent Training", "authors": "Yuetai Li; Huseyin A Inan; Xiang Yue; Wei-Ning Chen; Lukas Wutschitz; Janardhan Kulkarni; Radha Poovendran; Robert Sim; Saravan Rajmohan", "team": "Microsoft、University of Washington、Carnegie Mellon；正文署名", "doi": "10.48550/arxiv.2511.01824", "arxiv": "2511.01824", "url": "https://arxiv.org/abs/2511.01824", "fulltext_url": "https://arxiv.org/html/2511.01824v1", "first_published": "2025-11-03", "verified_version": "v1", "version_date": "2025-11-03", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-08T14:59:37.853140+08:00", "evidence": "Simia以工具schema、历史和少量种子轨迹提示推理模型生成反馈；合成轨迹过滤JSON与已知工具错误，另以模拟反馈和奖励支持训练。作者报告训练后下游任务收益。", "limitations": "语法/工具校验未保证语义正确；未接触实际testbed的模拟器可能合理地虚构实体。训练效用不能推出对指定服务的状态等价。", "read_first": "§2.1–2.3 → 附录D/E", "id": "arxiv:2511.01824", "appearances": [ "digests/agent-environments/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "GTM: Simulating the World of Tools for AI Agents", "authors": "Zhenzhen Ren; Xinpeng Zhang; Zhenxing Qian; Yan Gao; Yu Shi; Shuxin Zheng; Jiyan He", "team": "按arXiv列出的7名作者保存；页面人数描述不一致，机构未核不作归属", "doi": "10.48550/arxiv.2512.04535", "arxiv": "2512.04535", "url": "https://arxiv.org/abs/2512.04535", "fulltext_url": "https://arxiv.org/html/2512.04535v2", "first_published": "2025-12-04", "verified_version": "v2", "version_date": "2025-12-05", "type": "preprint_author_version", "classification": "supplement", "verification": "original_abstract_and_targeted_fulltext", "verified_at": "2026-10-08T14:59:37.853140+08:00", "evidence": "1.5B通用工具模型，以CARG合成20K工具/300领域数据，覆盖单轮、多轮和四类参数错误。Table 1借Qwen2.5-72B裁判报告三类综合平均89.4%。", "limitations": "89.4%是LLM判断的综合值，非真实服务执行等价率；多轮一致性不保证持久状态正确。速度对照的设备/接口条件不同，不能泛化成本优势。", "read_first": "§4.2 → §5.1/Table 1", "id": "arxiv:2512.04535", "appearances": [ "digests/agent-environments/2026/2026-10-08.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "CTE-Bench: Counterfactual Trace Evaluation for Stateful Software Simulators", "authors": "Xinran Zhang", "team": "Independent Researcher", "doi": "10.48550/arxiv.2609.36647", "arxiv": "2609.36647", "url": "https://arxiv.org/abs/2609.36647", "fulltext_url": "https://arxiv.org/html/2609.36647v1", "first_published": "2026-09-29", "verified_version": "v1", "version_date": "2026-09-29", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_sections", "verified_at": "2026-10-09T07:07:31.847364+08:00", "read_first": "Figure 1; §3; results tables", "id": "arxiv:2609.36647", "appearances": [ "digests/agent-environments/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Fault Injection in OpenAPI Specifications for Evaluating Black-Box Testing Effectiveness", "authors": "Hamza Bin Mazhar; Yuqing Wang; Mika V. Mäntylä", "doi": "10.48550/arxiv.2607.12101", "arxiv": "2607.12101", "url": "https://arxiv.org/abs/2607.12101", "fulltext_url": "https://arxiv.org/html/2607.12101v1", "first_published": "2026-07-13", "verified_version": "v1", "version_date": "2026-07-13", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_sections", "verified_at": "2026-10-09T07:07:31.847364+08:00", "read_first": "fault classes; Table III; RESTler discussion", "id": "arxiv:2607.12101", "appearances": [ "digests/agent-environments/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "AutoRestTest at the SBFT 2026 Tool Competition", "authors": "Tyler Stennett et al.", "team": "Georgia Tech / AWS AI Labs / IBM Research / University of Georgia", "doi": "10.1145/3786155.3795701", "arxiv": "2607.01063", "url": "https://arxiv.org/abs/2607.01063", "fulltext_url": "https://arxiv.org/html/2607.01063v1", "first_published": "2026-07-01", "verified_version": "v1", "version_date": "2026-07-01", "type": "conference_tool_report", "classification": "supplement", "verification": "original_abstract_and_key_method_results_sections", "verified_at": "2026-10-09T07:07:31.847364+08:00", "read_first": "§2.1–2.4; §3", "id": "arxiv:2607.01063", "appearances": [ "digests/agent-environments/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities", "authors": "Jiarui Lu et al.", "team": "Apple", "doi": "10.48550/arxiv.2408.04682", "arxiv": "2408.04682", "url": "https://arxiv.org/abs/2408.04682", "fulltext_url": "https://arxiv.org/html/2408.04682v1", "first_published": "2024-08-08", "verified_version": "v1", "version_date": "2024-08-08", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_sections", "verified_at": "2026-10-09T07:07:31.847364+08:00", "read_first": "Figure 1; §2.3", "secondary_original_url": "https://machinelearning.apple.com/research/toolsandbox-stateful-conversational-llm-benchmark", "id": "arxiv:2408.04682", "appearances": [ "digests/agent-environments/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Weblica: Scalable and Reproducible Training Environments for Visual Web Agents", "authors": "Oğuzhan Fatih Kar et al.", "team": "Apple", "doi": "10.48550/arxiv.2605.06761", "arxiv": "2605.06761", "url": "https://arxiv.org/abs/2605.06761", "fulltext_url": "https://arxiv.org/html/2605.06761v1", "first_published": "2026-05-07", "verified_version": "v1", "version_date": "2026-05-07", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_sections", "verified_at": "2026-10-09T07:07:31.847364+08:00", "read_first": "§3.2–3.3; §5.4; §6", "id": "arxiv:2605.06761", "appearances": [ "digests/agent-environments/2026/2026-10-09.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks", "authors": "Jeonghye Kim, Minseon Kim, Young Jin Kim, Matheus Pereira, Marc-Alexandre Côté, Alessandro Sordoni, Xingdi Yuan, Zhengyan Shi", "team": "KAIST; Microsoft Research Montréal; Microsoft AI", "arxiv": "2609.18805", "url": "https://arxiv.org/abs/2609.18805", "fulltext_url": "https://arxiv.org/html/2609.18805v1", "first_published": "2026-09-16", "verified_version": "v1", "version_date": "2026-09-16", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_limitations_sections", "verified_at": "2026-10-10T07:10:49.384914+08:00", "evidence": "12应用完整重建，590原子行为/413工作流；58.98和49.15为按应用均值，不当原始通过计数", "read_first": "§2.4; §3.3; Table 1", "secondary_original_url": "https://microsoft.github.io/debug-gym/blog/2026/09/programdistill/", "id": "arxiv:2609.18805", "appearances": [ "digests/agent-environments/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Differential Regression Testing for REST APIs", "authors": "Patrice Godefroid, Daniel Lehmann, Marina Polishchuk", "team": "Microsoft Research; University of Stuttgart", "doi": "10.1145/3395363.3397374", "url": "https://www.microsoft.com/en-us/research/publication/differential-regression-testing-for-rest-apis/", "fulltext_url": "https://patricegodefroid.github.io/public_psfiles/issta2020.pdf", "first_published": "2019-08", "verified_version": "ISSTA 2020 revised paper", "version_date": "2020-07", "type": "conference_paper", "classification": "supplement", "verification": "original_abstract_and_key_method_results_limitations_sections", "verified_at": "2026-10-10T07:10:49.384914+08:00", "evidence": "17网络API版本及DNS历史实验；人工确认14回归，5规格9服务；98%粗服务差异为空", "read_first": "§3.3–3.4; §4.5.1", "id": "doi:10.1145/3395363.3397374", "appearances": [ "digests/agent-environments/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "You Can REST Now: Automated REST API Documentation and Testing via LLM-Assisted Request Mutations", "authors": "Alix Decrop et al.", "team": "University of Namur; University of Luxembourg", "doi": "10.48550/arxiv.2402.05102", "arxiv": "2402.05102", "url": "https://arxiv.org/abs/2402.05102", "fulltext_url": "https://arxiv.org/html/2402.05102v2", "first_published": "2024-02-07", "verified_version": "v2", "version_date": "2025-07-15", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_limitations_sections", "verified_at": "2026-10-10T07:10:49.384914+08:00", "evidence": "v2 10 APIs每模型5次；DeepSeek V3平均88.62%路径89.25%参数；3本地未知API各10路径10参数，平均58.67%/71.00%", "read_first": "§3.3–3.4; §4.2; §5.1; Table 