{
"rubric_version": "PAPER-RUBRIC-1.0",
"paper_id": "ATT-0086",
"paper_title": "CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables",
"reviewer": "B",
"analysis_mode": "standard",
"paper_version": "2505.23116v1",
"version_date": "2025/05/29",
"first_public_date": "2025/05/29",
"assessment_date": "2026-11-08",
"reading_started_at": "2026-11-08T02:53:52.329294+00:00",
"reading_completed_at": "2026-11-08T02:58:55.062641+00:00",
"read_physical_pages": [
1,
2,
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13
],
"visual_checks_physical_pages": [
4,
8,
9,
12,
13
],
"pdf_sha256": "0235fd4cd7a51ae05b313532abdc7d0fe4dc86ac23bc29c1105f5e3cccb814f8",
"legacy_review_read_first": true,
"legacy_full_review_available": false,
"legacy_status": "仅有历史清单行或风险摘要；未提供完整旧评，不能声称复读完整旧评。",
"legacy_records": [
{
"source_record_id": "S5-086",
"source_id": "S5",
"source_library_file_id": "libfile_849759d66a808191813dbe9bd1605bb4",
"source_filename": "已粘贴的文本(5).txt",
"source_line": 28,
"source_number": 86,
"title": "CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables",
"source_url": "https://doi.org/10.1145/3711896.3736899",
"raw_line": "| 12 | [P086 CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables](https://doi.org/10.1145/3711896.3736899) | 59.2 | 100% | 55% | 53.7–65.0 | 18 | arXiv:2505.23116v1；本地 PDF；ACM 论文集版本未单独读取 | **重要问题**：主表及消融未报告重复运行次数、随机种子或误差区间；例如 ECL 多对一平均 MSE 的差距为 0.004，稳定性未知。；
**重要问题**：附录 B 给出 Sum 到 Cross Only 的卷积重参数化；单靠表 4 的结果无法区分参数化、初始化、训练动态和约束作用，也缺权重物理含义的独立验证。；
**重要问题**：插件在部分设置下出现退化：表 11 中 Traffic 96 步的 DLinear MSE 从 0.268 升至 0.338；ECL 的 DLinear 平均 MAE 从 0.457 升至 0.463。 | [评审](论文/P086/评审.md) · [评分数据](论文/P086/评审.json) · 无单独复核文件 |",
"raw_cells": [
"12",
"[P086 CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables](https://doi.org/10.1145/3711896.3736899)",
"59.2",
"100%",
"55%",
"53.7–65.0",
"18",
"arXiv:2505.23116v1；本地 PDF；ACM 论文集版本未单独读取",
"**重要问题**：主表及消融未报告重复运行次数、随机种子或误差区间；例如 ECL 多对一平均 MSE 的差距为 0.004，稳定性未知。；
**重要问题**：附录 B 给出 Sum 到 Cross Only 的卷积重参数化；单靠表 4 的结果无法区分参数化、初始化、训练动态和约束作用，也缺权重物理含义的独立验证。；
**重要问题**：插件在部分设置下出现退化：表 11 中 Traffic 96 步的 DLinear MSE 从 0.268 升至 0.338；ECL 的 DLinear 平均 MAE 从 0.457 升至 0.463。",
"[评审](论文/P086/评审.md) · [评分数据](论文/P086/评审.json) · 无单独复核文件"
],
"normalized_title": "crosslinearplugandplaycrosscorrelationembeddingfortimeseriesforecastingwithexogenousvariables",
"normalized_url": "doi.org/10.1145/3711896.3736899",
"record_type": "legacy_scored",
"legacy_rank": 12,
"legacy_score": "59.2",
"legacy_coverage": "100%",
"legacy_tentative_share": "55%",
"legacy_range": "53.7–65.0",
"citation_verbatim": "18",
"legacy_review_version": "arXiv:2505.23116v1；本地 PDF；ACM 论文集版本未单独读取",
"legacy_risks_verbatim": "**重要问题**：主表及消融未报告重复运行次数、随机种子或误差区间；例如 ECL 多对一平均 MSE 的差距为 0.004，稳定性未知。；
