{
"candidate_id": "ATT-0095",
"reviewer": "D",
"rubric": "PAPER-RUBRIC-1.0",
"review_status": "frozen",
"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
},
"analysis_date": "2026-11-08",
"reading_record": "2026-11-08 UTC：先读历史条目及身份，逐页阅读主PDF物理页1—9（正文1—7、参考文献8—9），视觉复核p4—7；另独立阅读全文及视觉检查作者仓库Appendices.pdf物理页1—5，创建日期2024-12-19，来源https://github.com/Eleanorkong/P-sLSTM/blob/main/Appendices.pdf。页码注明补充以与主文区分，未读取其他新评。",
"freeze_time_utc": "2026-11-08T02:56:08.204822+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": "d86078d19d8f0cdafdbc1b5942fc75982c496fd02a014de5749391daa65372da",
"records": [
{
"source_record_id": "S5-095",
"source_id": "S5",
"source_library_file_id": "libfile_849759d66a808191813dbe9bd1605bb4",
"source_filename": "已粘贴的文本(5).txt",
"source_line": 50,
"source_number": 95,
"title": "Unlocking the Power of LSTM for Long Term Time Series Forecasting",
"source_url": "https://ojs.aaai.org/index.php/AAAI/article/view/33303",
"raw_line": "| 34 | [P095 Unlocking the Power of LSTM for Long Term Time Series Forecasting](https://ojs.aaai.org/index.php/AAAI/article/view/33303) | 53.7 | 100% | 35% | 49.2–58.2 | 106 | AAAI 2025 会议论文终稿；作者仓库附录 2024-12-19 版本 | **重要问题**：正文声称首次将 CI 应用于 RNN 预测；SegRNN 2023 v1 已采用 CI、片段迭代和 GRU。该冲突影响宽泛首创表述，特定 sLSTM 组合仍需单独评价。；
**重要问题**：特定随机过程、噪声和遗忘门条件下的结论被扩展为较宽泛的短记忆论述；分块对长期依赖的解决未由所列定理直接建立。 | [评审](论文/P095/评审.md) · [评分数据](论文/P095/评审.json) · 无单独复核文件 |",
"raw_cells": [
"34",
"[P095 Unlocking the Power of LSTM for Long Term Time Series Forecasting](https://ojs.aaai.org/index.php/AAAI/article/view/33303)",
"53.7",
"100%",
"35%",
"49.2–58.2",
"106",
"AAAI 2025 会议论文终稿；作者仓库附录 2024-12-19 版本",
"**重要问题**：正文声称首次将 CI 应用于 RNN 预测；SegRNN 2023 v1 已采用 CI、片段迭代和 GRU。该冲突影响宽泛首创表述，特定 sLSTM 组合仍需单独评价。；
**重要问题**：特定随机过程、噪声和遗忘门条件下的结论被扩展为较宽泛的短记忆论述；分块对长期依赖的解决未由所列定理直接建立。",
"[评审](论文/P095/评审.md) · [评分数据](论文/P095/评审.json) · 无单独复核文件"
],
"normalized_title": "unlockingthepoweroflstmforlongtermtimeseriesforecasting",
"normalized_url": "ojs.aaai.org/index.php/aaai/article/view/33303",
"record_type": "legacy_scored",
"legacy_rank": 34,
"legacy_score": "53.7",
"legacy_coverage": "100%",
"legacy_tentative_share": "35%",
"legacy_range": "49.2–58.2",
"citation_verbatim": "106",
"legacy_review_version": "AAAI 2025 会议论文终稿；作者仓库附录 2024-12-19 版本",
"legacy_risks_verbatim": "**重要问题**：正文声称首次将 CI 应用于 RNN 预测；SegRNN 2023 v1 已采用 CI、片段迭代和 GRU。该冲突影响宽泛首创表述，特定 sLSTM 组合仍需单独评价。；