5", "id": "arxiv:2402.05102", "appearances": [ "digests/agent-environments/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "RESTOR: Automated Test Oracle Generation for RESTful APIs via Reinforcement Learning", "authors": "Xun Zhou et al.", "team": "Fudan University; ByteDance", "doi": "10.1145/3832185", "arxiv": "2607.23963", "url": "https://arxiv.org/abs/2607.23963", "fulltext_url": "https://arxiv.org/html/2607.23963v1", "first_published": "2026-07-27", "verified_version": "v1", "version_date": "2026-07-27", "type": "preprint_with_publication_metadata", "classification": "supplement", "verification": "original_abstract_and_key_method_results_limitations_sections", "verified_at": "2026-10-10T07:10:49.384914+08:00", "evidence": "超过2300流量样本、246服务15业务线；229留出样本；字段F1 85.42；采用率受人工审核", "read_first": "§3.2.2; §5.1; §7", "id": "arxiv:2607.23963", "appearances": [ "digests/agent-environments/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "WFC/WFD: Web Fuzzing Commons, Dataset and Guidelines to Support Experimentation in REST API Fuzzing", "authors": "Omur Sahin, Man Zhang, Andrea Arcuri", "team": "Erciyes University; Kristiania University College; Beihang University; Oslo Metropolitan University", "doi": "10.48550/arxiv.2509.01612", "arxiv": "2509.01612", "url": "https://arxiv.org/abs/2509.01612", "fulltext_url": "https://arxiv.org/html/2509.01612v1", "first_published": "2025-09-01", "verified_version": "v1", "version_date": "2025-09-01", "type": "preprint", "classification": "supplement", "verification": "original_abstract_and_key_method_results_limitations_sections", "verified_at": "2026-10-10T07:10:49.384914+08:00", "evidence": "36API、6工具、一小时10次名义2160组合；EmRest平台切换和2API不可运行；作者主动披露认证不对称", "read_first": "§4.1; §6; §6.2", "id": "arxiv:2509.01612", "appearances": [ "digests/agent-environments/2026/2026-10-10.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "FDRRestTest: Feedback-Driven Logical Testing for REST APIs", "authors": "Yutong Huo; Dongcheng Li", "team": "China University of Geosciences, Wuhan; Cal Poly Humboldt", "doi": "10.4230/lipics.esem.2026.20", "url": "https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.ESEM.2026.20", "fulltext_url": "https://drops.dagstuhl.de/storage/00lipics/lipics-vol394-esem2026/LIPIcs.ESEM.2026.20/LIPIcs.ESEM.2026.20.pdf", "first_published": "2026-10-05", "verified_version": "publisher proceedings", "version_date": "2026-10-05", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-11T07:08:40.340298+08:00", "evidence": "12 services, 600 seconds, 10 repetitions; FRE .9878 vs .9294 upper estimate; manual 96/120 true defects from 236 findings; lower throughput", "read_first": "§3.4–3.7; Table 2; §4.3.3", "id": "doi:10.4230/lipics.esem.2026.20", "appearances": [ "digests/agent-environments/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "FRE counts multiple signals per response; baseline logical signals reconstructed post hoc; confirmed 96/120 applies only to the reviewed sample, not all findings." }, { "title": "Automated CRUD Semantics Inference for REST APIs", "authors": "Michele Pasqua; Davide Corradini; Michele Perlotto; Mariano Ceccato", "team": "University of Verona; University of Luxembourg; University of Naples Federico II", "doi": "10.1145/3744916.3787826", "url": "https://iris.univr.it/handle/11562/1193047", "fulltext_url": "https://michelepasqua.github.io/publications/pdfs/ICSE26.pdf", "verified_version": "ICSE 2026 author manuscript", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-11T07:08:40.340298+08:00", "evidence": "13 APIs, 395 operations, 10 repetitions; ground truth independent source-code annotations and arbiter; precision .95 recall .65; LLM baseline recall .99", "read_first": "§3.2; Table 3; §4 discussion", "id": "doi:10.1145/3744916.3787826", "appearances": [ "digests/agent-environments/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "CRUD patterns do not cover non-CRUD operations; confirmation may reject legitimate service behavior such as repeat-delete semantics." }, { "title": "Checking Security Properties of Cloud Service REST APIs", "authors": "Vaggelis Atlidakis; Patrice Godefroid; Marina Polishchuk", "team": "Columbia University; Microsoft Research", "doi": "10.1109/icst46399.2020.00046", "url": "https://www.microsoft.com/en-us/research/publication/checking-security-properties-of-cloud-services-rest-apis/", "fulltext_url": "https://patricegodefroid.github.io/public_psfiles/icst2020.pdf", "first_published": "2019-02", "verified_version": "ICST 2020 author paper", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-11T07:08:40.340298+08:00", "evidence": "3 production services,13/19/18 request types,1-hour experiments; optimized preserves exhaustive-mode buckets in these services;7 unique reported bugs across experiments", "read_first": "§III-B; §V resource-leak example", "id": "doi:10.1109/icst46399.2020.00046", "appearances": [ "digests/agent-environments/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "Three services only; optimized mode preserving observed defect classes is not a universal completeness guarantee." }, { "title": "RESTler: Stateful REST API Fuzzing", "authors": "Vaggelis Atlidakis; Patrice Godefroid; Marina Polishchuk", "team": "Columbia University; Microsoft Research", "doi": "10.1109/icse.2019.00083", "arxiv": "1806.09739", "url": "https://www.microsoft.com/en-us/research/publication/restler-stateful-rest-api-fuzzing/", "fulltext_url": "https://www.microsoft.com/en-us/research/wp-content/uploads/2021/03/RESTler.pdf", "first_published": "2018-06-26", "verified_version": "ICSE 2019 full paper", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-11T07:08:40.340298+08:00", "evidence": "controlled blog service dependency/feedback ablations at length3;GitLab strategy experiments TableIII22 bugs vs abstract overall28 distinct scope", "read_first": "Figure 3; Figure 4", "id": "arxiv:1806.09739", "appearances": [ "digests/agent-environments/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "Invalid-prefix pruning can miss failed-call side effects; greater sequence depth need not improve code coverage." }, { "title": "RESTCov: A Tool for Structural Coverage Analysis of REST APIs", "authors": "Tolgahan Bardakci; Serge Demeyer", "team": "University of Antwerp; Flanders Make", "doi": "10.48550/arxiv.2608.28114", "arxiv": "2608.28114", "url": "https://arxiv.org/abs/2608.28114", "fulltext_url": "https://arxiv.org/html/2608.28114v1", "first_published": "2026-08-28", "verified_version": "v1", "version_date": "2026-08-28", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-11T07:08:40.340298+08:00", "evidence": "Petstore demonstration:14paths20operations25parameters100%,14/36 response statuses38.9%;no comprehensive cross-API evaluation in this paper", "read_first": "Petstore coverage table; unmatched request example", "id": "arxiv:2608.28114", "appearances": [ "digests/agent-environments/2026/2026-10-11.