{
  "candidate_id": "ATT-0173",
  "reviewer": "D",
  "rubric": "PAPER-RUBRIC-1.0",
  "review_status": "frozen",
  "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
  },
  "analysis_date": "2026-11-08",
  "reading_record": "2026-11-08，先读legacy_reference_only.json，再逐页核对PDF 1–34页及附录A–I；主要以全文抽取阅读，另视觉核对第18、29–32页。开始时间未单独留存，完成/冻结时间见freeze_time_utc。",
  "freeze_time_utc": "2026-11-08T02:28:33.289111+00:00",
  "review_type": "independent_full_text_review",
  "independence": "未读取其他新评；先读取旧历史条目，再检查本版本全文和附录。",
  "legacy": {
    "full_review_available": false,
    "counts_as_review": false,
    "source_sha256": "11818c82f171596d63285db78f1546104af994cf86d2b1ec98788ea73999dad8",
    "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
      }
    ]
  },
  "claims": [
    "CL01：在声明的潜变量与有限阶假设下，长历史可降低生成过程不确定性，W>P是达到噪声下界的必要条件。证据：3–5、13–16页；理论支持较充分，真实机制解释部分支持。",
    "CL02：更异质的预训练分布增加识别负担，可能需要更长窗口和更大容量。证据：6–7、16–17、25页；支持部分，真实基础模型比较含混杂。",
    "CL03：分离可缓存上下文编码与短窗口预测能改善重复推理成本收益。证据：7–8、21–22、25–32页；在上下文可复用条件下支持，非普遍无损提速。"
  ],
  "review_body": "本文值得优先阅读的原因是它把长上下文的用途分成两个不同问题：近期观测决定在已知生成机制下如何预测，较长历史则帮助判断当前序列究竟来自哪一个机制。输入仍然是目标序列的过去观测，输出是未来数值；作者不是宣称发现一种普遍优胜的新基础模型，而是提出一个解释框架，并以可缓存的上下文编码器展示设计含义。旧记录只有“尚未评审、引用低于门槛”，没有可核对的完整旧评。这个历史条目不含质量证据，不能影响本次评分；当前引用量和会议身份也不代替正文论证。\n\nIdea方面，式5的全条件期望把预测写成对潜在过程参数的后验平均，数学工具本身是经典的贝叶斯分解，实质新增在于将它具体用于区分全局预测的GPI与CF负担，并把这一分解联到输入长度、模型容量和缓存设计。最接近的线索包括作者引用的Montero-Manso与Hyndman的局部/全局预测关系、贝叶斯式ICL解释，以及既有全局局部模型和超网络。本文明确承认这些联系，所以我不会把“上下文识别任务”称为从无到有的新思想。前作只核对本文引述与关系定位，未逐篇取得原始前作全文，因此原创性保持暂定。该框架的可迁移价值较高：它提示读者不要把长窗口收益直接解释成很长的时间依赖。\n\n理论部分的主要结论较窄但有用。定理3.2在固定参数的平稳有限阶加性噪声模型、最后P个值不能排除条件均值不同的潜在过程等假设下，证明W=P的最优均方误差仍含一个正的条件方差项，因此要达到噪声方差下界必须W>P。附录B的全方差分解成立；W<P一段写得较口语化，但可以利用条件信息减少不会降低最优平方风险补齐。这里“必要”不是“充分”，论文没有证明任意有限长窗口都能识别过程，更没有给出一般分布下的收缩速率。真实非平稳数据也未自动满足假设。另有一个需保留的报告问题：附录E.1式27写噪声为N(0,0.2)，随后明确第二参数为方差，而相关图中的噪声MSE下界接近0.04；AR示例图注写σ²=0.1而已知参数预测的RMSE接近0.1，也更像使用标准差0.1。此处需澄清符号与实际生成参数的一致性，不能据此断定全部实验无效。我已对第18页视觉核对，AR稳定筛选条件包含绝对值，文本抽取容易丢失，不应误判为稳定性公式错误。\n\nBaseline方面，合成数据用同类GRU比较局部、全局和附加远端上下文，并改变过程阶数、隐藏维数及训练域组合，直接服务于机制问题，比只报一张SOTA表更合适。真实数据覆盖交通、能源、温度和辐照，架构包括PatchTST、TSMixer、ModernTCN和LRU，基础模型探针包括Moirai2与TimerXL。作者按序列和时间双轴划分，独立保留未见序列，并用样本内和样本外验证损失平均进行早停，能够区分跨序列泛化与同序列未来预测。公平性仍有两处边界：基础模型与领域内训练PatchTST的差异同时包含架构、训练分布、规模和优化，不足以把全部差距归因为GPI；解耦模型多了上下文编码器且获得额外历史，固定预测器配置也不等于总参数或训练预算完全匹配。论文的结论应是设计可行性和条件性优势，而不是已完成所有强模型的等预算排名。\n\n消融与归因方面，附录D比较线性与PatchTST上下文编码器，附录I扩展P=4和6、隐藏宽度、训练域组合、归纳与直推设置，证据覆盖较实在。长间隔上下文仍有用说明即时邻接不是获得收益的必要条件，但固定两周间隔不能排除周期、共同季节状态和尺度统计等替代解释；“不是短期依赖”不能直接推出编码器识别了真实生成参数。作者也观察到拼接非连续输入使预测整体上移或下移。缺少跨序列错误上下文、打乱时间但保留统计量、概念漂移强弱等能更直接区分机制的材料，故机制证据高于普通删模块消融，却还不到排除全部竞争解释的程度。\n\n效率证据有实际内容，但使用条件非常重要。式8通过一次编码C长度历史和多次处理W长度近期数据，将重复读取成本从m(C+W)转为C+mW；这个收益来自历史缓存可复用，冷启动和频繁刷新时会缩小。正文归一化Pareto图易遮蔽数据集差异，附录第29至32页的逐架构逐数据图与表6应同时阅读。表6允许相对最佳标准模型最多10%的MAE损失，不能把其倍数写成无精度代价提速；PatchTST在ClmUS-R的CUDA倍数标准为4.87、解耦为2.10，部分组合并不优胜。训练估计总计约1800个A5000小时，单实验可达4小时，说明单卡可容纳不等于整个研究低成本。\n\n可复现性应分层评价。全文给出数据来源、划分、模型尺寸、优化器、早停、三次运行与标准差，多处合成线性实验有五次运行；CEREn需要申请，气象采样位置的具体清单和完整代码入口在本次材料中没有核实。多数实验24GB以内、部分12GB以内属于作者报告，不能替代实际复现，也不能推出每个结果在任何消费卡都能重建。我的建议是将本文保留为机制导向的优先阅读候选，重点借鉴GPI/CF分离及缓存成本核算；对“基础模型长窗口成本明确由GPI造成”的较强表述和未澄清的噪声参数保持保留。个人具体研究问题、知识基础和算力用途未提供，因此不代填F组，也不生成看似精确的综合总分。",
