{ "candidate_id": "ATT-0095", "reviewer": "C", "rubric": "PAPER-RUBRIC-1.0", "frozen": true, "frozen_at": "2026-11-08T02:51:01.832221+00:00", "identity": { "candidate_id": "ATT-0095", "title": "Unlocking the Power of LSTM for Long Term Time Series Forecasting", "authors": [ "Kong, Yaxuan", "Wang, Zepu", "Nie, Yuqi", "Zhou, Tian", "Zohren, Stefan", "Liang, Yuxuan", "Sun, Peng", "Wen, Qingsong" ], "arxiv_id": "2408.10006", "arxiv_version": "v2", "first_arxiv_date": "2024/08/19", "latest_arxiv_date": "2025/02/24", "source_urls": [ "https://ojs.aaai.org/index.php/AAAI/article/view/33303" ], "publication_status": "AAAI 2025, DOI 10.1609/aaai.v39i11.33303 verified", "full_text_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0095/ATT-0095_paper.pdf", "metadata_source_path": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0095/ATT-0095_arxiv.html", "full_text_sha256": "e9eb97044f109147e3a05324ec712eb3a9f2379c36aebb1c45a01f0b7d420b3d", "review_state": "ready_for_five_new_independent_reviews", "completed_new_reviews": 0, "source_versions_not_asserted_identical": true }, "source_sha256_verified": "e9eb97044f109147e3a05324ec712eb3a9f2379c36aebb1c45a01f0b7d420b3d", "reading_started_at": "2026-11-08T02:48:28Z", "reading_finished_at": "2026-11-08T02:51:01.832221+00:00", "reading_time_note": "以UTC记录开始和完成时刻", "read_coverage": { "pdf_pages": 9, "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": [ 4, 6 ] }, "legacy_read_first": true, "legacy_full_reviews_missing": true, "other_new_reviews_read": false, "experiments_run": false, "external_predecessor_fulltext_checked": false, "resource_notes": "本次完整阅读所给9页PDF，内无附录；旧条目所称仓库旧附录未随包提供，本次未取得或阅读。代码入口未实际运行或核验，未实验。", "claims": [ { "claim_id": "CL01", "claim": "P-sLSTM相对传统LSTM提升预测", "judgment": "多任务支持且PEMS长预测存在反例", "pages": "6–7" }, { "claim_id": "CL02", "claim": "理论为分块修复长期记忆提供保证", "judgment": "理论对象与实际预测模型不同，证明及逻辑缺口导致当前担保不可用", "pages": "3–5" }, { "claim_id": "CL03", "claim": "CI减轻过拟合", "judgment": "单数据局部支持，但容量混杂且表格标签冲突", "pages": "6–7" } ], "risks": [ { "severity": "关键问题", "certainty": "已确认", "description": "充分条件被解释为必要条件，指数输入门被无依据界为1，暂停CL02理论担保", "pages": "3–5" }, { "severity": "重要问题", "certainty": "已确认", "description": "缺近邻结构对照及完整训练设置，无种子方差报告", "pages": "6–7" }, { "severity": "重要问题", "certainty": "已确认", "description": "CI表正文Solar、表格Weather，数值与Solar主表一致，不能算双数据验证", "pages": "6–7" } ], "scores": [ { "dimension_id": "A1", "dimension_name": "问题重要性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "重新评估循环网络在长期预测中的能力具有实际意义。", "evidence_pages": "1–2" }, { "dimension_id": "A2", "dimension_name": "原创性", "score": 50, "status": "tentative", "confidence": "中", "judgment_range": "40–65", "reason": "具体sLSTM组合有增量，但宽泛CI首创未获充分支持；SegRNN历史线索未外查。", "evidence_pages": "2,5–6" }, { "dimension_id": "A3", "dimension_name": "实质推进程度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "相对未分块RNN改善大，但缺最接近结构对照与更多强基线。", "evidence_pages": "6–7" }, { "dimension_id": "A4", "dimension_name": "知识与解释增量", "score": 45, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "提出统计记忆视角，但充分必要条件及预测能力之间连接不足。", "evidence_pages": "3–5" }, { "dimension_id": "B1", "dimension_name": "问题定义与评价有效性", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "有限窗口预测指标清楚，统计长记忆与学习能力混用。", "evidence_pages": "2–6" }, { "dimension_id": "B2", "dimension_name": "技术与论证正确性", "score": 35, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "漂移证明遗漏指数输入门界，生成噪声与有界输入不协调，理论不涵盖patch。", "evidence_pages": "3–5" }, { "dimension_id": "B3", "dimension_name": "数据与评测流程完整性", "score": 55, "status": "tentative", "confidence": "中", "judgment_range": "45–65", "reason": "引用标准划分但缺明确流程和主要配置，无法完整核验。", "evidence_pages": "6–7" }, { "dimension_id": "B4", "dimension_name": "比较对象与比较公平性", "score": 45, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "缺分块普通LSTM和近邻结构基线，超参预算未交代。", "evidence_pages": "6–7" }, { "dimension_id": "C1", "dimension_name": "核心主张的证据充分度", "score": 55, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "预测提升有表格支持，但长记忆解决与普遍SOTA不充分。", "evidence_pages": "4–7" }, { "dimension_id": "C2", "dimension_name": "机制归因与竞争解释区分度", "score": 45, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "单数据消融有限，无法区分短循环、全状态头和指数门贡献。", "evidence_pages": "5–7" }, { "dimension_id": "C3", "dimension_name": "结果稳定性与不确定性处理", "score": 25, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "未报告随机重复次数、方差和区间，细微优势不确定。", "evidence_pages": "6–7" }, { "dimension_id": "C4", "dimension_name": "适用范围与边界验证", "score": 55, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "五数据多长度并报告PEMS失败，但解释与机制边界未直接验证。", "evidence_pages": "6–7" }, { "dimension_id": "D1", "dimension_name": "过程透明度", "score": 40, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "输入输出规模和硬件明确，主要训练超参与时间口径不完整。", "evidence_pages": "5–7" }, { "dimension_id": "D2", "dimension_name": "研究资源可获得性", "score": 60, "status": "tentative", "confidence": "中", "judgment_range": "45–70", "reason": "有公开仓库入口，旧评所指补充附录未提供，未核验可运行性。", "evidence_pages": "1" }, { "dimension_id": "D3", "dimension_name": "主张克制与报告一致性", "score": 40, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "宽泛首创、理论充分必要性混用及CI表数据标签冲突。", "evidence_pages": "4–7" }, { "dimension_id": "D4", "dimension_name": "表达清晰度", "score": 65, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "架构图易读，理论对象和实验配置不够清楚。", "evidence_pages": "3–7" }, { "dimension_id": "E1", "dimension_name": "复杂度合理性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "分块共享骨干与线性读出简单，工程复杂度较低。", "evidence_pages": "4–5" }, { "dimension_id": "E2", "dimension_name": "资源效率与成本收益", "score": 