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "Petstore demonstration; structural coverage does not establish transition coverage or behavioral correctness." }, { "title": "REST API Testing with Verified LLM-Inferred Dependencies and Response-Driven Refinement", "authors": "Tu Nguyen; Thanh Nguyen; Huy Nguyen; Viet Nguyen; Tien N. Nguyen; Vu Nguyen", "doi": "10.48550/arxiv.2608.17546", "arxiv": "2608.17546", "url": "https://arxiv.org/abs/2608.17546", "fulltext_url": "https://arxiv.org/html/2608.17546v2", "first_published": "2026-08-18", "verified_version": "v2", "version_date": "2026-08-20", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-12T07:10:01.054798+08:00", "evidence": "16 services; GPT-4.1-mini temperature0 five runs; operation coverage92.3%;771/875generated workflows;not matched workflow sets;197 unique5xx not rootbugs", "read_first": "依赖验证与响应修订、操作覆盖表", "limitations": "验证调用也消耗预算；不同工具生成流程集合不同，独特5xx不等于确认根因。", "id": "arxiv:2608.17546", "appearances": [ "digests/agent-environments/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Detecting and Mitigating Flakiness in REST API Fuzzing", "authors": "Man Zhang; Chongyang Shen; Andrea Arcuri; Tao Yue", "doi": "10.48550/arxiv.2603.28452", "arxiv": "2603.28452", "url": "https://arxiv.org/abs/2603.28452", "fulltext_url": "https://arxiv.org/html/2603.28452v1", "first_published": "2026-03-30", "verified_version": "v1", "version_date": "2026-03-30", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-12T07:10:01.054798+08:00", "evidence": "36 APIs;31 exhibit flakiness;near3000 failing tests analyzed;handling one-hour10runs per mode,100test replays;not all services improve", "read_first": "5.1—5.3节与表5", "limitations": "同环境重跑可能遗漏跨环境波动；屏蔽断言减少检查，部分服务失败率仍未改善。", "id": "arxiv:2603.28452", "appearances": [ "digests/agent-environments/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Gecko: A Simulation Environment with Stateful Feedback for Refining Agent Tool Calls", "authors": "Zeyu Zhang; Guohao Li; Zhenchang Xing; Alexandros Apostolopoulos; Yu Lin Lee; Liang Zheng", "team": "ANU; CSIRO Data61 and coauthors", "doi": "10.48550/arxiv.2602.19218", "arxiv": "2602.19218", "url": "https://arxiv.org/abs/2602.19218", "fulltext_url": "https://raw.githubusercontent.com/mlresearch/v306/main/assets/zhang26cv/zhang26cv.pdf", "first_published": "2026-02-22", "verified_version": "arxiv v2; formal ICML2026 PDF", "version_date": "2026-06-09", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-12T07:10:01.054798+08:00", "evidence": "formal PDF verified;GPT4o BFCL overall76.93 to84.62;hybrid tau2 realread simulatedwrite;not service fidelity", "read_first": "第3节、GATS流程及附录H/I", "limitations": "下游任务表现不等于服务行为保真；τ²只读查询使用真实数据库，额外模型和重试成本需计入。", "id": "arxiv:2602.19218", "appearances": [ "digests/agent-environments/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Intelligent REST API Data Fuzzing", "authors": "Patrice Godefroid; Bo-Yuan Huang; Marina Polishchuk", "team": "Microsoft Research", "url": "https://www.microsoft.com/en-us/research/publication/intelligent-rest-api-data-fuzzing/", "fulltext_url": "https://patricegodefroid.github.io/public_psfiles/fse2020.pdf", "first_published": "2019-11", "verified_version": "ESEC/FSE2020", "version_date": "2020-11", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-12T07:10:01.054798+08:00", "evidence": "AzureDNS13request types4withbody;max1000variants/request/rulepair;hundreds500triaged17confirmedbugs across Azure services", "read_first": "3节与3.3错误类型、5节值来源", "limitations": "真实云服务探测；归一化错误类型不等于根因数，需经人工确认。", "id": "url:https://www.microsoft.com/en-us/research/publication/intelligent-rest-api-data-fuzzing/", "appearances": [ "digests/agent-environments/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Log-based, Business-aware REST API Testing", "authors": "Ding Yang; Ruixiang Qian; Zhao Wei; Zhenyu Chen; Chunrong Fang", "doi": "10.48550/arxiv.2604.08007", "arxiv": "2604.08007", "url": "https://arxiv.org/abs/2604.08007", "fulltext_url": "https://arxiv.org/html/2604.08007v1", "first_published": "2026-04-09", "verified_version": "v1", "version_date": "2026-04-09", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-12T07:10:01.054798+08:00", "evidence": "17 services8comparators20runs;smaller1h fullGitLab24h;352.7/1099operations80/109modules;participant simulatedusage logs", "read_first": "日志切片、补全、完整GitLab实验", "limitations": "日志由参与者模拟使用收集，非随机生产样本；完整GitLab操作覆盖约32.1%，非完整行为覆盖。", "id": "arxiv:2604.08007", "appearances": [ "digests/agent-environments/2026/2026-10-12.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Automated Synthesis of Cloud Emulators", "authors": "Archit Bhatnagar; Zhenning Yang; Sarah McClure; Yiming Qiu; Sylvia Ratnasamy; Ang Chen", "team": "UMich / UC Berkeley / HKU", "arxiv": "2608.23842", "url": "https://arxiv.org/abs/2608.23842", "fulltext_url": "https://arxiv.org/html/2608.23842v1", "first_published": "2026-08-24", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-13T09:36:49.872899+08:00", "read_first": "第3.3、4.1—4.3节", "evidence": "AWS参考端上50 CLI及50 Terraform用例；逐级比较错误码、字段及中间状态。另330用例LocalStack消融属于不同参照。", "limitations": "抽样用例不覆盖全部EC2输入；约1.24万行专用框架带有人工作业成本。真实AWS保真证据与LocalStack拟合消融应分开。", "id": "arxiv:2608.23842", "appearances": [ "digests/agent-environments/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Fidelity of Cloud Emulators: The Imitation Game of Testing Cloud-based Software", "authors": "Anna Mazhar; Saad Sher Alam; William X. Zheng; Yinfang Chen; Suman Nath; Tianyin Xu", "team": "Cornell / UIUC / Microsoft Research", "doi": "10.1109/ICSE55347.2025.00077", "url": "https://doi.org/10.1109/ICSE55347.2025.00077", "fulltext_url": "https://tianyin.github.io/pub/cloudtest.pdf", "first_published": "2025-04", "verified_version": "ICSE 2025", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-13T09:36:49.872899+08:00", "read_first": "第III、IV、VII节", "evidence": "五项云服务255个API中94个存在可触发差异，约37%；另用十个开源项目测试检查应用后果。", "limitations": "比例分母为接口，非每次请求的错误概率；研究版本及工作负载有限。已验证调用序列的复用不证明新组合安全。", "id": "doi:10.1109/icse55347.2025.00077", "appearances": [ "digests/agent-environments/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Toward Interactive Understanding of Code APIs", "authors": "Dhananjay Ashok; Jesse Thomason; Jonathan May", "team": "USC / ISI", "arxiv": "2609.32081", "url": "https://arxiv.org/abs/2609.32081", "fulltext_url": "https://arxiv.org/html/2609.32081v1", "first_published": "2026-09-25", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-13T09:36:49.872899+08:00", "read_first": "第2节、图2—3、附录C", "evidence": "741个无状态函数测试，至多25次查询；描述评分及重写代码一致率分别评估。正确描述交给重写模型仍约72.96%一致率。", "limitations": "重写指标包含代码生成误差；结构覆盖不等于语义穷尽。正文模型名和显著性图注有不一致，不能推导完整排名；无跨调用状态。", "id": "arxiv:2609.32081", "appearances": [ "digests/agent-environments/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Push-Button Reliability Testing for Cloud-Backed Applications with Rainmaker", "authors": "Yinfang Chen; Xudong Sun; Suman Nath; Ze Yang; Tianyin Xu", "team": "UIUC / Microsoft Research", "url": "https://www.usenix.org/conference/nsdi23/presentation/chen-yinfang", "fulltext_url": "https://www.usenix.org/system/files/nsdi23-chen-yinfang.pdf", "first_published": "2023-04", "verified_version": "NSDI 2023", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-13T09:36:49.872899+08:00", "read_first": "图3、第4.2.1节", "evidence": "11个.NET云应用73个新发现；USENIX正式论文报告55确认、51修复。HTTP代理区分请求未执行及已执行但响应丢失。", "limitations": "未证明覆盖全部故障及语言客户端；旧微软页面44/40与正式论文不同。研究真实应用故障注入，非生成模拟器。", "id": "url:https://www.usenix.org/conference/nsdi23/presentation/chen-yinfang", "appearances": [ "digests/agent-environments/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Benchmarking Combinations of Learning and Testing Algorithms for Automata Learning", "authors": "Bernhard K. Aichernig; Martin Tappler; Felix Wallner", "team": "TU Graz / Silicon Austria Labs", "doi": "10.1145/3605360", "url": "https://doi.org/10.1145/3605360", "fulltext_url": "https://tugraz.elsevierpure.com/ws/portalfiles/portal/77962395/3605360.pdf", "verified_version": "Formal Aspects of Computing 36(1), article 3", "type": "journal_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-13T09:36:49.872899+08:00", "read_first": "第4.1.4节、图6—7", "evidence": "六种学习与七种测试方法；153个已有模型，十次重复正确学习的最低预算；共同完成116个对象单独比较。", "limitations": "随机Wp抽样没有完整一致性测试的形式保证；模型集合、失败惩罚、12GB内存及实现限制影响比较。无限数据参数及异步行为需要额外抽象。 2024-03为所核期刊卷期，在线首发日期未单独核实。", "id": "doi:10.1145/3605360", "appearances": [ "digests/agent-environments/2026/2026-10-13.