**重要问题**：附录 B 给出 Sum 到 Cross Only 的卷积重参数化；单靠表 4 的结果无法区分参数化、初始化、训练动态和约束作用，也缺权重物理含义的独立验证。；
**重要问题**：插件在部分设置下出现退化：表 11 中 Traffic 96 步的 DLinear MSE 从 0.268 升至 0.338；ECL 的 DLinear 平均 MAE 从 0.457 升至 0.463。",
"legacy_material_links_verbatim": "[评审](论文/P086/评审.md) · [评分数据](论文/P086/评审.json) · 无单独复核文件",
"counts_as_independent_review": false
}
],
"new_peer_reviews_read": false,
"review_text": "先读可得历史摘要，再完整阅读2505.23116v1的十三页，包括附录A—E；旧完整评语缺失，不将历史总分或意见当成新判断。CrossLinear的核心是把目标变量和历史协变量堆叠，经过单层一维卷积后与目标自身作可学习加权，再进入分块线性映射及预测头。这是实用而低门槛的设计，尤其适合已有通道独立预测器需要借助其他变量的情形。需要区分三个命题：简单混合层能改善一些预测器、残差参数化有训练优势、卷积权重真实识别直接变量依赖。第一项有相当实验支持，第二项有现象但归因尚弱，第三项没有得到识别意义上的验证。\n\n第4页的符号明确输入只有历史内生与外生变量，不包括已知未来的日历、天气预报或价格驱动量。因此这里的exogenous主要是预测协变量，不能自动等同于计量经济学的外生性，更不能由性能推得因果关系。归一化固定后，所描述的卷积、残差、patch投影、位置向量与线性头构成仿射映射；RevIN统计量依赖样本，所以不宜把整个原始输入到输出都称为单一固定线性矩阵。该设计的价值更可能是参数共享、局部约束和训练偏置，而不是新的非线性表达族。位置编码是与数据无关的加性项，flatten后的预测头本已区分位置，其作用不能仅凭常见Transformer类比就认定为不可替代的时间次序建模。\n\n附录第12—13页的Sum与Cross Only等价证明是一个优点。在匹配padding、可容纳单位选择核、妥善处理偏置的条件下，K′=(1−α)K+αS确能吸收残差。它证明两个表达族的重参数化关系，不是同一组数值权重产生相同结果，更不是两个训练程序必然等价。因此Sum在ECL平均MSE为0.323而Cross Only为0.396，不能解释为Sum获得了后者原则上无法表示的额外信息。初始化、有限迭代、正则化和优化路径都可产生差距。作者第13页也承认重参数化与有限数据因素，应保留这份克制，而不将正文“防止学到不恰当表示”扩展成唯一机制。\n\n基线覆盖TimeXer、TiDE、多种Transformer及线性模型，较只对弱模型有明显优势；十二个数据集和两种输出范式也具有广度。不过接近最佳的许多差距很小：外生预测ECL均值0.323对TimeXer的0.327，ETTh1为0.072对0.073，且ECL的720步本作0.403反而差于0.365。EPF平均MSE优于TimeXer，但MAE为0.266，略差于0.265；多变量Traffic也不及iTransformer。统一历史长度96、十轮训练的设置未证明各模型在等预算下接近其最佳状态。附录说所有基线遵循官方配置，与统一轮数和学习率搜索的关系仍不完全清楚。没有随机种子次数、方差或配对区间，无法判断千分位排序的稳定性。Weather异常值与归一化导致很小误差虽有作者说明，不能把原始汇总指标自动当成已经验证的异常鲁棒性。\n\n消融比仅删除模块丰富：Endo Only、Cross Only、Concat以及α、β、窗口长度、kernel和隐藏维度都有结果。ECL去掉其他变量从0.323升到0.356，支持协变量有用；ETTh1从0.072到0.073及Traffic从0.152到0.155，则显示收益高度依赖数据。Concat与Sum的差距也混有不同输入宽度及参数预算。五种既有模型上的插件结果是最有可复用价值的部分，但不可概括成每种组合都提升。DLinear的Traffic在96步由0.268变为0.338，ECL平均MAE也由报告的0.457变成0.463。\n\n本次还逐项核对了比历史风险更重要的汇总算术问题。第12页表11中，DLinear原始Traffic四行MSE为0.268、0.302、0.298、0.340，算术平均是0.302，而表列0.323；四行MAE平均0.3845，而表列0.404。这两个不一致也见第6页主表。加入模块后的四行均值约0.308、0.391，和其汇总一致。于是按已印分项重算，原表宣称的两个平均改善方向都会反转。不能据此断定真实实验一定变差，因为也可能分项抄错；但“该模型该数据集平均改善”的报告目前不可核证。应将这一特定结论暂挂，而不是把所有主结果一并判无效。表标题多处把按预测长度求均值写成跨数据集，也降低了汇总口径透明度。\n\n随机缺失实验覆盖内生、外生、50%和100%、置零与噪声替代，较只报完整数据更有意义。但没有相同掩码的其他模型或专门缺失处理基线，不能证明相对鲁棒性领先。掩码是在归一化前还是后、训练阶段是否同样缺失等细节未充分交代；尤其目标全部掩掉时，是否保留来自完整目标的均值和方差会改变问题定义，当前只能标为未核，而不能直接指控泄漏。表6的NP标注96→96，与正文EPF的168→24不一致，需要区分表头错误和不同设置。ETTh1所有外生信息置零后MSE仍为0.055，也不支持“所有数据都显著恶化”的概括。\n\n解释性图是参数可视化，不是依赖图验证。多变量卷积核按滞后求和可能抵消正负作用，变量间相关及后续线性层也使单层权重不具有唯一归因；与气象常识一致只能提供表面合理性。效率图在ECL、同批量下报告每轮训练时间，有帮助但不是推理延迟。O(T)成立需固定变量数、kernel、隐藏宽度及输出长度，多变量卷积实际还含N平方因子，不能推得高维变量下成本近乎不变。没有端到端调参成本或边缘设备实测。\n\n官方README可读，提供Python版本、依赖、数据入口及两类任务脚本；论文另给Zenodo入口，本次没有核验该存档内容，也没有运行模型。整体上，这是值得纳入强简单基线的经验模块，而不是已证明直接依赖识别的理论方法。值得保留的问题是残差收益能否主要由同表达族的训练参数化解释，以及插件平均改善在纠正表格口径后还剩多少。优先亲读第4页结构与第12—13页等价证明，再交叉看第6、12页原始行和平均值；这是决定如何使用现有结论的证据判断，不向用户布置实验。",
