{
  "paper_id": "ATT-0173",
  "paper_type": [
    "实证分析",
    "理论",
    "方法设计"
  ],
  "reading_recommendation": "重点阅读，理论框架值得借鉴；经验因果归因与成本结论需保留条件。",
  "detailed_review_zh": "## 阅读判断\n这篇工作最值得读的不是又一种预测骨干，而是它把长历史的作用拆成“辨认当前是哪一个生成过程”和“给定过程后预测下一步”。对需要理解基础模型为何依赖长输入、为何小型领域模型有时反而更好的人，这一表述具有较高知识价值。A路建议重点阅读第3.1—3.2节及附录B，再定向查阅第4节和附录I.4。现有证据支持“长窗口可以服务于过程辨认”，尚不足以证明真实基础模型的长窗口收益主要或唯一由这一机制解释。首次公开时间按2026年6月1日计，不按会议海报或低引用量加减分。\n\n## 方法与新意\n输入是一组由不同参数产生的序列，模型参数跨序列共享，输出是未来数值。式(5)把最优条件均值写成按潜在过程后验加权的预测平均。这个恒等式使用标准条件期望法则，本身不应被当成全新的数学工具；真正的新增是把它明确用于全局时序预测的窗口需求，并从中推导可缓存历史表征的结构。作者在第5节、附录H承认与全局/局部模型、贝叶斯式上下文学习以及超网络有关。最接近的Montero-Manso与Hyndman关于全局模型窗口的工作以及贝叶斯ICL前作，本次仅核对本文对其关系的陈述及参考条目，未读取前作全文，因此原创性保留暂定，而不使用“首次解释”的绝对判断。\n\n第3.2节给定有限阶、平稳、加性独立噪声过程，并额外假定给定最近P点后，不同潜在过程的条件均值仍存在正的后验方差。附录B的方差分解直接显示P点预测风险严格大于噪声方差。这说明W>P是达到该下界的必要条件，不说明任何有限更大窗口一定够，也不提供后验收缩速度。W<P部分原文以一般性说明带过，但可以由信息嵌套下最优风险单调性补齐，所以我不把它判作定理失败。值得注意的是附录C式(23)把式(22)收集同一滞后项时的过程/系数下标交叉了；原图核对确认不是抽取错误。正确的第一系数应组合两个过程的第一系数，第二项同理。这更像局部代数排版错误，不能据此推翻正确的全方差论证，但会降低使用示例公式时的信任。\n\n## 证据与比较\n合成实验是本文最强的因果证据：真实阶数固定为2的NAR过程中，全局GRU使用更多历史才接近噪声下界，而加入隔开的历史段后，最近窗口只需达到阶数即可改善；附录I.2又改变阶数和隐藏维度。附录D比较局部/全局线性模型，并加上仅300个观察训练局部模型的控制，说明混合多个线性生成过程后的最优条件均值可以是非线性的。它比只展示一个真实数据榜单更能区分“长记忆”与“辨认过程”的作用。但附录E.1将噪声写为N(0,0.2)，紧接着明确第二参数是方差，而图3、图13等把不可约MSE画在0.04附近；附录C图8把sigma平方写0.1，已知过程预测RMSE却约0.1而非平方根0.1。这些一致指向标准差/方差记法混用，具体代码未读未运行，无法确认哪个记录正确。这里应保留为重要的报告与可重建性问题，不能把曲线坐标当成已核实的理论数值。\n\n真实数据实验使用交通、电力和气候序列，附录E.2明确同时按时间和序列划分，并用相互独立的留出序列验证和测试集。三次重复会改变序列划分和合成生成随机性，这些是明确优点。两种预训练模型在两周间隔之外仍从历史上下文受益，与过程辨认解释相容；然而两周间隔不是“没有任何相关信息”的证明，季节性、周期对齐、均值尺度估计都可以保留。不同预训练语料、骨干容量、优化训练量也使基础模型与领域PatchTST之间的差异不能全部归因于GPI。论文自身的受控NAR分析比该真实模型比较更适合支持机制结论。\n\n## 消融、成本与复现\n第4节把历史编码器G与短窗预测器F分离，固定历史表征能跨预测复用。四种F骨干、线性与PatchTST两种G，以及不同窗口、阶数和隐藏维度提供相当扎实的设计验证。它们仍未直接证明学习到的64维向量就是过程身份，也未充分量化分布漂移下何时重算缓存。附录G关于patch平移破坏缓存的论述适用于其指定逐时间步重切窗情形，并非所有流式patch方案都无缓存能力。\n\n成本结果必须按热缓存的重复调用读。第4节明确缓存固定时的处理量由每次C+W变成一次C加多次W，图7又分别按数据集和架构归一化；因此不能把聚合Pareto图解释成统一硬件预算下对所有模型的绝对统治。附录I.4保留分模型曲线和允许10% MAE损失的表6，很有帮助，也显示并非每个GPU时间或内存项都占优。表6里TSMixer两个不同数据集的解耦行四项数字完全相同，值得核对原始记录，但相同数字本身不足以认定结果错误。训练总费用报告约1800 A5000小时，单实验至多约4小时、24GB内；这不能简写成“整篇低成本可复现”。CEREn须申请，代码公开入口在本版全文未定位，数据版本与脚本完整性尚未独立核实。结论是过程和配置写得较详细，但可取得和可运行应分别保留未知。\n\n## 可带走的内容\n可迁移的认识是先问历史信息是否用于确定对象，再问它是否直接参与未来条件预测。若前者变化慢而后者更新快，分离编码与缓存就有具体理由。后续值得分析的是过程漂移时辨认信息的寿命，而不是单纯增加模块；若远历史经常代表已失效的过程，这种缓存设计的收益可能消失。用户具体研究问题、已有知识和资源未给出，F组保持未知，不用一般学术质量替代个人价值。旧文件只是“因低引用待处理”的摘要，没有完整旧评，不能视为负面质量证据。",
  "claims": [
    {
      "claim_id": "CL01",
      "claim": "最优全局预测可解释为潜在过程后验下的预测平均，某些有限阶过程需要超过阶数的输入",
      "claim_type": "理论/解释",
      "conditions": "平方损失、固定潜在过程、加性独立噪声及假设3.1",
      "support_status": "充分支持必要性，非充分性或速率",
      "location": "§3.1—3.2，式5—7，附录B pp13—14"
    },
    {
      "claim_id": "CL02",
      "claim": "真实基础模型的长上下文收益有过程辨认成分",
      "claim_type": "机制",
      "conditions": "Moirai2与TimerXL，四个小时级数据集，间隔336",
      "support_status": "部分支持；周期性等竞争解释未排除",
      "location": "图2、图5、附录F"
    },
    {
      "claim_id": "CL03",
      "claim": "分离GPI与CF并缓存可改善重复预测成本精度权衡",
      "claim_type": "效率",
      "conditions": "固定/缓变过程、热缓存、多次调用，四骨干",
      "support_status": "部分支持；不含完整冷启动与漂移刷新代价",
      "location": "§4，附录G、I.4，表6"
    }
  ],
  "risks": [
    {
      "risk_id": "R01",
      "severity": "重要问题",
      "certainty": "已确认记法和图值不一致；执行参数未知",
      "description": "N(0,0.2)被明确定义第二参数为方差，但不可约MSE图示约0.04；AR示例sigma平方0.1对应RMSE约0.1",
      "location": "p18式27、图3/13、p15图8",
      "question": "代码中noise参数是标准差还是方差，各图下界与生成参数的正确版本是什么？"
    },
    {
      "risk_id": "R02",
      "severity": "局部问题",
      "certainty": "已确认公式代数不一致",
      "description": "式23从式22收集同一滞后项时交叉了过程与系数下标；不影响全方差主定理",
      "location": "p15式22—23",
      "question": "是否应将两个过程的同一系数作后验加权？"
    },
    {
      "risk_id": "R03",
      "severity": "重要问题",
      "certainty": "证据归因局限",
      "description": "跨两周历史仍可能携带周期、尺度等预测信息；基础模型与领域模型训练条件不同",
      "location": "§3.1图2、§3.3图5",
      "question": "该证据支持GPI存在还是足以证明GPI主导？"
    },
    {
      "risk_id": "R04",