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Morest: Model-based RESTful API Testing with Execution Feedback", "authors": "Yi Liu; Yuekang Li; Gelei Deng; Yang Liu; Ruiyuan Wan; Runchao Wu; Dandan Ji; Shiheng Xu; Minli Bao", "team": "Nanyang Technological University / Huawei Cloud et al.", "doi": "10.1145/3510003.3510133", "arxiv": "2204.12148", "url": "https://arxiv.org/abs/2204.12148", "fulltext_url": "https://arxiv.org/html/2204.12148v1", "type": "conference_paper", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-14T07:07:08.897218+08:00", "read_first": "第3节、表2、CRUD序列过滤规则", "evidence": "6 services; 8h per run, 5 repeats; 5/6 significantly better; 44 detected issues, 13 unique to tool; not all developer-confirmed", "first_published": "2022-04-26", "verified_version": "v1", "limitations": "真实REST服务测试，未生成模拟器；CRUD过滤与服务集合限制覆盖范围。44项发现并非全部获开发者确认。可作为服务仿真的调用依赖验证参照。", "id": "arxiv:2204.12148", "appearances": [ "digests/agent-environments/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "DeepREST: Automated Test Case Generation for REST APIs Exploiting Deep Reinforcement Learning", "authors": "Davide Corradini; Zeno Montolli; Michele Pasqua; Mariano Ceccato", "team": "University of Verona", "arxiv": "2408.08594", "url": "https://arxiv.org/abs/2408.08594", "fulltext_url": "https://arxiv.org/html/2408.08594v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-14T07:07:08.897218+08:00", "read_first": "第4.1节、第5.2节、图5", "evidence": "11 APIs, 5 comparison tools; identical per-API request budget, 10 repeats and fresh filesystems; aggregate metrics highest; fault difference versus RESTler not significant", "first_published": "2024-08-16", "verified_version": "v1", "limitations": "成功计数向量不是内部状态真值；5XX不等于已确认根因。相对RESTler的故障差异不显著；提供真实API探索方法，不能视为模拟器保真证明。", "id": "arxiv:2408.08594", "appearances": [ "digests/agent-environments/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "AgentRx: Diagnosing AI Agent Failures from Execution Trajectories", "authors": "Shraddha Barke; Arnav Goyal; Alind Khare; Avaljot Singh; Suman Nath; Chetan Bansal", "team": "Microsoft Research / Microsoft / UIUC", "arxiv": "2602.02475", "url": "https://arxiv.org/abs/2602.02475", "fulltext_url": "https://arxiv.org/html/2602.02475v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-14T07:07:08.897218+08:00", "read_first": "第3节、表5、附录B", "evidence": "115 failed trajectories: 29 tau,42 Flash,44 Magentic; GPT-5 default, human labels; tau best one-shot step 32.2 to54.0 and category25.3 to40.2, 3 runs", "first_published": "2026-02-02", "verified_version": "v1", "limitations": "约束含模型语义判断，不等于机械证明。115条失败轨迹中的子集结果不可概括全部场景；关键未恢复失败与首次状态分歧是不同标签。", "id": "arxiv:2602.02475", "appearances": [ "digests/agent-environments/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Scaling Agentic Data for Long-Horizon Terminal Intelligence", "authors": "Shun Zou; Ziyu Ma; Yi Zou; Yong Wang; Lin Chen; Zehui Chen; Guanhua Chen; Xiangxiang Chu; Feng Zhao", "team": "Author team as listed in original preprint", "arxiv": "2610.13166", "url": "https://arxiv.org/abs/2610.13166", "fulltext_url": "https://arxiv.org/html/2610.13166v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-14T07:07:08.897218+08:00", "read_first": "第2.1—2.3节、图2", "evidence": "1535 environments,2852 trajectories,52.8 mean turns; TB2.1 89 tasks x3,3h/task, mean pass@1 46.82 to60.67; task verifiers", "first_published": "2026-10-09", "verified_version": "v1", "limitations": "整套数据训练的效果未完全拆分环境修整与渐进扩展；长轨迹不等于保真。终端执行与任务验证器证据不能改写为真实外部服务模拟器已通过差分验证。", "id": "arxiv:2610.13166", "appearances": [ "digests/agent-environments/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Dream-RSI: Recursive Self-Improvement through Evolving Worlds", "authors": "Tong Zheng; Xidong Wu; Zheng Zhang; Zhankui He; Chaoyi Zhang; Benjamin Coleman; Ruoqiao Wei; Di Bai; Haolin Liu; Rui Liu; Xue Wang; Yue Zhuan; Wang-Cheng Kang; Renkai Xiang; Heng Huang; Xinwu Cheng; Yunsong Guo", "team": "Google DeepMind / University of Maryland et al.", "arxiv": "2609.14858", "url": "https://arxiv.org/abs/2609.14858", "fulltext_url": "https://arxiv.org/html/2609.14858v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-14T07:07:08.897218+08:00", "read_first": "第3节Online rollout与Offline evaluation、图4", "evidence": "8 tasks; controlled Recursive Fixed Exploration same agent/init/protocol; Gemini3.1Pro on4 KernelBench tasks; comparable VGG16/LayerNorm performance with2.43x/1.79x fewer generations", "first_published": "2026-09-14", "verified_version": "v1", "limitations": "重放只支持已记录分支；缺失延续返回空集合，不能泛化到任意新候选。生成次数节省未包含所有离线策略改写、存储与重放成本。", "id": "arxiv:2609.14858", "appearances": [ "digests/agent-environments/2026/2026-10-14.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "TestPrism: Rethinking Test Evaluation Beyond a Single Reference", "authors": "Han Li; Lingxiang Hu; Jiacheng Huang; Ziqian Jiang; Jingkai Luo; Wei Gao; Yunfan Tan; Zun Wang; Jiaheng Liu", "team": "Nanjing University / Hong Kong Polytechnic University / Tencent", "arxiv": "2610.12289", "url": "https://arxiv.org/abs/2610.12289", "fulltext_url": "https://arxiv.org/html/2610.12289v1", "first_published": "2026-10-08", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-15T07:04:16.614471+08:00", "evidence": "300 tasks, 3000 implementations; best of 14 configurations: single-reference 59.67%, joint success 28.00%", "read_first": "第2.3—2.6节、表2", "id": "arxiv:2610.12289", "limitations": "14种配置中最佳单参考成功率59.67%，联合条件成功率28.00%，分母均为300任务。合法性标签依赖源验证器与专家审阅，不能视为形式正确性证明；本研究提供验证方法，而不是提出一个模拟器。", "appearances": [ "digests/agent-environments/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Memento3: Model-Based Recursive Self-Improvement through Reflective Rulebooks", "authors": "Haoyu Zhao; Zhengxu Yu; Zhiyuan He; Meng Fang; Rasul Tutunov; Haitham Bou Ammar; Weilin Luo; Jun Wang", "team": "UCL / Huawei Noah UK / University of Liverpool", "arxiv": "2610.11794", "url": "https://arxiv.org/abs/2610.11794", "fulltext_url": "https://arxiv.org/html/2610.11794v1", "first_published": "2026-10-08", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-15T07:04:16.614471+08:00", "evidence": "25 games, 183/183 levels, 7518 actions vs human reference 17135; rulebook ablation only 3 games: 617 vs 677 actions", "read_first": "规则本更新、历史重放、图2及消融", "id": "arxiv:2610.11794", "limitations": "作者报告25个游戏、183/183关完成、7518次真实动作，对照人类参考17135次。动作数不包含全部计算；规则本消融只有3个游戏，617对677次动作。历史重放一致不保证未见分支保真，公开排行榜比较也不全是同条件重跑。", "appearances": [ "digests/agent-environments/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments", "authors": "Xiaoyang Yuan; Ji Zhao; Xingyu Hu; Wenhan Wang; Yutao Hu; Qiaosheng Zhang; Pengfei Liu; Yu Qiao; Yi Bin; Wenqi Shao", "team": "Tongji University / Shanghai Innovation Institute / USTC / Southeast University", "arxiv": "2610.16867", "url": "https://arxiv.org/abs/2610.16867", "fulltext_url": "https://arxiv.org/html/2610.16867v1", "first_published": "2026-10-12", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-15T07:04:16.614471+08:00", "evidence": "30395 verified trajectories, 2260 invoked tool names; MCP-Atlas no-key subset 89 tasks: base16.9, SFT52.8, DPO56.2", "read_first": "第3节、附录A与B.1.4", "id": "arxiv:2610.16867", "limitations": "30,395条已验证轨迹和2260个被调用工具名称不等于环境数。Qwen3-Coder-30B-A3B 在89题免密 MCP-Atlas 子集上为16.9、SFT52.8、DPO56.2；消融同时改变来源配比和语料规模。自建契约一致不等于独立真实系统保真，EGTS 截断失败尾部也不能证明失败覆盖。", "appearances": [ "digests/agent-environments/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "When Simulation Lies: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents", "authors": "Xiaolin Zhou; Aojie Yuan; Zheng Luo; Zipeng Ling; Xixiao Pan; Yicheng Gao; Haiyue Zhang; Jiate Li; Shuli Jiang; Prince Zizhuang Wang; Zixuan Zhu; Jinbo Liu; Ryan A. Rossi; Hua Wei; Xiyang Hu", "team": "Arizona State University / USC / Adobe Research et al.", "arxiv": "2605.11928", "url": "https://arxiv.org/abs/2605.11928", "fulltext_url": "https://arxiv.org/html/2605.11928v2", "first_published": "2026-05-12", "verified_version": "v2", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-15T07:04:16.614471+08:00", "evidence": "199 clean +3522 perturbed=3721 instances;21 models;3B ToolRL clean .638 perturbed .404; transition gap .315 to .231 under static domain randomization", "read_first": "第4.3节、表2、第7节", "id": "arxiv:2605.11928", "limitations": "3B ToolRL 干净与扰动成绩为63.8%与40.4%；执行扰动的性能差距在静态域随机化后从31.5降至23.1个百分点。