"review_hanzi_count": 1704,
"claims": [
{
"id": "CL01",
"claim": "简单协变量混合在许多任务具竞争力",
"status": "supported_with_scope",
"pages": [
6,
7,
8
]
},
{
"id": "CL02",
"claim": "Sum和Cross Only可通过卷积核重参数化表达相同映射",
"status": "supported_with_conditions",
"pages": [
12,
13
]
},
{
"id": "CL03",
"claim": "DLinear Traffic插件平均MSE和MAE均改善",
"status": "suspended_due_to_internal_arithmetic_inconsistency",
"pages": [
6,
8,
12
]
},
{
"id": "CL04",
"claim": "静态卷积权重识别直接变量依赖",
"status": "not_established",
"pages": [
9
]
}
],
"scores": [
{
"dimension_id": "A1",
"dimension_name": "问题重要性",
"score": 85,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "低成本利用历史协变量是重要且常见的预测问题。",
"evidence_pages": [
1,
3,
4
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "A2",
"dimension_name": "原创性",
"score": 60,
"status": "tentative",
"confidence": "中",
"judgment_range": [
50,
70
],
"rationale": "卷积、残差、patch和线性头组合有实用新意，非全新表达族。",
"evidence_pages": [
3,
4,
13
],
"missing_information": "见理由与材料边界",
"source_ids": [
"S01"
]
},
{
"dimension_id": "A3",
"dimension_name": "实质推进程度",
"score": 70,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "广泛任务上竞争力好，插件复用价值高，但细小优势和错误平均限制增量。",
"evidence_pages": [
6,
7,
8,
12
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "A4",
"dimension_name": "知识与解释增量",
"score": 60,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "等价证明有解释价值，权重即直接依赖与训练机制未识别。",
"evidence_pages": [
9,
12,
13
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "B1",
"dimension_name": "问题定义与评价有效性",
"score": 75,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "历史协变量与输出范式明确；缺失归一化顺序及NP设置不清。",
"evidence_pages": [
3,
4,
5,
8
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "B2",
"dimension_name": "技术与论证正确性",
"score": 75,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "核重参数化论证条件下正确；静态权重与直接依赖的联系不足。",
"evidence_pages": [
4,
9,
12,
13
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "B3",
"dimension_name": "数据与评测流程完整性",
"score": 65,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "划分、输入长度与超参范围有披露，重复和缺失流程细节不足。",
"evidence_pages": [
5,
8,
11
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "B4",
"dimension_name": "比较对象与比较公平性",
"score": 70,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "强外生模型和多类基线齐备，但等搜索预算与10轮配置未充分对齐。",
"evidence_pages": [
5,
6,
11
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "C1",
"dimension_name": "核心主张的证据充分度",
"score": 65,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "支持多任务有效，部分平均算术与小幅无方差优势不支持广泛优越性。",
"evidence_pages": [
6,
7,
8,
12
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "C2",
"dimension_name": "机制归因与竞争解释区分度",
"score": 60,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "多种消融有价值，表达族等价仍容许优化、初始化与容量解释。",