  "review_chinese_character_count": 1661,
  "dimensions": [
    {
      "dimension": "A1",
      "name": "问题重要性",
      "score": 85,
      "status": "rated",
      "confidence": "高",
      "reason": "长上下文成本与全局泛化是重要瓶颈",
      "evidence_pages": "1–3",
      "judgment_range": null
    },
    {
      "dimension": "A2",
      "name": "原创性",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "reason": "GPI/CF联结有新意，但贝叶斯分解和ICL前作核对有限",
      "evidence_pages": "3,9,22–23",
      "judgment_range": [
        60,
        80
      ]
    },
    {
      "dimension": "A3",
      "name": "实质推进程度",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "解释与解耦设计有明确推进，非通用SOTA证据",
      "evidence_pages": "5–8,25–32",
      "judgment_range": null
    },
    {
      "dimension": "A4",
      "name": "知识与解释增量",
      "score": 80,
      "status": "rated",
      "confidence": "中",
      "reason": "区分条件预测与过程不确定性，能指导架构",
      "evidence_pages": "3–7,13–17",
      "judgment_range": null
    },
    {
      "dimension": "B1",
      "name": "问题定义与评价有效性",
      "score": 80,
      "status": "rated",
      "confidence": "高",
      "reason": "风险定义、有限阶假设与归纳评测相对应",
      "evidence_pages": "2–5,19",
      "judgment_range": null
    },
    {
      "dimension": "B2",
      "name": "技术与论证正确性",
      "score": 75,
      "status": "tentative",
      "confidence": "中",
      "reason": "方差证明成立，噪声参数记法与图示需澄清",
      "evidence_pages": "13–18,24,33–34",
      "judgment_range": [
        65,
        85
      ]
    },
    {
      "dimension": "B3",
      "name": "数据与评测流程完整性",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "序列/时间双轴划分明确，预训练重叠未全面核验",
      "evidence_pages": "19,21",
      "judgment_range": null
    },
    {
      "dimension": "B4",
      "name": "比较对象与比较公平性",
      "score": 65,
      "status": "rated",
      "confidence": "中",
      "reason": "合成对照针对性强，真实模型训练和容量未等预算",
      "evidence_pages": "6–8,20–21",
      "judgment_range": null
    },
    {
      "dimension": "C1",
      "name": "核心主张的证据充分度",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "理论、合成与真实多层支持，因果范围有限",
      "evidence_pages": "3–8,13–34",
      "judgment_range": null
    },
    {
      "dimension": "C2",
      "name": "机制归因与竞争解释区分度",
      "score": 65,
      "status": "rated",
      "confidence": "中",
      "reason": "远端历史与容量对照有力，季节统计等解释未排除",
      "evidence_pages": "4–6,17,24–25",
      "judgment_range": null
    },
    {
      "dimension": "C3",
      "name": "结果稳定性与不确定性处理",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "三次均值标准差和部分五次结果，未给全面显著性及误差分解",
      "evidence_pages": "5–8,16–17,24–34",
      "judgment_range": null
    },
    {
      "dimension": "C4",
      "name": "适用范围与边界验证",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "多阶数、多域和多架构；漂移与概率预测边界薄弱",
      "evidence_pages": "9,18,24–32",
      "judgment_range": null
    },
    {
      "dimension": "D1",
      "name": "过程透明度",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "配置划分透明，但噪声记法及数据清单有缺口",