固定首次错误不是自然故障频率，未测试失败后状态副作用或长序列保真。本次补入较早论文及v2，不宣称核实了实质版本增量。", "appearances": [ "digests/agent-environments/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "ARTIS: Agentic Risk-Aware Test-Time Scaling via Iterative Simulation", "authors": "Xingshan Zeng; Lingzhi Wang; Weiwen Liu; Liangyou Li; Yasheng Wang; Lifeng Shang; Xin Jiang; Qun Liu", "team": "Huawei / Harbin Institute of Technology Shenzhen / Shanghai Jiao Tong University", "arxiv": "2602.01709", "url": "https://arxiv.org/abs/2602.01709", "fulltext_url": "https://arxiv.org/html/2602.01709v3", "first_published": "2026-02-02", "verified_version": "v3", "type": "preprint", "verification": "original_abstract_and_key_method_results_discussion", "verified_at": "2026-10-15T07:04:16.614471+08:00", "evidence": "Simulator accuracy98.57→99.14 and failure macro-F1 81.39→93.42; strict unseen-tool task accuracy 17.00±1.90 vs16.23±1.65 over3 runs", "read_first": "第4.3节、失败分类结果、附录H—I", "id": "arxiv:2602.01709", "limitations": "整体模拟准确率98.57%至99.14%，失败宏平均F1为81.39%至93.42%，反映指标关注点不同。严格8组未见工具、3次运行的成绩17±1.90对16.23±1.65，差距有限；主要评测存在工具重叠，相同试次K不代表相同总计算预算。较早论文v3的补入不表示核实了版本增量。", "appearances": [ "digests/agent-environments/2026/2026-10-15.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "SABER: Small Actions, Big Errors — Safeguarding Mutating Steps in LLM Agents", "authors": "Alejandro Cuadron; Pengfei Yu; Yang Liu; Arpit Gupta", "team": "Amazon AGI Foundations", "arxiv": "2512.07850", "url": "https://arxiv.org/abs/2512.07850", "fulltext_url": "https://arxiv.org/html/2512.07850", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-16T07:14:29.229284+08:00", "evidence": "Qwen3-Thinking airline success 49.3→63.3 over three runs; manual annotation issues31/50 airline,53/115 retail; gains not universal", "read_first": "任务修订部分、表2与动作偏差定义", "id": "arxiv:2512.07850", "first_published": "2025-11-26", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "作者手工标注的任务问题不是任何判据下都必然错误。动作数量差回归不能定位首次因果错误，收益并非全部模型和领域一致；本研究改善代理行为，并未重建黑盒服务。" }, { "title": "τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment", "authors": "Victor Barres; Honghua Dong; Soham Ray; Xujie Si; Karthik Narasimhan", "team": "Sierra / University of Toronto / Vector Institute", "arxiv": "2506.07982", "url": "https://arxiv.org/abs/2506.07982", "fulltext_url": "https://arxiv.org/html/2506.07982", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-16T07:14:29.229284+08:00", "evidence": "Telecom114 sampled from2285; four runs per task; three critical user errors in50 manually inspected telecom traces", "read_first": "电信任务构造、共享状态示例及人工质量表", "id": "arxiv:2506.07982", "first_published": "2025-06-09", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "NoUser对照同时改变通信、信息与工具权限。50条电信轨迹的表格和正文有3条严重模拟错误，与表注冲突；当前代码不一定对应论文版本。生成环境的自建断言不等于独立真实服务保真。" }, { "title": "Identifying the Risks of LM Agents with an LM-Emulated Sandbox", "authors": "Yangjun Ruan; Honghua Dong; Andrew Wang; Silviu Pitis; Yongchao Zhou; Jimmy Ba; Yann Dubois; Chris J. Maddison; Tatsunori Hashimoto", "team": "Stanford University / University of Toronto", "arxiv": "2309.15817", "url": "https://arxiv.org/abs/2309.15817", "fulltext_url": "https://arxiv.org/html/2309.15817v2", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-16T07:14:29.229284+08:00", "evidence": "100cases yield200trajectories; four human raters; failure precision72.5±7.1% standard,68.8±6.7% adversarial; six of seven terminal failures reproduced", "read_first": "模拟器输入、隐含状态与人工失败验证规则", "id": "arxiv:2309.15817", "first_published": "2023-09-25", "verified_version": "v2", "appearances": [ "digests/agent-environments/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "72.5%与68.8%是人工支持的失败告警精度，非API保真率；7套工具无现成公共API。终端7例中6例重现仅支持该窄样本，隐含状态仍需独立验证。" }, { "title": "Handling Web Service Interactions in Fuzzing with Search-Based Mock-Generation", "authors": "Susruthan Seran; Man Zhang; Onur Duman; Andrea Arcuri", "team": "Kristiania / software testing research", "doi": "10.1145/3731558", "url": "https://doi.org/10.1145/3731558", "fulltext_url": "https://www.kristiania.no/contentassets/0459de09219c48b8ba3d064d66239d10/susruthan-seran-nett.pdf", "type": "journal_paper", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-16T07:14:29.229284+08:00", "evidence": "Five real REST APIs;30 repeats per configuration,1h each; online pay-publicapi line coverage11.1→31.8%; gains nonuniform", "read_first": "作者学位论文第2章mock生成过程、表2.4—2.5", "id": "doi:10.1145/3731558", "first_published": "2025", "verified_version": "author-thesis-chapter-2", "appearances": [ "digests/agent-environments/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "出版社正文受限，证据来自作者学位论文第2章同篇全文。卷期元数据为2026-01-21，学位论文列2025，首次在线日期未解决。白盒插桩帮助生成mock；覆盖提升不等于外部服务保真，HTTP500计数不等于确认的独立缺陷。" }, { "title": "WorldSimProbe: Diagnosing Simulator Faithfulness in Action-Conditioned World Models for Embodied Manipulation", "authors": "Peterson Co; Sicheng Hu; Chunxuan Jiao; Hongyang Cheng; Yulin Luo; Yijie Xu; Sixiang Chen; Zhongxia Zhao; Zihao Wang; DaFeng Chi; Peidong Liu; YuTong Chen; Henghua Liu; Zhihao Yuan; Huizhu Jia; Yuzheng Zhuang; Tianle Zhang; Liang Lin; Huajie Tan; Shanghang Zhang", "team": "Peking University / EvoPhysAI and collaborators", "arxiv": "2608.09298", "url": "https://arxiv.org/abs/2608.09298", "fulltext_url": "https://arxiv.org/html/2608.09298", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-16T07:14:29.229284+08:00", "evidence": "Six models over18000 instances;500OOD videos VLM action-following91.7% vs human42.2%", "read_first": "第4节局部扰动与无接触测试、第5.5节评判器验证", "id": "arxiv:2608.09298", "first_published": "2026-08-10", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-16.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "物理运动检查不能直接换算成API状态指标。视觉评判在500个分布外视频上的91.7%与人类42.2%差异，仅适用于此动作跟随判断；不能推广为所有模型裁判失效。" }, { "title": "Inferring data model from service interactions for response generation in service virtualization", "authors": "Md. Arafat Hossain; Jiaojiao Jiang; Jun Han; Muhammad Ashad Kabir; Jean-Guy Schneider; Chengfei Liu", "team": "Swinburne / UNSW / Charles Sturt / Deakin", "doi": "10.1016/j.infsof.2021.106803", "url": "https://doi.org/10.1016/j.infsof.2021.106803", "fulltext_url": "https://www.sciencedirect.com/science/article/pii/S0950584921002391", "type": "journal_article", "verification": "original_abstract_intro_section_excerpts_only", "verified_at": "2026-10-17T07:11:27.787841+08:00", "evidence": "Original introduction describes entity/attribute and key inference, live data model plus control model; full empirical tables not verified; no numeric accuracy claim", "read_first": "第2节人员信息例子、第3节模型构造；第4节原实验表待核", "id": "doi:10.1016/j.infsof.2021.106803", "first_published": null, "verified_version": null, "appearances": [ "digests/agent-environments/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "Issue date 2022-05; first online date unresolved；Accepted manuscript HTTP403; publisher direct open intermittent; abstract, introduction and section excerpts only" }, { "title": "GTDD: Generative Test-Driven Development for AI Coding Agents with Adversarial Testing", "authors": "Masahiro Kato", "team": "Mizuho-DL Financial Technology", "arxiv": "2610.02952", "url": "https://arxiv.org/abs/2610.02952", "fulltext_url": "https://arxiv.org/html/2610.02952v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-17T07:11:27.787841+08:00", "evidence": "30 paired blocks; 2000 held-out traces of10–40 actions; dynamic informed27.4% vs static40.5% trajectory failure; property29.8%, no significant dynamic