{ "candidate_id": "ATT-0594", "reviewer": "C", "rubric": "PAPER-RUBRIC-1.0", "frozen": true, "frozen_at": "2026-11-08T02:34:22.728597+00:00", "identity": { "candidate_id": "ATT-0594", "title": "ChronosX: Adapting Pretrained Time Series Models with Exogenous Variables", "authors": [ "Arango, Sebastian Pineda", "Mercado, Pedro", "Kapoor, Shubham", "Ansari, Abdul Fatir", "Stella, Lorenzo", "Shen, Huibin", "Senetaire, Hugo", "Turkmen, Caner", "Shchur, Oleksandr", "Maddix, Danielle C.", "Bohlke-Schneider, Michael", "Wang, Yuyang", "Rangapuram, Syama Sundar" ], "arxiv_id": "2503.12107", "arxiv_version": "v1", "first_arxiv_date": "2025/03/15", "latest_arxiv_date": "2025/03/15", "source_urls": [ "https://proceedings.mlr.press/v258/arango25a.html" ], "publication_status": "AISTATS 2025, PMLR v258 arango25a verified", "full_text_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0594/ATT-0594_paper.pdf", "metadata_source_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0594/ATT-0594_arxiv.html", "full_text_sha256": "cf934ba841056d2119d61ced731a0ac5966f584ec36146ce46d8ad9b97f69855", "review_state": "ready_for_five_new_independent_reviews", "completed_new_reviews": 0, "source_versions_not_asserted_identical": true }, "source_sha256_verified": "cf934ba841056d2119d61ced731a0ac5966f584ec36146ce46d8ad9b97f69855", "reading_started_at": "2026-11-08T02:30:24Z", "reading_finished_at": "2026-11-08T02:34:22.728597+00:00", "reading_time_note": "以UTC记录开始和完成时刻", "read_coverage": { "pdf_pages": 53, "scope": "正文、参考文献及全部随附附录的逐页文本；关键图式按需要视觉复核", "visual_pages": [ 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53 ] }, "legacy_read_first": true, "legacy_full_reviews_missing": true, "other_new_reviews_read": false, "experiments_run": false, "external_predecessor_fulltext_checked": false, "resource_notes": "本次未执行代码、训练或推理；未将仓库入口存在等同资源完整或可运行。", "claims": [ { "claim": "轻量协变量接口可改善多类预训练主干", "judgment": "支持，限目标任务训练与给定未来信息条件", "pages": "4–9,20–24" }, { "claim": "冻结适配总体具有竞争力", "judgment": "WQL支持，MASE及逐数据集并非全胜", "pages": "6–8,23–24" }, { "claim": "适配成本低于全微调", "judgment": "现有训练加测试时长及更新参数量支持，不能外推全部部署成本", "pages": "19" } ], "risks": [ { "risk": "真实未来协变量可得性未逐项说明", "classification": "证据缺口，不等于已证实泄漏" }, { "risk": "种子均值缺区间、调参预算不齐", "classification": "比较和稳定性限制" }, { "risk": "完整旧评缺失", "classification": "历史证据缺口" } ], "scores": [ { "dimension_id": "A1", "dimension_name": "问题重要性", "score": 85, "status": "scored", "confidence": "高", "judgment_range": "80–90", "reason": "外生信息接入是预训练预测模型的重要实际缺口。", "evidence_pages": "1–3" }, { "dimension_id": "A2", "dimension_name": "原创性", "score": 65, "status": "tentative", "confidence": "中", "judgment_range": "55–75", "reason": "创新集中于通用接口及实证，残差适配本身成熟，未全文核验最近前作。", "evidence_pages": "2–5" }, { "dimension_id": "A3", "dimension_name": "实质推进程度", "score": 75, "status": "scored", "confidence": "高", "judgment_range": "70–80", "reason": "三类主干和真实数据改善具有实际推进，未普遍超过专用模型。", "evidence_pages": "6–9,20–24" }, { "dimension_id": "A4", "dimension_name": "知识与解释增量", "score": 