      "severity": "重要问题",
      "certainty": "成本条件明确、外推未证",
      "description": "热缓存收益未系统纳入冷启动、缓存刷新和漂移误差",
      "location": "§4图7；附录G、表6",
      "question": "在明确刷新频率下收益边界是什么？"
    },
    {
      "risk_id": "R05",
      "severity": "局部问题",
      "certainty": "重复值已观察，是否错误未知",
      "description": "TSMixer在CEREn与LargeST-SD解耦行四个成本改进值完全相同",
      "location": "p26表6",
      "question": "能否对应到各数据集的原始成本记录？"
    }
  ],
  "reading_record": {
    "pdf_pages": 34,
    "text_pages_read": "1—34（含参考文献与全部附录）",
    "visual_checked_pages": [
      15,
      18,
      26,
      27,
      29,
      30,
      31,
      32,
      33,
      34
    ],
    "legacy_read": "仅legacy_reference_only.json；旧评全文缺失",
    "theory_checked": "条件期望、全方差证明及AR2示例代数；未作形式化验证",
    "code_execution": "未执行",
    "external_predecessors": "仅本文相关工作与参考条目，未读取前作全文",
    "dataset_version": "CEREn SME subset；LargeST-SD；ClmUS POWER Daily 2.3.5 2023-02-26；Electricity/Traffic/Weather按附录E",
    "resource_check": "全文数据入口已读；代码实际可取得性未独立验证"
  },
  "rubric_version": "PAPER-RUBRIC-1.0",
  "reviewer_id": "A",
  "paper_title": "Why Do Time Series Models Need Long Context Windows?",
  "paper_version": "2606.01999v1",
  "first_public_date": "2026/06/01",
  "version_date": "2026/06/01",
  "assessment_date": "2026-11-08",
  "analysis_mode": "standard",
  "independence": {
    "other_new_reviews_read": false,
    "legacy_read_before_fulltext": true,
    "legacy_full_reviews_available": false,
    "legacy_records": [
      {
        "source_record_id": "S5-173",
        "source_id": "S5",
        "source_library_file_id": "libfile_849759d66a808191813dbe9bd1605bb4",
        "source_filename": "已粘贴的文本(5).txt",
        "source_line": 199,
        "source_number": 173,
        "title": "Why Do Time Series Models Need Long Context Windows?",
        "source_url": "https://neurips.cc/virtual/2026/poster/149819",
        "raw_line": "| P173 | [Why Do Time Series Models Need Long Context Windows?](https://neurips.cc/virtual/2026/poster/149819) | 0 | 待处理 | 尚未评审；引用较低（0 次，低于 10 次门槛） |",
        "raw_cells": [
          "P173",
          "[Why Do Time Series Models Need Long Context Windows?](https://neurips.cc/virtual/2026/poster/149819)",
          "0",
          "待处理",
          "尚未评审；引用较低（0 次，低于 10 次门槛）"
        ],
        "normalized_title": "whydotimeseriesmodelsneedlongcontextwindows",
        "normalized_url": "neurips.cc/virtual/2026/poster/149819",
        "record_type": "legacy_unranked",
        "citation_verbatim": "0",
        "legacy_status_verbatim": "待处理",
        "legacy_unranked_reason_verbatim": "尚未评审；引用较低（0 次，低于 10 次门槛）",
        "counts_as_independent_review": false
      }
    ],
    "legacy_counts_as_review": false
  },
  "identity": {
    "candidate_id": "ATT-0173",
    "title": "Why Do Time Series Models Need Long Context Windows?",
    "authors": [
      "Butera, Luca",
      "De Felice, Giovanni",
      "Cini, Andrea",
      "Alippi, Cesare"
    ],
    "arxiv_id": "2606.01999",
    "arxiv_version": "v1",
    "first_arxiv_date": "2026/06/01",
    "latest_arxiv_date": "2026/06/01",
    "source_urls": [
      "https://neurips.cc/virtual/2026/poster/149819"
    ],
    "publication_status": "NeurIPS 2026 official poster record; not yet conference proceedings",
    "full_text_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0173/ATT-0173_paper.pdf",
    "metadata_source_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0173/ATT-0173_arxiv.html",
    "full_text_sha256": "9226e890eaf0c0306caec014475d6c123b4185d6727e08da42b26b453351a0c8",
    "review_state": "ready_for_five_new_independent_reviews",
    "completed_new_reviews": 0,
    "source_versions_not_asserted_identical": true
  },
  "user_context": {
    "research_question": null,
    "prior_knowledge": null,
    "constraints": null
  },
  "aggregation": {
    "weights": null,
    "weighted_score_observed": null,