advantage; budget attribution unresolved", "read_first": "第6.1–6.3节、表5/6、独立审计抽样条件", "id": "arxiv:2610.02952", "first_published": "2026-10-02", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "Development did not use audit as an actual acceptance gate；Finite-population fresh-sample guarantee is conditional, not all-sequence correctness" }, { "title": "MCP-Universe: Benchmarking Large Language Models with Real-World Model Context Protocol Servers", "authors": "Ziyang Luo; Zhiqi Shen; Wenzhuo Yang; Zirui Zhao; Prathyusha Jwalapuram; Amrita Saha; Doyen Sahoo; Silvio Savarese; Caiming Xiong; Junnan Li", "team": "Salesforce AI Research", "arxiv": "2508.14704", "url": "https://arxiv.org/abs/2508.14704", "fulltext_url": "https://arxiv.org/html/2508.14704v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-17T07:11:27.787841+08:00", "evidence": "231 tasks;6 domains;11 servers;133 tools;84 distinct evaluators; GPT-5 historical task success101/231=43.72%, average evaluator satisfaction60.23%; intentionally hard task selection", "read_first": "第3.4节和表2/3", "id": "arxiv:2508.14704", "first_published": "2025-08-20", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "Paper-version results only, not current leaderboard；Official blog and paper merged; evolving repository not claimed as verified substantive update" }, { "title": "Which Metadata Matters for LLM Tool Selection in MCP? Evidence from Controlled Factor-Level Interventions", "authors": "Xiaofan Li; Fengchun Qiao; Xing Gao", "team": "University of Delaware / University of South Florida", "arxiv": "2610.14334", "url": "https://arxiv.org/abs/2610.14334", "fulltext_url": "https://arxiv.org/html/2610.14334v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-17T07:11:27.787841+08:00", "evidence": "367174 tool metadata inventory; controlled5tasks5models; PlaceOrder 2-tool GPT5.4mini flat selection97% with first-position51%; no tool execution measured", "read_first": "第4节、第5.2节位置效应、第5.4节明确请求条件", "id": "arxiv:2610.14334", "first_published": "2026-10-09", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "First tool selection only; no task execution or backend state fidelity measured." }, { "title": "Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance?", "authors": "Abhishek Kadian; Joanne Truong; Aaron Gokaslan; Alexander Clegg; Erik Wijmans; Stefan Lee; Manolis Savva; Sonia Chernova; Dhruv Batra", "team": "Facebook AI Research / Georgia Tech / Oregon State / Simon Fraser", "doi": "10.1109/LRA.2020.3013848", "arxiv": "1912.06321", "url": "https://arxiv.org/abs/1912.06321", "fulltext_url": "https://arxiv.org/pdf/1912.06321", "type": "journal_article", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-17T07:11:27.787841+08:00", "evidence": "9models ×3configs ×5start-goal pairs ×3repeats=405 runs per side; SRCCsuccess.18→.844, SPL.603→.875 after simulator adjustment; same lab/model set used for tuning", "read_first": "第IV-E节及第V-A/V-B节图5–7", "id": "arxiv:1912.06321", "first_published": "2019-12-13", "verified_version": "v2, 2020-08-17; RA-L2020", "appearances": [ "digests/agent-environments/2026/2026-10-17.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "HTML affiliation corruption resolved with original PDF page1；Tuning result on same lab and model suite, not unseen-environment validation" }, { "title": "Mining Service Behavior for Stateful Service Emulation", "authors": "Md Arafat Hossain; Jun Han; Muhammad Ashad Kabir; Steve Versteeg; Jean-Guy Schneider; Jiaojiao Jiang", "team": "Swinburne / Charles Sturt / Monash / UNSW", "arxiv": "2510.18519", "url": "https://arxiv.org/abs/2510.18519", "fulltext_url": "https://arxiv.org/html/2510.18519v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-18T07:05:04.711409+08:00", "evidence": "十折按记录划分；LDAP完全相同87.98%对Whole-Cluster OSV33%；SOAP43.67%，不能正确演化余额。", "read_first": "第2节、第4.2—4.3节、第6.4节银行结果", "id": "arxiv:2510.18519", "first_published": "2025-10-21", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "先读状态构造及银行实验，检查记录间依赖和未见交互。该2025年arXiv版本源于更早博士研究，不能以预印本日期宣称思想首次提出；它也不同于后续关系数据模型论文，本次分别去重保留。" }, { "title": "Fara1.5 – A family of frontier computer use agent models", "authors": "Ahmed Awadallah et al.", "team": "Microsoft Research", "url": "https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/", "fulltext_url": "https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/", "type": "official_technical_article", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-18T07:05:04.711409+08:00", "evidence": "表5同Qwen3.5-9B基础模型四域迁移73.4%→83.4%；比较训练数据不同，不能单独归因副本保真。", "read_first": "Creating Synthetic Environments及表5", "id": "url:https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/", "first_published": "2026-05-21", "verified_version": "official article; May21; July release note not classified as material update", "appearances": [ "digests/agent-environments/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "这里收录的是5月21日官方文章，七月发布说明没有被认定为本次实质版本更新。文章属于公司技术披露，未独立复现；可先读Creating Synthetic Environments与表5，追问副本状态、验收规则与迁移条件如何对应。" }, { "title": "Trajectory-Level Speculation with Semantic Verification for Tool-Using Agents", "authors": "Zilin Wang; Yinhong Liu; Chao Chen; Qun Liu; Zhijiang Guo", "team": "LARK/HKUST(GZ), Yinwang, Cambridge, HKUST", "arxiv": "2610.21463", "url": "https://arxiv.org/abs/2610.21463", "fulltext_url": "https://arxiv.org/html/2610.21463v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-18T07:05:04.711409+08:00", "evidence": "τ²每题3轨迹，航空50题零售114题；真实MCP101题中Qwen组72.3%→59.4%，105s→120s，非普遍同质提速。", "read_first": "第3.2节、算法1及表3", "id": "arxiv:2610.21463", "first_published": "2026-10-15", "verified_version": "v1, 2026-10-15T08:06:32Z", "appearances": [ "digests/agent-environments/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "这项工作提供执行验证方法，未提出已经证明保真的通用模拟器。阅读第3.2节、算法1及表3时，应确认检查失败后究竟撤销什么、环境是否支持回滚，以及成功率与时间是否来自相同任务和配置。" }, { "title": "Semantic Representation for Agentic Self-Play in Automated Laboratories", "authors": "Yves Delley; John Coller et al.", "team": "Google Research / Google Cloud / Google DeepMind", "url": "https://www.gstatic.com/science-ai/verid_semantic_lab.pdf", "fulltext_url": "https://www.gstatic.com/science-ai/verid_semantic_lab.pdf", "type": "technical_paper", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-18T07:05:04.711409+08:00", "evidence": "三个染料案例；案例1人工参与根因诊断，修正后逐步稀释1.92–1.99倍；读板失败后未重做已完成移液。", "read_first": "第2.2—2.4节、案例1及图2", "id": "url:https://www.gstatic.com/science-ai/verid_semantic_lab.pdf", "first_published": "2026-10-16", "verified_version": "original PDF dated 2026-10-16", "appearances": [ "digests/agent-environments/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "案例数量有限，人工诊断不能从自主性描述中省略，局部稀释结果也不是总体任务成功率。先读第2.2–2.4节、案例1与图2，分别检查状态表示、模拟检查和实际设备反馈。本文为原始技术论文，未独立复现。" }, { "title": "Evaluating LLMs as Agentic Workflow Executors in Real-World Customer Support Skills", "authors": "Md Tahmid Rahman Laskar; Gundeep Singh; Xue-Yong Fu et al.", "team": "Dialpad", "arxiv": "2610.14346", "url": "https://arxiv.org/abs/2610.14346", "fulltext_url": "https://arxiv.org/html/2610.14346v2", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-18T07:05:04.711409+08:00", "evidence": "最终80技能，500场景，281工具，17模型5裁判；10对裁判κ均值.524，无独立人类参考不能视一致率为准确率。", "read_first": "第3.1.4节、第4.2节、表1及表3", "id": "arxiv:2610.14346", "first_published": "2026-10-09", "verified_version": "v2, 2026-10-14T16:00:21Z", "appearances": [ "digests/agent-environments/2026/2026-10-18.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "该工作可为工具服务仿真的场景构造和评测提供参照，但并未据此证明生产保真。此次首次入库的是v2（10月14日），首发10月9日；没有对v1至v2进行实质差异认定。