70, "status": "scored", "confidence": "中", "judgment_range": "65–75", "reason": "揭示冻结、注入位置和模型规模的交互，但机制解释有限。", "evidence_pages": "7–9,25–36" }, { "dimension_id": "B1", "dimension_name": "问题定义与评价有效性", "score": 80, "status": "scored", "confidence": "高", "judgment_range": "75–85", "reason": "任务和多指标明确，几何平均与逐集结果可互查，未来输入可得性仍需补充。", "evidence_pages": "3,5–7,17" }, { "dimension_id": "B2", "dimension_name": "技术与论证正确性", "score": 80, "status": "scored", "confidence": "中", "judgment_range": "75–85", "reason": "结构和接口可理解，未见关键公式错误，未经代码验证。", "evidence_pages": "3–5,13" }, { "dimension_id": "B3", "dimension_name": "数据与评测流程完整性", "score": 60, "status": "tentative", "confidence": "中", "judgment_range": "50–70", "reason": "真实未来协变量发布时点、预训练重叠未充分审计，不能据此断言泄漏。", "evidence_pages": "5–6,17–18" }, { "dimension_id": "B4", "dimension_name": "比较对象与比较公平性", "score": 65, "status": "scored", "confidence": "高", "judgment_range": "60–70", "reason": "基线丰富但调参预算、预训练数据及概率输出能力不完全一致。", "evidence_pages": "5–7,18" }, { "dimension_id": "C1", "dimension_name": "核心主张的证据充分度", "score": 80, "status": "scored", "confidence": "高", "judgment_range": "75–85", "reason": "广泛合成和真实任务支持有效适配，全面优越性不成立。", "evidence_pages": "6–9,20–36" }, { "dimension_id": "C2", "dimension_name": "机制归因与竞争解释区分度", "score": 75, "status": "scored", "confidence": "高", "judgment_range": "70–80", "reason": "无协变量和接口消融有针对性，仍缺等参数随机信息对照。", "evidence_pages": "7–9,25–30,34–36" }, { "dimension_id": "C3", "dimension_name": "结果稳定性与不确定性处理", "score": 55, "status": "scored", "confidence": "高", "judgment_range": "50–60", "reason": "三种子均值有报告，多数缺方差及配对区间，预测区间不能替代实验误差条。", "evidence_pages": "18,20–36,37–53" }, { "dimension_id": "C4", "dimension_name": "适用范围与边界验证", "score": 70, "status": "scored", "confidence": "高", "judgment_range": "65–75", "reason": "多主干、多规模、多数据，但缺未来协变量误差、缺失与延迟测试。", "evidence_pages": "6–9,15–18,31–36" }, { "dimension_id": "D1", "dimension_name": "过程透明度", "score": 75, "status": "scored", "confidence": "中", "judgment_range": "70–80", "reason": "主要设置明确，未来信息构造和确定性基线分位数处理需补充。", "evidence_pages": "13,15–18" }, { "dimension_id": "D2", "dimension_name": "研究资源可获得性", "score": 65, "status": "tentative", "confidence": "中", "judgment_range": "55–75", "reason": "有官方代码分支入口，未核验快照、依赖及可运行性。", "evidence_pages": "1,11,13" }, { "dimension_id": "D3", "dimension_name": "主张克制与报告一致性", "score": 75, "status": "scored", "confidence": "高", "judgment_range": "70–80", "reason": "承认零样本限制，整体措辞仍需区分指标最佳和逐任务一致提升。", "evidence_pages": "6–9" }, { "dimension_id": "D4", "dimension_name": "表达清晰度", "score": 80, "status": "scored", "confidence": "高", "judgment_range": "75–85", "reason": "图示和全量表格清晰，附录充分但旋转表阅读成本高。", "evidence_pages": "4,6–9,20–36" }, { "dimension_id": "E1", "dimension_name": "复杂度合理性", "score": 80, "status": "scored", "confidence": "高", "judgment_range": "75–85", "reason": "模块轻量且可插接，单侧模块有时更优，配置仍应按任务裁剪。", "evidence_pages": "4–5,25–27" }, { "dimension_id": "E2", "dimension_name": "资源效率与成本收益", "score": 70, "status": "tentative", "confidence": "中", "judgment_range": "60–80", "reason": "适配训练代价较低，图测训练加测试而非独立在线延迟，主干存储仍保留。", "evidence_pages": "13,19" }, { "dimension_id": "E3", "dimension_name": "可迁移与可复用程度", "score": 