      "evidence_pages": "18–21",
      "judgment_range": null
    },
    {
      "dimension": "D2",
      "name": "研究资源可获得性",
      "score": 45,
      "status": "tentative",
      "confidence": "中",
      "reason": "正文附录可得，代码入口未核，CEREn受控申请",
      "evidence_pages": "18,21",
      "judgment_range": [
        35,
        65
      ]
    },
    {
      "dimension": "D3",
      "name": "主张克制与报告一致性",
      "score": 65,
      "status": "rated",
      "confidence": "中",
      "reason": "承认未训练新基础模型，但部分GPI归因和优势措辞偏强",
      "evidence_pages": "7–9,25–26",
      "judgment_range": null
    },
    {
      "dimension": "D4",
      "name": "表达清晰度",
      "score": 80,
      "status": "rated",
      "confidence": "高",
      "reason": "框架清楚，图表丰富，部分噪声与符号一致性需修正",
      "evidence_pages": "3–8,18",
      "judgment_range": null
    },
    {
      "dimension": "E1",
      "name": "复杂度合理性",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "独立上下文模块具有明确缓存目的，需考虑额外编码器",
      "evidence_pages": "7,20–22",
      "judgment_range": null
    },
    {
      "dimension": "E2",
      "name": "资源效率与成本收益",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "reason": "可复用上下文时收益可观，冷启动及刷新成本未充分量化",
      "evidence_pages": "21–22,25–32",
      "judgment_range": [
        60,
        80
      ]
    },
    {
      "dimension": "E3",
      "name": "可迁移与可复用程度",
      "score": 85,
      "status": "rated",
      "confidence": "中",
      "reason": "分离识别与预测的原则可迁移到多架构",
      "evidence_pages": "7,20,22–23",
      "judgment_range": null
    },
    {
      "dimension": "E4",
      "name": "后续研究启发性",
      "score": 85,
      "status": "rated",
      "confidence": "中",
      "reason": "漂移下识别、上下文选择和元数据条件化形成具体问题",
      "evidence_pages": "7–9,22–23",
      "judgment_range": null
    },
    {
      "dimension": "F1",
      "name": "与当前问题的相关性",
      "score": null,
      "status": "unknown",
      "confidence": "低",
      "reason": "具体研究问题未提供",
      "evidence_pages": "不适用页码：缺少用户背景",
      "judgment_range": null
    },
    {
      "dimension": "F2",
      "name": "相对于已有知识的学习收益",
      "score": null,
      "status": "unknown",
      "confidence": "低",
      "reason": "已有知识基础未提供",
      "evidence_pages": "不适用页码：缺少用户背景",
      "judgment_range": null
    },
    {
      "dimension": "F3",
      "name": "当前使用可行性",
      "score": null,
      "status": "unknown",
      "confidence": "低",
      "reason": "个人资源与用途未提供",
      "evidence_pages": "不适用页码：缺少用户背景",
      "judgment_range": null
    },
    {
      "dimension": "F4",
      "name": "当前决策信息价值",
      "score": null,
      "status": "unknown",
      "confidence": "低",
      "reason": "当前具体待决策事项未提供",
      "evidence_pages": "不适用页码：缺少用户背景",
      "judgment_range": null
    }
  ],
  "missing_and_limits": [
    "未取得完整旧评，仅历史待评行。",
    "未逐字对照会议终稿与arXiv；本评分只针对2606.01999v1。",
    "未独立检索并全文核对最近前作及后续复现，原创性暂定。",
    "未核验作者代码发布状态，不等于断言代码未公开。",
    "未执行训练、推理或代码复现；作者成本数字不等于实测。",
    "噪声方差/标准差记法和图示下界存在需澄清的一致性问题。"
  ],
  "experiments_executed": false,
  "aggregate_score": null
}