先读第3.1.4节、第4.2节与表1/3。" }, { "title": "Systematic API Testing Through Model Checking and Executable Contracts", "authors": "Ana Ribeiro; Margarida Mamede; Carla Ferreira", "team": "NOVA University Lisbon", "arxiv": "2604.08633", "url": "https://arxiv.org/abs/2604.08633", "fulltext_url": "https://arxiv.org/html/2604.08633v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-19T07:08:36.899104+08:00", "evidence": "三系统；Petstore完整状态但74%—84%转移覆盖；97为操作级错误次数；12值比赛模型467GB/16小时以上。", "read_first": "第IV-C、IV-D节，表IV/V", "id": "arxiv:2604.08633", "first_published": "2026-04-09", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "有限状态模型需人工编写；完整状态覆盖不保证完整转移覆盖，状态爆炸限制可行规模。真实API测试为仿真行为验证提供方法，不是服务仿真器。" }, { "title": "Multi-Agent LLM-based Metamorphic Testing for REST APIs", "authors": "Shehroz Khan; Abdullah Mughees; Gaadha Sudheerbabu; Tanwir Ahmad; Dragos Truscan", "team": "Åbo Akademi University", "arxiv": "2605.28321", "url": "https://arxiv.org/abs/2605.28321", "fulltext_url": "https://arxiv.org/html/2605.28321v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-19T07:08:36.899104+08:00", "evidence": "两系统各10会话、1000请求/30分钟；117失败中92作者判真问题，204中116；非独立缺陷数或缺陷召回。", "read_first": "图1，第III节与第IV-C/IV-D节", "id": "arxiv:2605.28321", "first_published": "2026-05-27", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "失败测试中的人工真问题比例不等于独立缺陷数或缺陷召回；变形关系和前置状态均须核实。真实REST服务执行，不是服务仿真。会议接收按作者稿归属。" }, { "title": "Cowork-Gym: Synthesizing Interactive Work Environments at Scale", "authors": "Anni Zou; Cheng Chen; Jiaji Deng; Weikang Zhou; Xinji Mai; Yue Wu; Jianing Lin; Lipeng Xie; Lixiong Qin; Qingyue Zhao; Que Shen; Shuchang Tao; Xuchen Pan; Yaliang Li; Yunpeng Zhai; Yushuo Chen; Zhihai Wang; Zhuo Zhang; Zhaoyang Liu; Bolin Ding; Jingren Zhou", "team": "Alibaba Token Foundry", "arxiv": "2610.17670", "url": "https://arxiv.org/abs/2610.17670", "fulltext_url": "https://arxiv.org/html/2610.17670v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-19T07:08:36.899104+08:00", "evidence": "64888环境；77729轨迹中过滤有效24732、SFT17312、RL2873环境；六benchmark平均分增8.7/6.4/4.1点，17/18模型任务组合改善。", "read_first": "第3.2节、第4节，表1/2", "id": "arxiv:2610.17670", "first_published": "2026-10-13", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "训练增益不等于对既有真实服务的行为保真；环境与轨迹过滤损耗、不同benchmark评分和个别退步必须保留。" }, { "title": "AutoEnvScaling: Automating the Data Flywheel with Terminal Agents", "authors": "Simon Yu; Baolin Peng; Bo Liu; Zhengyan Shi; Zhaoyang Wang; Hao Cheng; Qianhui Wu; Wenlin Yao; Natasha Jaques; Weiyan Shi; Jianfeng Gao", "team": "Microsoft; Northeastern; Stanford; University of Washington; UNC Chapel Hill", "arxiv": "2610.14163", "url": "https://arxiv.org/abs/2610.14163", "fulltext_url": "https://arxiv.org/html/2610.14163v1", "type": "preprint", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-19T07:08:36.899104+08:00", "evidence": "八提出模型三方式；Sol有效环境41.1%到86.3%，成本0.90到0.59美元/有效环境；两模型成本反增；附录F同任务开关自检有效。", "read_first": "第3.2节，表2，附录F", "id": "arxiv:2610.14163", "first_published": "2026-10-09", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "有效环境要求构建成功、参考解通过及空解失败；控制更新步数不等于匹配总计算。部分模型成本增加，自检仍可能误判。" }, { "title": "Demystifying evals for AI agents", "authors": "Mikaela Grace; Jeremy Hadfield; Rodrigo Olivares; Jiri De Jonghe", "team": "Anthropic", "url": "https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents", "fulltext_url": "https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents", "type": "official_technical_article", "verification": "original_official_article_fulltext", "verified_at": "2026-10-19T07:08:36.899104+08:00", "evidence": "官方工程经验总结；环境状态与交互记录、隔离运行和评分器校准，无统一分母和受控实验。", "read_first": "The structure of an evaluation；From zero to one的任务歧义/隔离/评分", "id": "https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents", "first_published": "2026-01-09", "verified_version": "official article dated 2026-01-09", "appearances": [ "digests/agent-environments/2026/2026-10-19.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "官方工程经验总结，非受控实验；任务完成、环境状态和服务行为保真是不同评价对象。" }, { "title": "EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL", "arxiv": "2605.18703", "authors": "Minrui Xu; Zilin Wang; Mengyi Deng; Zhiwei Li; Zhicheng Yang; Xiao Zhu; Yinhong Liu; Boyu Zhu; Baiyu Huang; Chao Chen; Heyuan Deng; Fei Mi; Lifeng Shang; Xingshan Zeng; Zhijiang Guo", "team": "HKUST (Guangzhou); Huawei; Cambridge; UCL", "url": "https://arxiv.org/abs/2605.18703", "fulltext_url": "https://arxiv.org/html/2605.18703v1", "type": "preprint", "evidence": "85环境7领域；Qwen3-4B BFCL多轮准确率33.50%至48.50%，MCPAtlas通过率4.12%至9.97%；下游训练收益不等于环境忠实度。", "read_first": "3.2—3.3节、图1、表1", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-20T07:10:28.972380+08:00", "id": "arxiv:2605.18703", "first_published": "2026-05-18", "verified_version": "v1", "appearances": [ "digests/agent-environments/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "环境自检及下游训练收益不能替代独立参考服务行为保真；生成方案数据量不同，非同预算组件对照。" }, { "title": "Generating Mock Skeletons for Lightweight Web-Service Testing with Description Logic Learning", "doi": "10.1007/s11761-026-00495-w", "authors": "Thilini Bhagya; Jens Dietrich; Hans Guesgen", "team": "Authors and affiliations verified on Springer original article", "url": "https://link.springer.com/article/10.1007/s11761-026-00495-w", "fulltext_url": "https://link.springer.com/article/10.1007/s11761-026-00495-w", "type": "journal_article", "evidence": "OCEL描述逻辑二分类，120秒/目标，两折验证；Google Tasks 503有2正1122负，准确率0.9893而精确率和召回率为0。", "read_first": "5.2历史表示；第6节503结果", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-20T07:10:28.972380+08:00", "id": "doi:10.1007/s11761-026-00495-w", "first_published": "2026-05-22", "verified_version": "published fulltext", "appearances": [ "digests/agent-environments/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "响应字段二分类不等于共享状态执行；部分样本抽样、两折验证；稀有失败类别可能完全漏识别。" }, { "title": "Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments", "arxiv": "2602.11964", "authors": "Romain Froger; Pierre Andrews; Matteo Bettini; et al. (original author list linked)", "team": "Meta Superintelligence Labs", "url": "https://arxiv.org/abs/2602.11964", "fulltext_url": "https://arxiv.org/html/2602.11964v1", "type": "preprint_with_acceptance_note", "evidence": "12类应用101工具；450人工标注轨迹中结构化验证器一致率0.98、仅LLM判断0.72；测任务评分而非服务重建。", "read_first": "第4节和表1，再读第3节", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-20T07:10:28.972380+08:00", "id": "arxiv:2602.11964", "first_published": "2026-02-12", "verified_version": "v1; original page states ICLR 2026 Oral", "appearances": [ "digests/agent-environments/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "验证450条轨迹的任务完成判定，而非模拟器对真实服务的保真；自由文本仍使用模型评分。" }, { "title": "Automated REST API Black-box Test Generation in Practice: An Experience Report from Industry", "doi": "10.1109/ICST69053.2026.00064", "authors": "Davide Corradini; Seung Yeob Shin; Domenico Bianculli", "team": "University of Luxembourg SnT / SVV; industry collaboration", "url": "https://orbilu.uni.lu/handle/10993/68146", "fulltext_url": "https://orbilu.uni.lu/bitstream/10993/68146/1/ICST-2026-REST-test-generation-in-practice.pdf", "type": "conference_paper", "evidence": "5 API共204操作；每工具每API10运行、每运行200×操作数请求；各API中位覆盖合计151/204，两工具相同；故障中位汇总17/18。", "read_first": "V-A/B、表III；VI-B", "verification": "original_metadata_and_key_fulltext_method_results", "verified_at": "2026-10-20T07:10:28.972380+08:00", "id": "doi:10.1109/icst69053.2026.00064", "first_published": "2026-07", "verified_version": "author postprint; ICST 2026 institutional record; institutional deposit 2026-04-03, first online date unknown", "appearances": [ "digests/agent-environments/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "2026-04-03仅为机构库入库日；论文集为2026年7月，首次在线日未确立。