**重要问题**：特定随机过程、噪声和遗忘门条件下的结论被扩展为较宽泛的短记忆论述；分块对长期依赖的解决未由所列定理直接建立。",
"legacy_material_links_verbatim": "[评审](论文/P095/评审.md) · [评分数据](论文/P095/评审.json) · 无单独复核文件",
"counts_as_independent_review": false
}
]
},
"claims": [
"将sLSTM与patching、channel independence及展平预测头结合用于长程预测。",
"条件性随机过程分析用于解释潜在短记忆问题。",
"五数据集与分块、记忆混合、CI消融展示实践收益。"
],
"review_body": "论文的出发点值得肯定：不能因为裸LSTM在标准预测榜单上落后，就忽略经过现代输入表示和训练设计后的循环模型。P-sLSTM把原本用于语言建模的标量记忆sLSTM与分块、通道独立和展平预测头结合，结构清晰，工程实现门槛相对可控。在长历史中减少递推步数，同时让预测头直接看到各块表征，是合理的归纳偏置。其贡献更适合定位为对现成骨干的任务适配和经验验证，而不是新的记忆单元或已经证明解决了长记忆问题。指数门、归一化记忆和多头混合主要来自已有xLSTM工作，不能全部算作本文原创。\n\n理论部分需要与实践收益分开评价。第三至五页讨论的是没有外部输入、以过去输出驱动自身并加入独立噪声的随机过程，以及遗忘门统一小于一时的几何遍历性。即使论证成立，也只是特定生成过程的条件性短记忆结论，不等价于训练后的条件预测器不能使用远处输入。分块网络的最终线性头还直接连接所有块表征，可能绕过最后状态的记忆瓶颈；长窗口收益因此不能单独证明循环状态已经获得统计学长记忆。正文把“不能保证”推进到“分块解决短记忆”，中间缺少相应定理和区分机制的证据。\n\n我还发现证明本身存在可定位的缺口。第四页推导用常数一向量替代了输入门的新增量，但第三页明确规定输入门使用指数函数，它并不天然被一所界定。因此取常数项为M加三倍维度的现有不等式不能直接从已给假设得到。在有界输入和隐状态下，固定有限参数可能允许另取输入门上界，从而修补某个漂移界；这与当前证明正确是两回事。正文又同时使用遍处正密度噪声和输出严格缩放到有限区间，二者如何在同一个自回归过程内共存没有说明。大于一的遗忘门讨论主要是趋势性解释，尚不能由一个时刻或一个坐标超过一推出整个归一化系统必然指数发散；输入门和分母共同变化也会影响结论。这些问题不证明实测结果错误，但显著降低理论支撑的权重。补充第二页提供稳定化门，可以避免直接指数溢出，故不能忽略实际实现的稳定化机制。不过其Case2条件直接给出稳定遗忘门小于一，后文近似为原始遗忘门还需额外条件；不能一般把二者等同。稳定化与原模型隐藏输出可保持等价，并不要求门值逐个相同。\n\n主表覆盖Weather、Electricity、Solar、ETTm1和PEMS03，四种预测跨度，并纳入iTransformer、DLinear、RLinear、TimesNet等有竞争力的模型。P-sLSTM在前四类数据的MSE常有优势，例如Electricity九十六步为零点一三零，而裸sLSTM为零点二八一；这说明完整配置确实改善了该基线。然而裸LSTM和裸sLSTM落后很大，本身不足以证明指数门组合相对于其他分块循环模型有独特优势。正文没有SegRNN；不过补充材料第五页确实给出分块LSTM和PatchTST比较，不能说这些对照完全缺失。P-LSTM在ETTm1七百二十步MSE为零点四二三，P-sLSTM为零点四二一，说明这个条件下大部分增益来自分块，指数门的增量很小；Electricity则仍有较大差距。补充未充分说明P-LSTM的容量和其他部件是否完全匹配，GRU近邻亦未覆盖。第六页“首次将CI用于RNN预测”的宽泛表述也缺乏足够先行工作审计，旧历史条目指出SegRNN前例，但本次没有获得该前例全文，不能把历史摘要当作独立核验完成。\n\n消融提供了有用但有限的归因信息。记忆混合在Weather四个跨度上的收益较小，九十六步MSE从零点一五一降至零点一四九，七百二十步MSE同为零点三二零；无重复运行分散程度，不能断言这些小差异稳健。通道独立与混合的训练、验证、测试误差对比符合正则化解释，但输入维度和模型容量同时改变，无法唯一归因为避免跨通道过拟合。该表还存在内部标注问题：正文说使用Solar，表四数据集标签却写Weather，而测试数值逐项对应主表Solar。分块长度图也不是简单单峰，二十六处变差、五十六附近再次变好；把曲线概括为先提高后下降掩盖了局部不规则。论文承认最优分块依赖数据集，这是合理边界。\n\n失败情形值得认真保留。PEMS03四十八步和九十六步时，P-sLSTM的MSE分别为零点一六三和零点二零九，均差于裸sLSTM的零点一三八和零点一六三；后者也显示不是所有长预测都因分块受益。作者将此归因于噪声但没有直接噪声控制证据，所以只能视为假说。多个数据上MSE最优而MAE未最优，也不支持统一的“准确率”提升措辞。实验没有展示跨域、缺失数据和外生变量使用场景，不能据五个基准推断通用长期依赖能力。\n\n复现方面，本次除九页主文外，也独立阅读全文作者仓库五页附录，文件创建日期为二〇二四年十二月十九日，并未假定它与当前arXiv版本完全同步。附录给出Adam、初始学习率千分之一、逐任务批量、分块、隐藏宽度、层数、头数、卷积核和dropout，比仅看主文完整得多；它还列出搜索范围与稳定化消融，后者差异约千分之一至千分之二。但搜索表的头数只有一、二、三，最终ETTm1和PEMS配置却有四和六，搜索空间记录不一致；轮数、重复次数、种子及误差条仍缺。资源表仅比较两数据上的P-sLSTM与iTransformer，约五十二点五四对七十九点一二秒、九十三点一三对一百一十四点二五秒，却未明确单轮或全训练计时口径，也无DLinear等轻模型对照。附录预测图承认对极端波动仍不足，属于有用边界。公开代码没有在本评审中运行。总体判断是有工程参考价值，补充材料改善了基线和透明度评价，但理论、严格近邻比较和不确定性处理仍限制更强推荐。",