"evidence_pages": [
7,
9,
11,
13
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "C3",
"dimension_name": "结果稳定性与不确定性处理",
"score": 35,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "未见重复次数、标准差或区间，千分位差距稳定性未知。",
"evidence_pages": [
6,
7,
11,
12
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "C4",
"dimension_name": "适用范围与边界验证",
"score": 75,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "12数据集、多对一/多对多、缺失测试覆盖较广，真实部署边界仍弱。",
"evidence_pages": [
5,
6,
7,
8
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "D1",
"dimension_name": "过程透明度",
"score": 60,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "附录完整表格和硬件配置可用，掩码顺序、NP表头及均值口径不足。",
"evidence_pages": [
8,
11,
12
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "D2",
"dimension_name": "研究资源可获得性",
"score": 80,
"status": "tentative",
"confidence": "中",
"judgment_range": [
70,
85
],
"rationale": "官方README、环境及脚本可访问；存档、完整代码和执行复现未核。",
"evidence_pages": [
1
],
"missing_information": "见理由与材料边界",
"source_ids": [
"S01"
]
},
{
"dimension_id": "D3",
"dimension_name": "主张克制与报告一致性",
"score": 50,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "报告允许重参数化解释，但部分插件平均改善因算术冲突暂不成立。",
"evidence_pages": [
6,
8,
9,
12,
13
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "D4",
"dimension_name": "表达清晰度",
"score": 75,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "结构和证明清楚，表头、混合轴表述及平均标签影响精确阅读。",
"evidence_pages": [
4,
8,
12,
13
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "E1",
"dimension_name": "复杂度合理性",
"score": 85,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "单层混合与分块线性头简单，额外机制较少。",
"evidence_pages": [
4,
5
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "E2",
"dimension_name": "资源效率与成本收益",
"score": 65,
"status": "tentative",
"confidence": "中",
"judgment_range": [
50,
75
],
"rationale": "同批量ECL每轮时间支持局部效率，O(T)固定其他维度且未给推理全成本。",
"evidence_pages": [
9
],
"missing_information": "见理由与材料边界",
"source_ids": [
"S01"
]
},
{
"dimension_id": "E3",
"dimension_name": "可迁移与可复用程度",
"score": 85,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "可插件化接入五种模型，两输出范式与公开入口降低复用门槛。",
"evidence_pages": [
4,
8,
12
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "E4",
"dimension_name": "后续研究启发性",
"score": 80,
"status": "rated",
"confidence": "高",
"judgment_range": null,
"rationale": "同表达族不同训练参数化、变量混合与可解释性边界值得继续理解。",
"evidence_pages": [
9,
12,
13
],
"missing_information": "",
"source_ids": [
"S01"
]
},
{
"dimension_id": "F1",
"dimension_name": "与当前问题的相关性",
"score": null,
"status": "unknown",
"confidence": "低",
"judgment_range": null,
"rationale": "未给出用户具体研究问题、已有知识、资源与当前决策，不能从论文的一般价值推断。",
"evidence_pages": [],
"missing_information": "个人背景未提供",
"source_ids": []
},
{
"dimension_id": "F2",
