{
  "candidate_id": "ATT-0367",
  "reviewer": "D",
  "rubric": "PAPER-RUBRIC-1.0",
  "review_status": "frozen",
  "identity": {
    "candidate_id": "ATT-0367",
    "title": "TimeDistill: Efficient Long-Term Time Series Forecasting with MLP via Cross-Architecture Distillation",
    "authors": [
      "Ni, Juntong",
      "Liu, Zewen",
      "Wang, Shiyu",
      "Jin, Ming",
      "Jin, Wei"
    ],
    "arxiv_id": "2502.15016",
    "arxiv_version": "v3",
    "first_arxiv_date": "2025/02/20",
    "latest_arxiv_date": "2026/01/07",
    "source_urls": [
      "https://doi.org/10.1145/3770854.3780218"
    ],
    "publication_status": "KDD 2026, DOI 10.1145/3770854.3780218 verified in Crossref",
    "full_text_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0367/ATT-0367_paper.pdf",
    "metadata_source_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0367/ATT-0367_arxiv.html",
    "full_text_sha256": "8f7b3e2db65e4a21253330bb8720204dfcbd4838ad500dde23e2a53408721d67",
    "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，按旧记录→身份→正文1–9及附录11–27顺序阅读全文；参考文献9–10页核对相关条目；视觉核对11–12页证明。阅读期间未读他路新评；完成时间见冻结字段。",
  "freeze_time_utc": "2026-11-08T02:31:26.617065+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": "bbbe03056cb102aa89fca63eff545b2da995725fc0c5c756db4baca352a2a187",
    "records": []
  },
  "claims": [
    "CL01：跨架构多尺度/多周期输出与特征蒸馏提升轻量学生。多数实验支持，非所有教师/数据组合；4–8、14–25页。",
    "CL02：蒸馏目标可由特殊mixup理论解释，并解释泛化与优化收益。原A.2归一化和等式不成立，A.1未覆盖真实多尺度维度；5、11–12页。",
    "CL03：部署仅使用MLP而降低推理和参数开销。部署指标部分支持，训练总成本未知、参数表不一致；2、17、21页。"
  ],
  "review_body": "TimeDistill面向已有高精度预测器不便部署的情形，在训练阶段用较复杂的教师指导轻量学生，部署时只保留学生MLP。输入为多变量历史，输出为固定预测长度；主要学生采用通道独立、分解加两层MLP结构。它并非推理阶段融合多个大模型，而是把额外成本转移到离线教师训练和蒸馏。旧材料records为空，完整旧评和摘要均未提供，因此本次没有可以沿用的历史质量结论。评判对象限定为2026年1月7日arXiv v3，原创性时间基准应回到2025年2月首次公开，而不是因KDD 2026身份把思想视为2026新提出。\n\nIdea的可识别增量是同时在输出与中间特征上做多尺度和多周期匹配。前者对教师、学生预测逐层下采样并计算平方距离；后者对FFT幅值去直流后以低温softmax形成频率分布，再做KL匹配。不同架构通过回归器对齐特征宽度。这是一种有合理时序归纳偏置的蒸馏组合，而不是知识蒸馏本身的新发明。论文列出CAKD、LightCTS*和DE-TSMCL等近邻，并在附录M补了后两者比较，改善了与“普通KD”相比的定位。不过未逐篇核对前作原文，关于首次跨架构蒸馏的措辞不能由本次有限阅读确认。尤其潜在特征维的排列并不天然具有时间顺序，沿隐藏维FFT解释为物理周期，需要比输出频谱更谨慎。\n\n实证收益总体有可用价值。主表统一720历史长度，八数据集、四预测长度，TimeDistill多数平均指标优于未蒸馏MLP及多种强教师。ETTm2平均MSE从0.300到0.244，约18.6%的学生改善与表9对应；但不能把该最大值理解成所有任务平均提升。Traffic上iTransformer平均MSE0.379优于默认TimeDistill的0.387，Solar用PatchTST或Autoformer蒸馏可弱于原MLP，FITS在ETTm1也有小幅退化。附录给出这些负例是优点。选择教师、学生宽度、历史窗口和归一化都会改变结论，不能泛化成“任何教师都能让学生优于教师”。\n\nBaseline与协议方面，附录C明确同用Time-Series-Library，并设置drop_last=False保证测试样本一致；附录P还比较无归一化、non-stationary与RevIN，说明作者意识到预处理混杂。包含MLP、Transformer、CNN和其他KD参照，已不只是用弱旧模型衬托。仍需限制两个说法：固定720历史对所有模型相同，回答的是该输入协议下的表现，不代表各架构最佳部署点；教师用其默认训练配置，学生搜索蒸馏权重，整体调参和训练预算没有完整核平。归一化按验证选取是合理做法，但附录表15的主文措辞“始终优于”过强，例如无归一化Weather的TimeMixer略优。已读材料未显示测试泄漏，不能把未做代码审计写成发现泄漏。\n\n消融覆盖比一般模块拼接论文充分：表4及附录T分别去输出层、特征层、多尺度、多周期、监督项，换教师后也重复；附录I直接匹配真值模式，用于区分教师信息与仅改变损失。多尺度去除在ETT上造成明显损失，特征匹配通常也有益。然而默认ModernTCN下去多周期时ECL、ETT平均和Solar几乎不变，多周期的必要性依赖教师，不能仅以两个漂亮频谱例图证成普遍机制。附录T用教师到真值的KL差解释这种差异是有用线索，但单任务两个数字不足以确认因果解释。通道独立学生预测的相关矩阵接近教师，也不意味着它在推理时获得了跨通道信息输入；这只能说明输出统计被训练塑造，而不能突破其函数输入限制。