"review_chinese_character_count": 1778,
"dimensions": [
{
"dimension": "A1",
"name": "问题重要性",
"score": 80,
"status": "rated",
"confidence": "high",
"reason": "重新评估循环架构的长程预测价值重要",
"evidence_pages": "1—2",
"judgment_range": null
},
{
"dimension": "A2",
"name": "原创性",
"score": 55,
"status": "tentative",
"confidence": "mid",
"reason": "sLSTM结合分块与CI，基础部件已有，宽泛首创缺充分先行审计",
"evidence_pages": "2,5—6",
"judgment_range": [
40,
65
]
},
{
"dimension": "A3",
"name": "实质推进程度",
"score": 60,
"status": "rated",
"confidence": "mid",
"reason": "多个数据上明显强于裸循环基线，但近邻适配循环模型缺席",
"evidence_pages": "6—7",
"judgment_range": null
},
{
"dimension": "A4",
"name": "知识与解释增量",
"score": 40,
"status": "rated",
"confidence": "high",
"reason": "条件短记忆分析与分块解决主张之间存在逻辑鸿沟",
"evidence_pages": "3—6",
"judgment_range": null
},
{
"dimension": "B1",
"name": "问题定义与评价有效性",
"score": 70,
"status": "rated",
"confidence": "mid",
"reason": "预测长度与误差度量明确，但统计长记忆与条件预测能力混用",
"evidence_pages": "2—6",
"judgment_range": null
},
{
"dimension": "B2",
"name": "技术与论证正确性",
"score": 35,
"status": "rated",
"confidence": "high",
"reason": "证明将指数输入门按一界定未获假设支持，过程的有界与噪声条件需协调",
"evidence_pages": "3—5",
"judgment_range": null
},
{
"dimension": "B3",
"name": "数据与评测流程完整性",
"score": 65,
"status": "tentative",
"confidence": "mid",
"reason": "补充有逐任务参数与搜索表，但头数范围矛盾，重复记录缺失",
"evidence_pages": "6—7；补充3—5",
"judgment_range": [
55,
75
]
},
{
"dimension": "B4",
"name": "比较对象与比较公平性",
"score": 60,
"status": "rated",
"confidence": "mid",
"reason": "补充包含P-LSTM与PatchTST，仍缺SegRNN和容量严格匹配说明",
"evidence_pages": "6—9；补充5",
"judgment_range": null
},
{
"dimension": "C1",
"name": "核心主张的证据充分度",
"score": 60,
"status": "rated",
"confidence": "mid",
"reason": "主表及补充支持实践增益，理论和独特机制证据仍弱",
"evidence_pages": "3—7；补充5",
"judgment_range": null
},
{
"dimension": "C2",
"name": "机制归因与竞争解释区分度",
"score": 60,
"status": "rated",
"confidence": "mid",
"reason": "增加P-LSTM、稳定化消融，但容量、CI和预测头替代解释未全区分",
"evidence_pages": "5—7；补充4—5",
"judgment_range": null
},
{
"dimension": "C3",
"name": "结果稳定性与不确定性处理",
"score": 30,
"status": "rated",
"confidence": "high",
"reason": "主文和五页补充均未给重复运行误差，稳定化小差异无显著性证据",
"evidence_pages": "6—7；补充5",
"judgment_range": null
},
{
"dimension": "C4",
"name": "适用范围与边界验证",
"score": 55,
"status": "rated",
"confidence": "mid",
"reason": "五域四跨度含PEMS失败，但边界原因仍为假说",
"evidence_pages": "6—7",
"judgment_range": null
},
{
"dimension": "D1",
"name": "过程透明度",
