{ "candidate_id": "ATT-0173", "reviewer": "C", "rubric": "PAPER-RUBRIC-1.0", "frozen": true, "frozen_at": "2026-11-08T02:27:53.633763+00:00", "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 }, "source_sha256_verified": "9226e890eaf0c0306caec014475d6c123b4185d6727e08da42b26b453351a0c8", "reading_started_at": null, "reading_finished_at": "2026-11-08T02:27:53.633763+00:00", "reading_time_note": "以UTC记录；首篇开始时间未精确打点", "read_coverage": { "pdf_pages": 34, "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": [ 15, 18 ] }, "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": [ { "id": "CL01", "claim": "长上下文支持过程识别而不只是时间依赖", "status": "充分支持（理论及受控条件）", "pages": "3—5、13—17" }, { "id": "CL02", "claim": "异质预训练增加识别开销", "status": "部分支持", "pages": "6—7、21" }, { "id": "CL03", "claim": "GPI/CF分离改善重复预测成本", "status": "部分支持，固定缓存条件", "pages": "7—8、22、26、29—32" } ], "risks": [ { "id": "R1", "issue": "噪声标准差/方差标注与最低误差、RMSE不一致", "pages": "5、15、18、34", "certainty": "已确认文本数字关系存在不一致；实现原因未知", "impact": "绝对误差尺度与过程重建" }, { "id": "R2", "issue": "真实基础模型差距不能唯一归因GPI", "pages": "6—7", "certainty": "明确比较条件差异", "impact": "CL02" }, { "id": "R3", "issue": "缓存速度不含刷新频率与漂移代价", "pages": "8、22、26", "certainty": "明确条件限制", "impact": "CL03" } ], "scores": [ { "dimension_id": "A1", "dimension_name": "问题重要性", "score": 85, "status": "rated", "confidence": "高", "judgment_range": "无", "reason": "全局及基础预测器上下文成本是重要认识和部署问题。", "evidence_pages": "1—3" }, { "dimension_id": "A2", "dimension_name": "原创性", "score": 70, "status": "tentative", "confidence": "中", "judgment_range": "60—80", "reason": "将成熟后验平均连接时序GPI/CF和缓存设计；前作全文核对有限。", "evidence_pages": "3、9、22—23" }, { "dimension_id": "A3", "dimension_name": "实质推进程度", "score": 75, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "已知阶数合成与多骨干验证支持有用认识及条件性效率推进。", "evidence_pages": "5—8、26" }, { "dimension_id": "A4", "dimension_name": "知识与解释增量", "score": 80, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "将过程身份与时间依赖区分，给出可预测架构行为的框架。", "evidence_pages": "3—5、16—17" }, { "dimension_id": "B1", "dimension_name": "问题定义与评价有效性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": "无", "reason": "任务、潜变量与均方误差目标清楚；真实GPI代理仍有局限。", "evidence_pages": "2—5" }, { "dimension_id": "B2", "dimension_name": "技术与论证正确性", "score": 75, "status": "tentative", "confidence": "中", "judgment_range": "65—85", "reason": "全方差证明成立，必要性不意味着充分性；噪声记号一致性待澄清。", "evidence_pages": "13—15、18" }, { "dimension_id": "B3", "dimension_name": "数据与评测流程完整性", "score": 80, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "时间及序列双轴切分、双验证集和随机划分清楚。", "evidence_pages": "18—19" }, { "dimension_id": "B4", "dimension_name": "比较对象与比较公平性", "score": 70, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "合成局部/全局参照合理，基础与域模型混有架构和训练差异。", "evidence_pages": "5—8、20—21" }, { "dimension_id": "C1", "dimension_name": "核心主张的证据充分度", "score": 75, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "理论、合成、真实与成本证据齐备，基础模型归因仍部分支持。", "evidence_pages": "3—8、13—34" }, { "dimension_id": "C2", "dimension_name": "机制归因与竞争解释区分度", "score": 65, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "已知阶数和非连续上下文有针对性，真实周期和容量未完全排除。", "evidence_pages": "4—8、17、24—25" }, { "dimension_id": "C3", "dimension_name": "结果稳定性与不确定性处理", "score": 75, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "三/五次运行及标准差充分可见，重复单位混合多个波动来源。", "evidence_pages": "5、16、19、24—26" }, { "dimension_id": "C4", "dimension_name": "适用范围与边界验证", "score": 65, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "涵盖多骨干多域，强时变、缓存刷新和长跨度边界仍有限。", "evidence_pages": "9、22、25" }, { "dimension_id": "D1", "dimension_name": "过程透明度", "score": 75, "status": "rated", "confidence": "高", "judgment_range": "无", "reason": "详细配置、划分与预算可重建多数过程，但噪声尺度有歧义。", "evidence_pages": "18—21" }, { "dimension_id": "D2", "dimension_name": "研究资源可获得性", "score": 45, "status": "tentative", "confidence": "低", "judgment_range": "30—65", "reason": "PDF和数据说明可得，CEREn受控申请，核心公共代码本次未核实。", "evidence_pages": "18、21" }, { "dimension_id": "D3", "dimension_name": "主张克制与报告一致性", "score": 60, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "承认未建基础模型，但GPI真实归因措辞偏强且噪声方差标注矛盾。", "evidence_pages": "7、9、15、18、34" }, { "dimension_id": "D4", "dimension_name": "表达清晰度", "score": 80, "status": "rated", "confidence": "高", "judgment_range": "无", "reason": "框架清楚、图和补充实验组织良好，个别符号报告需澄清。", "evidence_pages": "3—8、13—34" }, { "dimension_id": "E1", "dimension_name": "复杂度合理性", "score": 80, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "分离编码与短预测的额外复杂度有明确目的和替代比较。", "evidence_pages": "7—8、17、20" }, { "dimension_id": "E2", "dimension_name": "资源效率与成本收益", "score": 75, "status": "tentative", "confidence": "中", "judgment_range": "60—85", "reason": "固定缓存下Pareto改善明确，刷新及完整训练成本应独立计算。", "evidence_pages": "8、21—22、26、29—32" }, { "dimension_id": "E3", "dimension_name": "可迁移与可复用程度", "score": 85, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "GPI/CF分离与缓存原则可迁移多个骨干和静态属性接口。", "evidence_pages": "7、20、22—23" }, { "dimension_id": "E4", "dimension_name": "后续研究启发性", "score": 85, "status": "rated", "confidence": "中", "judgment_range": "无", "reason": "识别收缩、非平稳刷新与基础模型设计是清楚且有价值的边界问题。", "evidence_pages": "5、9、22—23" }, { "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": "## 总体判断与历史复核\n\n这是一篇以实证解释为主、兼有理论命题和架构设计的论文，值得优先精读其问题拆分，而不是把它当作已经验证的新基础模型。历史条目只说“尚未评审；引用较低”，没有可恢复的完整评语或实质论据；该低引用门槛不进入本次质量判断。本文首次公开于2026年6月，当前所读为34页v1，官方会议记录只用于身份确认，不能把后续录用当作内容正确性的证据。\n\n作者研究的是同一个全局预测器要服务许多不同生成过程时，为何很长的输入窗口有用。输入是一个序列的近期观测，输出是后续数值；困难不是每个单独过程必然具有长记忆，而是预测器事先不知道它面对哪个过程。第3页式(5)用条件期望的塔式法则，把最优均方误差预测分成对潜在过程参数的后验加权与给定过程后的条件预测。这一拆分本身源自成熟的贝叶斯思想，新增价值在于把它清楚地用于解释全局时序模型的上下文开销，并把历史上下文与即时预测窗口分离。前作全文本次未补齐，因此我只认可这项有意义的连接，不作“首创贝叶斯解释”结论。\n\n## 主张、idea和理论\n\nCL01：长上下文可以用于生成过程识别，而不只是捕获远距离依赖。第3—5页与附录B、C提供直接支持。对于加性独立零均值噪声、平稳且给定参数后为P阶的过程，第14页全方差分解把最优风险写为噪声方差加上过程不确定性项。当最近P个观测仍无法区分条件均值不同的过程时，W=P不可能达到噪声下界。W