    "weighted_coverage": null,
    "tentative_weight_share": null,
    "notes": "保留24维原始评分，不预设总分；F组个人背景未知。"
  },
  "scores": [
    {
      "dimension_id": "A1",
      "dimension_name": "问题重要性",
      "score": 85,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "长窗成本和全局模型认识是重要瓶颈",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§1、§3"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "A2",
      "dimension_name": "原创性",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        60,
        80
      ],
      "rationale": "把经典贝叶斯分解用于窗口和缓存设计有新认识；最邻近前作全文未核",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§5、附录H"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "A3",
      "dimension_name": "实质推进程度",
      "score": 75,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        65,
        80
      ],
      "rationale": "合成理论与多骨干验证形成明显推进；真实模型归因与噪声记法待核",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "图3—7、附录B/I"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "A4",
      "dimension_name": "知识与解释增量",
      "score": 85,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "明确区分过程辨认与条件预测并产生可检验设计",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "式5—8、附录D"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "B1",
      "dimension_name": "问题定义与评价有效性",
      "score": 80,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "理论目标清晰，真实数据仅代理检验GPI",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§2—3"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "B2",
      "dimension_name": "技术与论证正确性",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        60,
        80
      ],
      "rationale": "主方差分解成立，示例下标与噪声尺度存在局部错误",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录B/C/E.1"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "B3",
      "dimension_name": "数据与评测流程完整性",
      "score": 75,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        65,
        85
      ],
      "rationale": "双轴留出和验证合理；预训练重叠与生成尺度未完全核对",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录E.2/F.1.3"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "B4",
      "dimension_name": "比较对象与比较公平性",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        60,
        80
      ],
      "rationale": "同骨干解耦对照有力；基础模型比较数据和容量不统一",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "图5—6、表4—5"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "C1",
      "dimension_name": "核心主张的证据充分度",
      "score": 75,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        65,
        80
      ],
      "rationale": "必要性理论强，真实主导机制和完整成本证据较弱",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录B、图2/7"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "C2",
      "dimension_name": "机制归因与竞争解释区分度",
      "score": 70,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "已知阶数及间隔实验有针对性，但真实周期性和参数容量仍混合",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "图3/4/11"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "C3",
      "dimension_name": "结果稳定性与不确定性处理",
      "score": 75,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "主要结果三次均值标准差，AR线性实验五次；无系统显著性分析",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "图3—6、图9—10、附录E.2"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "C4",
      "dimension_name": "适用范围与边界验证",
      "score": 70,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "多领域与多个阶数、骨干；漂移和长期缓存边界不足",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录I、§6"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "D1",
      "dimension_name": "过程透明度",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        60,
        80
      ],
      "rationale": "配置、划分、训练时间清楚，但噪声记法和计时细节需澄清",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录E/F、表6"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "D2",
      "dimension_name": "研究资源可获得性",
      "score": 50,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        35,
        70