"score": 65,
"status": "rated",
"confidence": "high",
"reason": "补充明确多数参数，但搜索头数与最终设置不一致、主文表四标签矛盾",
"evidence_pages": "6—7；补充3—5",
"judgment_range": null
},
{
"dimension": "D2",
"name": "研究资源可获得性",
"score": 65,
"status": "tentative",
"confidence": "mid",
"reason": "代码公开且作者补充可得；代码可执行性未验证",
"evidence_pages": "1；补充1—5",
"judgment_range": [
50,
80
]
},
{
"dimension": "D3",
"name": "主张克制与报告一致性",
"score": 40,
"status": "rated",
"confidence": "high",
"reason": "解决短记忆、首次CI与最低成本表述过强，有部分失败披露",
"evidence_pages": "2,5—7",
"judgment_range": null
},
{
"dimension": "D4",
"name": "表达清晰度",
"score": 60,
"status": "rated",
"confidence": "mid",
"reason": "架构图明晰，理论定义与符号、表四标签问题增加歧义",
"evidence_pages": "3—7",
"judgment_range": null
},
{
"dimension": "E1",
"name": "复杂度合理性",
"score": 70,
"status": "rated",
"confidence": "mid",
"reason": "适配结构简洁，展平预测头的规模和长度限制仍需注意",
"evidence_pages": "4—5",
"judgment_range": null
},
{
"dimension": "E2",
"name": "资源效率与成本收益",
"score": 50,
"status": "tentative",
"confidence": "mid",
"reason": "两数据快于单一对手，未明确计时单位且无轻模型成本比较",
"evidence_pages": "7",
"judgment_range": [
35,
65
]
},
{
"dimension": "E3",
"name": "可迁移与可复用程度",
"score": 70,
"status": "rated",
"confidence": "mid",
"reason": "分块与通道独立容易复用，附录补齐多数任务配置",
"evidence_pages": "4—6；补充3—5",
"judgment_range": null
},
{
"dimension": "E4",
"name": "后续研究启发性",
"score": 70,
"status": "rated",
"confidence": "mid",
"reason": "能启发公平比较现代循环骨干与预测头，理论仍待厘清",
"evidence_pages": "3—7",
"judgment_range": null
},
{
"dimension": "F1",
"name": "与当前问题的相关性",
"score": null,
"status": "unknown",
"confidence": "low",
"reason": "未知用户当前研究问题",
"evidence_pages": "用户背景未提供",
"judgment_range": null
},
{
"dimension": "F2",
"name": "相对于已有知识的学习收益",
"score": null,
"status": "unknown",
"confidence": "low",
"reason": "未知用户已掌握知识",
"evidence_pages": "用户背景未提供",
"judgment_range": null
},
{
"dimension": "F3",
"name": "当前使用可行性",
"score": null,
"status": "unknown",
"confidence": "low",
"reason": "未知用户设备、时间和数据约束",
"evidence_pages": "用户背景未提供",
"judgment_range": null
},
{
"dimension": "F4",
"name": "当前决策信息价值",
"score": null,
"status": "unknown",
"confidence": "low",
"reason": "未知用户具体选择与机会成本",
"evidence_pages": "用户背景未提供",
"judgment_range": null
}
],
"missing_and_limits": [
"旧完整评语缺失，只有历史行记录。",
"另读作者仓库五页附录，创建日期2024-12-19，未假定与arXiv v2完全同步。",
"未读取SegRNN全文；相关历史风险仅为待核线索，未宣称完成先行工作核验。",
"未执行代码、训练、推理或复现；公开仓库代码及提交版本未核查。",
"缺少随机种子、重复次数、误差条及完整训练/计时口径。",
"F组用户背景未知。"
],
"experiments_executed": false,
"aggregate_score": null
}