55, "status": "tentative", "confidence": "中", "judgment_range": "40–65", "reason": "两个数据快于iTransformer但计时口径和其他基线缺失。", "evidence_pages": "7" }, { "dimension_id": "E3", "dimension_name": "可迁移与可复用程度", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "分块与全状态读出可复用，理论保证不宜随模块迁移。", "evidence_pages": "4–5" }, { "dimension_id": "E4", "dimension_name": "后续研究启发性", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "reason": "短循环路径与直接历史读出是可区分且具体的研究问题。", "evidence_pages": "5–7" }, { "dimension_id": "F1", "dimension_name": "与当前问题的相关性", "score": null, "status": "unknown", "confidence": "低", "judgment_range": null, "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": null, "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": null, "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": null, "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": "## 阅读判断与独立详评\n\nP-sLSTM 是面向长期预测的轻量方法论文，带有较长的统计记忆论证。若研究兴趣在循环网络与分块读出结构，值得定向查阅；理论部分暂不作为已经证明“分块修复长记忆”的依据。本轮先阅读历史条目，再阅读第二〇二五年二月第二版全部九页，其中末两页为参考文献，没有随 PDF 附加的实验附录。历史条目提到的作者仓库旧附录并未随当前材料提供，本次没有读到它，不能据此声称已经核验所有仓库补充材料。旧完整评语缺失，旧分数不沿用。\n\n方法本身相当直接：每条变量序列独立共享一个骨干，切成片段并线性嵌入，经过具有指数门和归一化记忆的 sLSTM 块，将所有片段的隐藏表示展平后用线性层预测未来。这里最实际的价值是降低循环展开步数，并让输出层直接利用多个历史位置。与只读取最后状态的传统循环模型相比，这种读出方式可以绕开最后状态的压缩瓶颈，却不等于改变了循环过程本身的统计记忆阶。本文未用保持读出方式一致的对照将两种解释分开。\n\n原创性应按具体组合而不是“首次把通道独立用于循环预测”来理解。正文第六页确实提出后一个宽泛首创声明，而相关工作与基线没有覆盖分块循环预测的直接对手。历史摘录指出 SegRNN 的优先权冲突，本次没有另读其原始全文，因此把它保留为待核线索，而不是假装已独立完成首创权裁定。即使该宽泛首创不成立，将 sLSTM 门控、分块以及共享通道骨干组合的特定效果仍应单独评价，不宜因此将整篇价值归零。\n\n核心理论需要明显收窄。文章把无外部驱动、带独立噪声的自回归生成过程写成 Markov 链，并在遗忘门一致小于一时使用漂移条件论证几何遍历性。该结论即使全部技术条件补足，也描述生成过程的统计依赖衰减，不直接等于有限窗口监督预测的学习能力，更没有涵盖分块后的模型或最终全状态展平头。文中从一个充分条件转而说几何遍历性“只在”门值小于一时成立，把充分性用成必要性，逻辑上没有建立。以此进一步说明分块解决长记忆问题，超出了所列定理的覆盖范围。\n\n证明还有可定位的技术缺口。第四页把输入门贡献用单位向量上界并取常数 M+3q，但第三页明确定义输入门为指数函数，并非总小于一。即使在有界输入和固定参数下可以找到其他有限上界，稿中也没有给出这个上界或把它代入漂移常数。输出叠加处处正密度的噪声，与后文把生成输入限制在负一到一的有界区间，也需要明确截断、变换或不同随机过程才能一致，不能只靠“已缩放”消除。进一步，归一化状态的正性、初始状态与使隐藏向量有界的可达状态空间没有完整交代。\n\n大于一的门值讨论同样不能只看原始遗忘门。把 cell 状态除以 normalizer 后，历史平均值的实际权重是 f乘上前一 normalizer，再除以 f乘上前一 normalizer 加输入门；它同时取决于两种门和历史归一化量。某些时间门值大于一，不能单独推出所有状态始终指数增长，也不保证新输入长期被压制。文章承认比值可能稳定是有帮助的，但关于溢出、梯度裁剪必然形成门值上界的推论没有严格证明，亦未与稳定化实现的行为对应。这些问题影响理论担保，而不是自动否定已经报告的误差表。\n\n经验结果确实显示较强的改进：五个数据集、四种预测跨度，电力和 ETTm1 相对未分块 sLSTM 与 LSTM 的差距很大，Solar 也较一致。本文同时给出不利结果值得肯定：PEMS03 的较长预测明显不如 iTransformer，九十六步甚至落后未分块 sLSTM 和 LSTM，因此“任何情形都受益”不成立。作者以噪声解释失败，但没有噪声控制或去噪对照，这只是可能原因。主表若干 MSE 优势并没有对应 MAE 优势，也说明指标上的取舍应保留。\n\n基线选择不足以充分建立更广泛的最先进结论。既然最重要改动是分块和通道独立，缺少 PatchTST、相近分块循环模型，以及给普通 LSTM 使用相同分块与展平头的对照，会使收益究竟来自 sLSTM 还是预测头配置难以识别。相同回看长度并不自动代表相同参数量或调参机会。论文只给出设备与总体输入输出长度，未完整列各模型的学习率、层数、隐藏维度、批量、优化器、训练轮数和选择程序；因此无法确认性能差距是否在合理调参后保留。\n\n消融提供了一些局部证据，但覆盖窄。memory mixing 去除后 Weather 四种预测长度的差距较小，最长跨度 MSE 完全相同，不宜赋予它主要增益解释。通道独立与通道混合在一个数据集的训练、验证、测试误差显示了过拟合差异，但同时改变维度和容量，不能归因于“独立”这一因素本身。该表还有直接标签矛盾：正文说 Solar，表四写 Weather，而测试列零点一六七、零点一八零、零点一九零、零点一九六与主表 Solar 行逐一对应。按证据更可能是表格标签错误，不能把它当成两个数据集都验证过。\n\n图二的片长曲线并非单一光滑先好后坏，图三回看曲线也有局部反转，因而它们支持参数会影响表现，而不是精确证实作者给出的记忆机制。全文没有随机重复次数、方差或置信区间，千分级消融差距的稳定性未知。效率仅比较两个数据上的 iTransformer，且秒数未充分明确是单轮、固定步数还是总训练；不能扩大为所有 Transformer、MLP 或循环模型中成本最低。共享通道、较少循环步有合理效率动机，但仍有串行依赖，作者也承认并行化限制。\n\n复用上，分块骨干加全状态线性读出非常简单，代码入口公开，但入口存在不代表依赖、补充配置与版本快照均已核验。可带走的问题是：性能改善主要来自更短的循环路径，还是使预测头直接访问全部片段状态，以及二者与指数门的交互。建议优先阅读第四至五页结构与理论条件，再读第七页完整结果和消融，保持理论与经验结论分开。用户背景不足，个人适配四维保持未知。" }