真实API测试非服务仿真；操作覆盖不等于分支或状态覆盖。" }, { "title": "Four evaluation engineering lessons learned from building agent plugins", "authors": "Casey West", "team": "Google Cloud", "url": "https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons/", "fulltext_url": "https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons/", "type": "official_technical_article", "evidence": "初始11提示中3个误判的工程案例；技能加载与文档搜索同回合并发解释了指令不可见；非受控普遍结论。", "read_first": "Lesson 2及Lesson 1的原始日志案例", "verification": "original_official_fulltext", "verified_at": "2026-10-20T07:10:28.972380+08:00", "id": "https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons/", "first_published": "2026-10-16", "verified_version": "official article dated 2026-10-16", "appearances": [ "digests/agent-environments/2026/2026-10-20.md" ], "provenance": "ai-assisted-literature-notes", "limitations": "官方工程案例非受控实验；初始11条提示中3个误判不构成普遍误判率；计划运行不能证明真实工具状态正确。" }, { "title": "Agent World Model in the Wild: Computer-Use Agent Environment Synthesis", "authors": "Zhaoyang Wang et al.", "team": "Microsoft Research, UNC-Chapel Hill et al.", "arxiv": "2610.26828", "url": "https://arxiv.org/abs/2610.26828", "fulltext_url": "https://arxiv.org/html/2610.26828v1", "first_published": "2026-10-19", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_html_key_sections", "verified_at": "2026-10-21T07:09:53.036691+08:00", "read_first": "任务合成五道质量门、表3", "evidence": "36个人物、640个状态、341环境任务对；6轨迹和三层判定；8项OOD平均收益不等于服务保真。", "id": "arxiv:2610.26828", "limitations": "多评判一致仍可能共同漏掉约束；训练迁移成绩不是固定真实服务逐接口保真测量。", "appearances": [ "digests/agent-environments/2026/2026-10-21.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy", "authors": "Ke Zhang et al.", "team": "UC Riverside, UC Irvine", "arxiv": "2610.26836", "url": "https://arxiv.org/abs/2610.26836", "fulltext_url": "https://arxiv.org/html/2610.26836v1", "first_published": "2026-10-19", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_html_key_sections", "verified_at": "2026-10-21T07:09:53.036691+08:00", "read_first": "图1、表1、第4节冻结模拟实验", "evidence": "158题、11007模拟调用重放；GPT5.2冻结模拟3次配对损失9.4pp；主比较多重校正未显著。", "id": "arxiv:2610.26836", "limitations": "GPT-5.2主比较多重校正后未显著；冻结模拟实验的配对结果与该主比较不是同一统计检验。范围限PHREEQC任务和相应交付方式，不外推所有API。", "appearances": [ "digests/agent-environments/2026/2026-10-21.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Building to the Test: Coding Agents Deliver What You Check, Not What You Requested", "authors": "Yanuo Ma, Ben Kereopa-Yorke, Ben Schultz", "team": "Microsoft", "arxiv": "2606.28430", "url": "https://arxiv.org/abs/2606.28430", "fulltext_url": "https://arxiv.org/html/2606.28430v1", "first_published": "2026-06-26", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_html_key_sections", "verified_at": "2026-10-21T07:09:53.036691+08:00", "read_first": "第3节机械审计与无操作替换", "evidence": "2模型×3条件×3重复、222测试；两运行均满分而停用缩放库后29/29与17/29，支持依赖审计。", "id": "arxiv:2606.28430", "limitations": "单一UI重实现任务的受控案例研究，不能估计全部编码任务的发生率。内部架构只有在交付契约要求时才应评分；本研究并非服务模拟器。", "appearances": [ "digests/agent-environments/2026/2026-10-21.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Qwen-AgentWorld: Language World Models for General Agents", "authors": "Qwen Team", "team": "Qwen / Alibaba", "arxiv": "2606.24597", "url": "https://arxiv.org/abs/2606.24597", "fulltext_url": "https://arxiv.org/html/2606.24597v1", "first_published": "2026-06-23", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_html_key_sections", "verified_at": "2026-10-21T07:09:53.036691+08:00", "read_first": "第2.2节输入格式、第4节评测协议", "evidence": "7域真实参考观测评分；35B平均47.73到56.39，397B54.74到58.71，为LLM五维归一评分而非正确率。", "id": "arxiv:2606.24597", "limitations": "五维参考辅助模型评分不是接口正确率；轮级真实历史条件阻断此前模拟误差累积，不能替代自回归连续状态保持测试。", "appearances": [ "digests/agent-environments/2026/2026-10-21.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Use", "authors": "Zijian Wu et al.", "team": "NUS TRAIL, EvalSys, LobeHub, SJTU, Fudan", "arxiv": "2509.24002", "url": "https://arxiv.org/abs/2509.24002", "fulltext_url": "https://arxiv.org/html/2509.24002v1", "first_published": "2025-09-28", "verified_version": "v1", "type": "preprint", "verification": "original_abstract_and_html_key_sections", "verified_at": "2026-10-21T07:09:53.036691+08:00", "read_first": "第2节初始状态与验证器、第3.1节指标", "evidence": "127任务、38初态、5服务、4独立运行；GPT5medium pass@1平均52.56%、pass^4 33.86%。", "id": "arxiv:2509.24002", "limitations": "真实服务上的智能体任务成绩混合规划、随机性和服务行为；并未比较参考服务与模拟服务，不能单凭该成绩归因模拟保真。", "appearances": [ "digests/agent-environments/2026/2026-10-21.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents", "authors": "Yuyao Ge et al.", "team": "CAS ICT et al.", "arxiv": "2610.20344", "url": "https://arxiv.org/abs/2610.20344", "fulltext_url": "https://arxiv.org/html/2610.20344v1", "type": "preprint", "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": "2026-10-22T07:10:59.255180+08:00", "evidence": "19模型各3构建；328测试，130带对照；最佳53.05%，分母不含对照。", "read_first": "第2.4节、图2", "id": "arxiv:2610.20344", "first_published": "2026-10-14", "verified_version": "v1", "limitations": "完整规格合成不同于未知服务重建；合法对照不额外计入328项分母。", "appearances": [ "digests/agent-environments/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "TRACE: Diagnosing Verifier Brittleness in Agentic Evaluation", "authors": "Radhika Gaonkar", "team": "Prime Intellect", "arxiv": "2610.11678", "url": "https://arxiv.org/abs/2610.11678", "fulltext_url": "https://arxiv.org/html/2610.11678v1", "type": "preprint", "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": "2026-10-22T07:10:59.255180+08:00", "evidence": "25合成任务；88新公共任务每条件3次，7/8比较落入±0.10；重跑翻转15–36%。", "read_first": "第3.1节、第5节", "id": "arxiv:2610.11678", "first_published": "2026-10-08", "verified_version": "v1", "limitations": "扩大实验并未支持普遍显著下降；等效界不等于完全相同；重跑存在较大逐任务波动。", "appearances": [ "digests/agent-environments/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "TRAP: Tool-result Replay Attack for Probing Misbinding in LLM Agents", "authors": "Yutong Zhi, Jie Zhang, Xiuqi Lu, Jiayuan Xu, Jing Shao", "team": "Tsinghua University; Shanghai AI Laboratory", "arxiv": "2610.16333", "url": "https://arxiv.org/abs/2610.16333", "fulltext_url": "https://arxiv.org/html/2610.16333v1", "type": "preprint", "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": "2026-10-22T07:10:59.255180+08:00", "evidence": "8模型×3框架×7场景=168组合，通常各10有效试验；最终状态错误率均值0.649，非自然故障率。", "read_first": "第3.3节、附录E", "id": "arxiv:2610.16333", "first_published": "2026-10-12", "verified_version": "v1", "limitations": "受控注入条件概率并非自然故障率；忠实仿真以参考服务语义为准。", "appearances": [ "digests/agent-environments/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "EnvHarness: Awakening Static Worlds for Agent Learning", "authors": "Chengsong Huang et al.", "team": "Google Cloud AI Research et al.", "arxiv": "2608.19880", "url": "https://arxiv.org/abs/2608.19880", "fulltext_url": "https://arxiv.org/html/2608.19880v1", "type": "preprint", "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": "2026-10-22T07:10:59.255180+08:00", "evidence": "ALFWorld OOD61.4→70.4；SWE407问题平均步数55.01→49.61；两个基准不混合；生成另耗轨迹。", "read_first": "第2.2节、第4.1节", "id": "arxiv:2608.19880", "first_published": "2026-08-20", "verified_version": "v1", "limitations": "主自动流程排除Chain；两个基准指标不可合并；环境生成额外消耗交互。", "appearances": [ "digests/agent-environments/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" }, { "title": "How to test agent skills without hitting real APIs", "authors": "Waldek Mastykarz", "team": "Microsoft for Developers", "url": "https://developer.microsoft.com/blog/how-to-test-agent-skills-without-hitting-real-apis/", "fulltext_url": "https://developer.microsoft.com/blog/how-to-test-agent-skills-without-hitting-real-apis/", "type": "official_technical_article", "verification": "上游已核官方技术文章正文；本汇编未独立运行示例", "verified_at": "2026-10-22T07:10:59.255180+08:00", "evidence": "官方配置含getAll/getOne/merge/delete；无受控实验；750调用是成本假设算例。", "read_first": "A lighter-weight approach", "id": "url:https://developer.microsoft.com/blog/how-to-test-agent-skills-without-hitting-real-apis", "first_published": "2026-07-17", "verified_version": null, "limitations": "官方技术文章而非受控实验；无真实后端差分验证；750次调用为假设算例。", "appearances": [ "digests/agent-environments/2026/2026-10-22.md" ], "provenance": "ai-assisted-literature-notes" } ]