\n\n理论部分存在比表述松散更实质的问题，必须单列。定理A.1把不同下采样长度的教师输出同未下采样学生直接做差，未明确上采样或统一空间；权重索引也与声称总和不一致。普通同维平方损失确实可导出加权目标平均的上界，但当前证明没有直接覆盖式4真正的多尺度损失。更明确的是定理A.2：混合目标写成Qy+ηQt，总质量为1+η，却作为概率分布进入KL；第12页log-sum之后分母本应为(1+η)Qs，下一行却写成对Qs的KL，等号不成立。取教师、学生和真值完全相同，监督误差与教师KL都为零，按文中所给求和定义右边却为(1+η)log(1+η)，η>0时立即反例。这不是要求跑实验才能确认的疑点。此外未乘温度平方、未说明中心化及FFT归一化就把高温KL极限直接等同MSE，且实际温度为0.5，也不能支撑其有限温度结论。因此“理论证明泛化改善和优化稳定”不能接受；工程蒸馏效果则应依据实验单独判断，不因理论错误自动归零。\n\n报告一致性也需扣分。表5中ETTh2 Win Keep为69.10%，正文却说全部超过76.61%；去监督的Traffic MSE0.506明显差于MLP0.434及教师0.413，与无监督仍优于两者的概括冲突。效率图2和表11同称ECL跨预测长度平均，但PatchTST参数前者约4.1M、后者约24.95M，MICN也有差异，当前版本未充分解释配置区别。这些不一致不证明作者动机，但会增加读者核查成本。\n\n可复现方面有五随机种子均值标准差、数据规模、训练轮数、学习率、权重搜索、批量及归一化配置，公开链接在PDF可解析为LingFengGold/TimeDistill。此次未拉取仓库或运行，不能说代码已验证。表11给出同批量16的推理时间和参数，学生与普通MLP同样约1.0956毫秒，支持推理结构不额外加重；但教师预训练、蒸馏前向与搜索开销未完整报告，不能把部署高效写成全生命周期低成本。总体建议作为轻量预测与KD基线保留，实验资料较丰富，理论部分不宜作为可靠引用，报告细节需要仔细复核。F组因用户具体问题、知识和资源不足保持未知。",
  "review_chinese_character_count": 1610,
  "dimensions": [
    {
      "dimension": "A1",
      "name": "问题重要性",
      "score": 80,
      "status": "rated",
      "confidence": "高",
      "reason": "轻量部署和准确率取舍实际重要",
      "evidence_pages": "1–2",
      "judgment_range": null
    },
    {
      "dimension": "A2",
      "name": "原创性",
      "score": 60,
      "status": "tentative",
      "confidence": "中",
      "reason": "时序模式KD组合有增量，首次性与近邻全文核对有限",
      "evidence_pages": "2–5,18",
      "judgment_range": [
        50,
        70
      ]
    },
    {
      "dimension": "A3",
      "name": "实质推进程度",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "多任务学生改善明确，但非所有教师和数据获益",
      "evidence_pages": "7,14–15,20",
      "judgment_range": null
    },
    {
      "dimension": "A4",
      "name": "知识与解释增量",
      "score": 50,
      "status": "rated",
      "confidence": "中",
      "reason": "教师模式与消融有启发，理论解释存在实质错误",
      "evidence_pages": "3–5,11–12,24–25",
      "judgment_range": null
    },
    {
      "dimension": "B1",
      "name": "问题定义与评价有效性",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "长期点预测与轻量部署目标基本匹配",
      "evidence_pages": "3,6–7,17",
      "judgment_range": null
    },
    {
      "dimension": "B2",
      "name": "技术与论证正确性",
      "score": 35,
      "status": "rated",
      "confidence": "高",
      "reason": "A.2归一化及KL推导有反例；算法训练目标仍可定义",
      "evidence_pages": "5,11–12",
      "judgment_range": null
    },
    {
      "dimension": "B3",
      "name": "数据与评测流程完整性",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "验证选归一化和统一测试样本明确，未静态审计全部处理",
      "evidence_pages": "12–13,18–19",
      "judgment_range": null
    },
    {
      "dimension": "B4",
      "name": "比较对象与比较公平性",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "同库同输入含KD参照，预算控制仍不充分",
      "evidence_pages": "6–7,13,18–19",
      "judgment_range": null
    },
    {
      "dimension": "C1",
      "name": "核心主张的证据充分度",
      "score": 60,
      "status": "rated",
      "confidence": "中",
      "reason": "性能支持较多，理论与部分机制主张不成立或不足",
      "evidence_pages": "7,11–12,14–25",
      "judgment_range": null
    },
    {