      ],
      "rationale": "全文和数据入口可读，CEREn需申请，代码完整入口本次未确认",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录E/F"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "D3",
      "dimension_name": "主张克制与报告一致性",
      "score": 65,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "有局限章节，但真实GPI归因及聚合Pareto措辞较宽，数值记法不一致",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§3.3/4/6、式27"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "D4",
      "dimension_name": "表达清晰度",
      "score": 80,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "主线清楚，公式和实例相连；个别下标错误",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "式5—8、附录C"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "E1",
      "dimension_name": "复杂度合理性",
      "score": 80,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "双组件对应慢变身份与快变预测，复杂性有目的",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§4、附录G"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "E2",
      "dimension_name": "资源效率与成本收益",
      "score": 75,
      "status": "tentative",
      "confidence": "中",
      "judgment_range": [
        60,
        85
      ],
      "rationale": "热缓存的多骨干收益明确，训练1800小时与刷新成本须另算",
      "missing_information": "见理由及风险记录",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "附录F.2/I.4"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "E3",
      "dimension_name": "可迁移与可复用程度",
      "score": 85,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "GPI/CF区分可独立迁移到全局和自适应预测设计",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§3—4、附录H"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "E4",
      "dimension_name": "后续研究启发性",
      "score": 85,
      "status": "rated",
      "confidence": "高",
      "judgment_range": null,
      "rationale": "缓存寿命、漂移与辨认样本复杂度形成明确问题",
      "missing_information": null,
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "§6、附录G"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "F1",
      "dimension_name": "与当前问题的相关性",
      "score": null,
      "status": "unknown",
      "confidence": null,
      "judgment_range": null,
      "rationale": "缺少用户具体研究问题",
      "missing_information": "缺少用户具体研究问题",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "用户背景未提供"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "F2",
      "dimension_name": "相对于已有知识的学习收益",
      "score": null,
      "status": "unknown",
      "confidence": null,
      "judgment_range": null,
      "rationale": "缺少已有知识清单",
      "missing_information": "缺少已有知识清单",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "用户背景未提供"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "F3",
      "dimension_name": "当前使用可行性",
      "score": null,
      "status": "unknown",
      "confidence": null,
      "judgment_range": null,
      "rationale": "缺少用途、算力及数据权限",
      "missing_information": "缺少用途、算力及数据权限",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "用户背景未提供"
      ],
      "claim_ids": []
    },
    {
      "dimension_id": "F4",
      "dimension_name": "当前决策信息价值",
      "score": null,
      "status": "unknown",
      "confidence": null,
      "judgment_range": null,
      "rationale": "缺少当前具体选择",
      "missing_information": "缺少当前具体选择",
      "source_ids": [
        "S01"
      ],
      "evidence_locations": [
        "用户背景未提供"
      ],
      "claim_ids": []
    }
  ],
  "sources": [
    {
      "source_id": "S01",
      "source_type": "原论文全文及附录",
      "location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0173/ATT-0173_paper.pdf",
      "version_or_date": "2606.01999v1",
      "sha256": "9226e890eaf0c0306caec014475d6c123b4185d6727e08da42b26b453351a0c8",
      "access_status": "全文可读；未运行代码或模型"
    },
    {
      "source_id": "S02",
      "source_type": "历史摘要，非完整旧评",
      "location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0173/legacy_reference_only.json",
      "access_status": "先读；仅历史待核线索，不作新评票"
    }
  ],
  "review_chinese_char_count": 1768
}