"dimension_name": "相对于已有知识的学习收益",
"score": null,
"status": "unknown",
"confidence": "低",
"judgment_range": null,
"rationale": "未给出用户具体研究问题、已有知识、资源与当前决策，不能从论文的一般价值推断。",
"evidence_pages": [],
"missing_information": "个人背景未提供",
"source_ids": []
},
{
"dimension_id": "F3",
"dimension_name": "当前使用可行性",
"score": null,
"status": "unknown",
"confidence": "低",
"judgment_range": null,
"rationale": "未给出用户具体研究问题、已有知识、资源与当前决策，不能从论文的一般价值推断。",
"evidence_pages": [],
"missing_information": "个人背景未提供",
"source_ids": []
},
{
"dimension_id": "F4",
"dimension_name": "当前决策信息价值",
"score": null,
"status": "unknown",
"confidence": "低",
"judgment_range": null,
"rationale": "未给出用户具体研究问题、已有知识、资源与当前决策，不能从论文的一般价值推断。",
"evidence_pages": [],
"missing_information": "个人背景未提供",
"source_ids": []
}
],
"risks": [
{
"severity": "major",
"scope": "CL03",
"issue": "DLinear Traffic原始行均值MSE0.302、MAE0.3845不等于所列0.323、0.404，改变插件平均改善方向。",
"pages": [
6,
8,
12
],
"action": "暂挂该平均改善，不推断原始行或均值哪一个是真实实验。"
},
{
"severity": "moderate",
"scope": "causal_interpretation",
"issue": "核权重不是可识别的依赖强度；等价表达族消融不能唯一分离机制。",
"pages": [
9,
13
],
"action": "限为预测混合和训练参数化现象。"
},
{
"severity": "moderate",
"scope": "reproduction",
"issue": "无种子方差；掩码与RevIN先后、NP预测设置不明。",
"pages": [
8,
11
],
"action": "保留统计和流程未知。"
}
],
"aggregation": {
"weights": null,
"weighted_score_observed": null,
"weighted_coverage": null,
"tentative_weight_share": null,
"notes": "保留24维原始分，不擅自合成总分；未知不作0。"
},
"sources": [
{
"source_id": "S01",
"source_type": "arxiv PDF",
"location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0086/ATT-0086_paper.pdf",
"version": "2505.23116v1"
},
{
"source_id": "S02",
"source_type": "待核历史记录",
"location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0086/legacy_reference_only.json"
}
],
"limits": {
"no_model_execution": true,
"no_training_or_inference": true,
"no_full_code_audit": true,
"external_closest_prior_full_text_check": "本次未独立取得并阅读全文，原创性为暂定判断。",
"author_result_not_independent_reproduction": true
},
"resource_read": {
"url": "https://github.com/mumiao2000/CrossLinear/blob/main/README.md",
"blob_sha": "31d276de7510b04c0585b135f2dbf68f0350d161",
"notes": "已读环境、数据入口与两类任务脚本说明；未执行、未核验Zenodo存档。"
},
"arithmetic_checks": [
{
"table": 11,
"physical_page": 12,
"model": "DLinear",
"dataset": "Traffic",
"metric": "MSE",
"original_rows": [
0.268,
0.302,
0.298,
0.34
],
"recomputed_mean": 0.302,
"reported_mean": 0.323,
"embedding_rows": [
0.338,
0.26,
0.285,
0.349
],
"embedding_recomputed_mean": 0.308
},
{
"table": 11,
"physical_page": 12,
"model": "DLinear",
"dataset": "Traffic",
"metric": "MAE",
"original_rows": [
0.351,
0.387,
0.384,
0.416
],
"recomputed_mean": 0.3845,
"reported_mean": 0.404,
"embedding_rows": [
0.414,
0.356,
0.378,
0.417
],
"embedding_recomputed_mean": 0.39125
}
]
}