      "dimension": "C2",
      "name": "机制归因与竞争解释区分度",
      "score": 60,
      "status": "rated",
      "confidence": "中",
      "reason": "覆盖损失与真值替代，隐藏维频率及跨通道解释偏强",
      "evidence_pages": "8–9,16,24–25",
      "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": "五种子标准差明确，基线和细消融配对波动不足",
      "evidence_pages": "12,14",
      "judgment_range": null
    },
    {
      "dimension": "C4",
      "name": "适用范围与边界验证",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "多教师学生窗口及短期PEMS，负迁移边界有报告",
      "evidence_pages": "14–15,18–20,23",
      "judgment_range": null
    },
    {
      "dimension": "D1",
      "name": "过程透明度",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "训练配置详尽，但下采样及理论对应细节欠清",
      "evidence_pages": "4–5,11–13",
      "judgment_range": null
    },
    {
      "dimension": "D2",
      "name": "研究资源可获得性",
      "score": 65,
      "status": "tentative",
      "confidence": "中",
      "reason": "代码链接明确，未核验核心脚本及权重完整性",
      "evidence_pages": "1,12–13",
      "judgment_range": [
        50,
        80
      ]
    },
    {
      "dimension": "D3",
      "name": "主张克制与报告一致性",
      "score": 45,
      "status": "rated",
      "confidence": "高",
      "reason": "Win Keep、去监督概括和参数计数有内部矛盾",
      "evidence_pages": "2,7–8,17,19,24",
      "judgment_range": null
    },
    {
      "dimension": "D4",
      "name": "表达清晰度",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "流程图清楚，证明与版本内部数字增加理解负担",
      "evidence_pages": "4–5,11–12,17",
      "judgment_range": null
    },
    {
      "dimension": "E1",
      "name": "复杂度合理性",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "离线损失设计不增加推理路径，部分多周期贡献小",
      "evidence_pages": "4–7,24–25",
      "judgment_range": null
    },
    {
      "dimension": "E2",
      "name": "资源效率与成本收益",
      "score": 70,
      "status": "tentative",
      "confidence": "中",
      "reason": "同批量推理优势可信但全训练成本和参数差异未澄清",
      "evidence_pages": "2,12,17,21",
      "judgment_range": [
        55,
        80
      ]
    },
    {
      "dimension": "E3",
      "name": "可迁移与可复用程度",
      "score": 75,
      "status": "rated",
      "confidence": "中",
      "reason": "多个教师学生接口可复用，依赖隐藏表示匹配",
      "evidence_pages": "4–5,14–15,20",
      "judgment_range": null
    },
    {
      "dimension": "E4",
      "name": "后续研究启发性",
      "score": 70,
      "status": "rated",
      "confidence": "中",
      "reason": "教师选择负迁移与表示空间模式含义值得研究",
      "evidence_pages": "14–18,24–25",
      "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": "无：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": "无：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": "无：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": "无：缺少用户背景",
      "judgment_range": null
    }
  ],
  "missing_and_limits": [
    "旧records为空，完整旧评和历史摘要均缺失。",
    "未与ACM最终出版PDF逐字对照。",
    "代码链接明确但未检查仓库完整性或执行。",
    "近邻前作与后续复现外部检索未开展，原创性暂定。",
    "A.2数学问题已通过原PDF视觉复核；不把理论错误扩大为全部实证无效。",
    "有些参数数目及表间数字可能涉及未解释配置变化，未推测动机。"
  ],
  "experiments_executed": false,
  "aggregate_score": null
}