85, "status": "scored", "confidence": "高", "judgment_range": "80–90", "reason": "三类模型接口和官方实现入口便于后续移植。", "evidence_pages": "4–5,13" }, { "dimension_id": "E4", "dimension_name": "后续研究启发性", "score": 75, "status": "scored", "confidence": "高", "judgment_range": "70–80", "reason": "未来输入可靠性和冻结微调边界有明确研究空间。", "evidence_pages": "7–9" }, { "dimension_id": "F1", "dimension_name": "与当前问题的相关性", "score": null, "status": "unknown", "confidence": "低", "judgment_range": "无", "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" }, { "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": "无", "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" }, { "dimension_id": "F3", "dimension_name": "当前使用可行性", "score": null, "status": "unknown", "confidence": "低", "judgment_range": "无", "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" }, { "dimension_id": "F4", "dimension_name": "当前决策信息价值", "score": null, "status": "unknown", "confidence": "低", "judgment_range": "无", "reason": "未提供具体当前研究目标、知识背景、资源和决策，不以通用价值替代个人评价。", "evidence_pages": "不适用页码：用户背景缺失" } ], "body_zh": "## 独立详评\n\n本轮先查看旧材料入口，未见可重读的完整历史评语，因而不把历史筛选顺序当成质量证据。正文及随附附录共五十三页均已阅读。ChronosX 的研究问题很实际：预训练时间序列模型往往只接收目标变量，而真实业务预测需要节假日、价格、天气或其他外生信息。作者通过输入注入块和输出注入块，在保留预训练主干的条件下引入历史及未来协变量，并把方法扩展到 Chronos、TimesFM 和 MOMENT 三类不同主干。其价值在于可复用的接入方式和系统实验，不在于提出新的基础模型或完全免训练的预测器。适配器本身仍须在目标任务上学习，作者第九页也明确承认失去原模型的零样本属性。\n\n技术结构清楚：输入侧在嵌入空间做残差修正，输出侧结合解码器状态、预测输出及协变量做残差修正。对于编码器、解码器和编码器解码器模型，接口的对应关系也有说明。主干冻结与全参数微调均被测试。这个设计没有明显的公式级矛盾，但以现有论文引用为限，尚不能断言其相对所有条件化适配器的原理创新程度。原始预测器加轻量条件模块是成熟思路，因此原创性主要落在时间序列预训练模型中的具体实现、通用接口和实验证据，而非残差适配本身。\n\n证据规模是优点：三十二种合成设置由四种信号、四种外生过程和加法或乘法关系组合，另有十八个真实数据集，覆盖能源、零售及出行。真实任务的归一化几何平均加权分位数损失为零点八七八，相对 Chronos Small 有明确提升，而且优于表中其他方法的该项总体指标。但不能将此解释为全部指标或每个数据集都最好。其总体 MASE 为零点九一一，落后于 NBEATSx、NHiTS 和 PatchTSTx；M5、部分电力与 ETT 数据也出现不如原始主干的结果。合成任务中，TFT 或 PatchTSTx 在若干组别明显更好，说明从头训练的专用模型仍有竞争力。\n\n机制消融相对扎实。无协变量但保留适配器的对照，有助于区分新增参数与新增信息的作用；仅隐状态、简单残差及不同非线性结构的比较，也显示完整接口在真实数据总体上有收益。不过无协变量版本删去了相关投影，并非严格等参数的安慰剂实验，仍可增加随机打乱协变量或无信息协变量对照。输入、输出注入的作用也不宜概括为两者同时使用总是最佳：输出注入单独使用在部分合成条件大幅优于双注入，真实数据 MASE 也略好于双注入。模块必要性依赖任务的信息结构，而不是由总平均损失自动证明。\n\n最需要补清楚的是未来协变量的业务可得性。形式化任务明确假设未来输入可见，合成数据也满足该条件，因而不能凭变量名称指控泄漏。然而真实数据表及处理说明没有逐数据集交代每类变量在预测时点是否已经公布、是实测未来值还是提前预测值、是否有发布延迟。对于负荷和天气相关任务，这会影响结果能否部署，应把论文结论限定为所给未来信息条件下的性能。预训练数据与评估数据的重合也未在当前材料中独立审计，不应把所有对比都自动视为严格未知域测试。\n\n比较覆盖许多经典、深度及基础模型，但训练预算不完全对称。作者对自身方法搜索三种学习率，许多基线采用默认设置；预训练主干所用数据和规模也不同。因此这里更像合理工程配置下的广泛比较，而非控制所有自由度的算法优越性证明。部分基线只产生点预测，而表格仍列出分位数损失，其确定性输出如何形成分位数需更明确。MASE 提供了有用补充，避免只凭概率模型擅长的指标下结论。三个随机种子求平均值得肯定，但大量表格没有标准差、配对置信区间或显著性分析；附录预测图中的百分之八十预测区间不是重复实验的误差条，也不能替代模型排名的不确定性估计。\n\n复现材料提供官方仓库分支、主要隐藏维度、训练步数、验证选点规则及部分硬件说明，达到了可着手实现的程度。本次没有检查仓库内容或运行代码，因而资源可运行性仍属待验证。效率图记录训练加测试的平均时间和可更新参数量，支持冻结适配比全微调更便宜，却不能直接推出在线推理延迟、显存峰值或总模型存储降低到同样比例。预训练主干仍需要加载，历史预训练成本也不包含在适配成本内。\n\n综合看，这是一篇具有直接工程价值、消融覆盖较好的适配研究。最可信的结论是，在明确给定外生变量且允许目标数据训练的条件下，冻结主干的轻量条件化可以有效改善多类预训练模型。后续优先补做发布时点审计、未来协变量噪声与缺失压力测试、等预算调参、等参数无信息对照和跨种子置信区间；这些比再增加几个总平均结果更能确定方法边界。未获得用户当前具体研究任务、算力和知识背景，故个人价值四维保留未知，不用论文通用价值代替个人适配度。" }