{ "rubric_version": "PAPER-RUBRIC-1.0", "paper_id": "ATT-0095", "paper_title": "Unlocking the Power of LSTM for Long Term Time Series Forecasting", "reviewer": "B", "analysis_mode": "standard", "paper_version": "2408.10006v2", "version_date": "2025/02/24", "first_public_date": "2024/08/19", "assessment_date": "2026-11-08", "reading_started_at": "2026-11-08T02:50:21.538798+00:00", "reading_completed_at": "2026-11-08T02:53:32.619544+00:00", "read_physical_pages": [ 1, 2, 3, 4, 5, 6, 7, 8, 9 ], "visual_checks_physical_pages": [ 4, 5, 6, 7 ], "pdf_sha256": "e9eb97044f109147e3a05324ec712eb3a9f2379c36aebb1c45a01f0b7d420b3d", "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": "S5-095", "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 } ], "new_peer_reviews_read": false, "review_text": "先读历史摘要，然后通读2408.10006v2九页主文。由于主文不含独立附录，本次从官方仓库取得Appendices.pdf，另读全部五页；该文件元数据显示2024年12月19日，不能直接称其与2025年2月主文逐字同版。没有完整旧评。论文用分块、通道独立共享骨干和sLSTM构成长预测器，直接输出整个预测区间，动机实用：减少循环步数并保留局部形状，可让RNN重新成为竞争性基线。不过需要把可用的工程组合、具体增益及统计记忆定理分开评价。\n\n历史首创风险已由先行原文定点核实，而不再仅引述旧评。SegRNN的2308.11200v1第3页明确采用通道独立策略、GRU以及片段级迭代，早于本作。因此主文第6页“首次把CI用于RNN预测”的宽泛表述不能成立。这个结论不取消sLSTM特定骨干、指数门与分块的组合价值，但最近RNN基线SegRNN缺席，也使性能增量缺少关键参照。这里仅核读先行文的相关段落，没有声称完成全部先行文献综述。\n\n理论的适用范围比主文一般化论述窄得多。第4页把外部输入设为上一期生成值，讨论带独立同分布噪声的自生成Markov过程；实际预测器则接收外部历史片段，并把多个隐状态展开后直接预测。两者不是同一个映射。遗忘门一致小于某个a<1只是给出短记忆的充分条件，论文第5页写成“only holds when”不能把它变成必要条件。没有定理证明分块把几何遍历过程变成长记忆过程，较长历史下预测误差改善也不是同一命题。缩短循环路径与更好的有限样本拟合是合理经验解释，不是已证明克服所有长期依赖限制。\n\n当前证明还有具体缺口。第4页漂移界用常数1代替输入门贡献，但sLSTM输入门是指数函数，不自动小于1；即使输入和隐状态都限于紧集，也只能得到依赖权重的有限上界，常数b=M+3q不能直接沿用。另一方面，假设噪声密度处处为正与生成输入始终落在[-1,1]之间需要另行说明缩放或截断过程，这会影响转移函数。f>1情形把过去项一定压过新输入作为主要解释，却未控制输入门i的相对增长。梯度裁剪也不自动对激活门值施加严格上界。因此本次暂停现有证明作为普遍记忆保证的用途，并不声称任何带补充条件的修正版都不可能成立。\n\n补充附录有重要正面信息：它明示指数门的稳定化状态，解释实现如何抑制溢出，且给稳定化消融。故不能只凭主文讨论无界状态就断言实际代码必然溢出。附录第2页由输出等价推出“稳定状态无梯度”的说法需要区分整体重参数化不变性与计算图的局部导数；第二种情形用当前log(i_t)替代前期稳定状态也额外假设时间变化可忽略，未被一般性证明。稳定版本门值本就被最大值归一化约束，主文未把它与原始Markov分析完整连接。\n\n实验包含五个数据集和四种预测长度，Weather、Electricity、Solar、ETTm1表现有竞争力，但不是全设置最优。PEMS03的48步P-sLSTM MSE为0.163，差于无分块sLSTM的0.138；96步为0.209，差于0.163甚至普通LSTM的0.196。这个失败反例说明分块有条件收益。作者将其解释为噪声较大只是可能性，没有噪声干预证据来区分信息损失、最佳块长、模型容量或优化因素。附录补入PatchTST和P-LSTM比较，不能误说完全没有这些对照；ETTm1在720步P-LSTM已达0.423，与P-sLSTM的0.421很接近，说明相当一部分巨大传统LSTM差距可由分块本身解释。\n\n消融的证据量仍不足以支持全部机制。记忆混合在Weather只有千分位增益且无方差；CI和CM改变输入维度与参数容量，验证集改善支持较少过拟合现象，不单独识别跨通道关联本身的价值。表4标题列Weather，正文说Solar，测试值0.167、0.180、0.190、0.196也对应Solar主表，应将此列为数据标签不一致。图2块长效果非单峰，6、16与56都较低，不能概括成简单先好后差。附录有详细搜索范围但候选头数仅1、2、3，实际配置又出现4与6；基线是否获得同样预算、选择集、种子与停止规则没有完整披露。\n\n效率比较只有Weather和ETTm1对iTransformer的训练秒数，未明确计时是一轮还是全训练、批量与预热，也没有SegRNN、简单线性或推理端完整成本，故“最低计算成本”只能限这两个比较。当前README提供两套CUDA12.1兼容环境、ninja和脚本入口，资源可访问不等于已复现；本次没有训练推理。整体结论是值得作为分块RNN经验候选，但宽泛首创、理论解决长记忆和普遍最优不能接受。值得保留的问题是分块收益来自更短优化路径还是可识别的记忆改变；如果普通P-LSTM在同条件已解释大部分增益，sLSTM机制的独立贡献就更窄。优先亲读主文第4—5页条件与第7页失败设置，再看附录第4—5页P-LSTM及稳定化消融。这些是当前证据限制，不要求用户执行实验。", "review_hanzi_count": 1442, "claims": [ { "id": "CL01", "claim": "分块sLSTM在若干基准有经验收益", "status": "supported_with_scope", "pages": [ 7, "附录5" ] }, { "id": "CL02", "claim": "首次把CI用于RNN时间序列预测", "status": "contradicted_by_prior_primary_source", "pages": [ 6 ], "prior": "SegRNN 2308.11200v1 p3" }, { "id": "CL03", "claim": "现有理论证明分块解决普遍长记忆问题", "status": "suspended", "pages": [ 4, 5 ] }, { "id": "CL04", "claim": "实际稳定化实现必然溢出", "status": "not_claimed_and_not_established", "pages": [ "附录2", "附录5" ] } ], "scores": [ { "dimension_id": "A1", "dimension_name": "问题重要性", "score": 80, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "恢复高效循环网络长预测能力值得研究。", "evidence_pages": [ 1, 2 ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "A2", "dimension_name": "原创性", "score": 45, "status": "tentative", "confidence": "中", "judgment_range": [ 35, 60 ], "rationale": "CI用于RNN的首创被SegRNN v1第3页否定，特定sLSTM组合仍有价值。", "evidence_pages": [ 5, 6 ], "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": 60, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "分块sLSTM表现强，但缺最近SegRNN对照且P-LSTM解释部分增益。", "evidence_pages": [ 7, "附录5" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "A4", "dimension_name": "知识与解释增量", "score": 45, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "记忆分析与实际预测器存在落差，定理及稳定化推导有缺口。", "evidence_pages": [ 3, 4, 5, "附录2" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "B1", "dimension_name": "问题定义与评价有效性", "score": 70, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "预测任务与长度明确，但统计长记忆定义不能直接对应预测误差。", "evidence_pages": [ 3, 4, 6 ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "B2", "dimension_name": "技术与论证正确性", "score": 35, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "指数输入门按1界定、噪声紧支持冲突及充分条件必要化。", "evidence_pages": [ 4, 5, "附录2" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "B3", "dimension_name": "数据与评测流程完整性", "score": 55, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "附录提供配置，但重复、选择程序与划分细节核验不足。", "evidence_pages": [ 6, "附录3", "附录4", "附录5" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "B4", "dimension_name": "比较对象与比较公平性", "score": 50, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "多类基线和补充P-LSTM/PatchTST有价值，缺SegRNN及等预算证据。", "evidence_pages": [ 6, 7, "附录5" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "C1", "dimension_name": "核心主张的证据充分度", "score": 60, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "支持部分数据有效，不支持理论彻底解决长记忆。", "evidence_pages": [ 5, 6, 7 ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "C2", "dimension_name": "机制归因与竞争解释区分度", "score": 50, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "有块长、CI、记忆与P-LSTM消融，容量和优化竞争解释仍强。", "evidence_pages": [ 6, 7, "附录5" ], "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": "未见随机种子重复/方差，千分位消融不支持稳定排序。", "evidence_pages": [ 7, "附录5" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "C4", "dimension_name": "适用范围与边界验证", "score": 55, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "五组常用数据，主动报告高波动失败，但机制测试很窄。", "evidence_pages": [ 6, 7, "附录4" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "D1", "dimension_name": "过程透明度", "score": 55, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "附录超参较多，表4标签、搜索头数与实际配置不一致。", "evidence_pages": [ 7, "附录3", "附录4", "附录5" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "D2", "dimension_name": "研究资源可获得性", "score": 75, "status": "tentative", "confidence": "中", "judgment_range": [ 65, 85 ], "rationale": "官方README和独立附录可访问，环境与脚本齐备但未执行。", "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": 40, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "首创表述错误、充分条件必要化、解决短记忆和最低成本过宽。", "evidence_pages": [ 2, 5, 6, 7 ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "D4", "dimension_name": "表达清晰度", "score": 60, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "流程可理解，理论域、符号及数据标签影响准确性。", "evidence_pages": [ 4, 5, 7 ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "E1", "dimension_name": "复杂度合理性", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "分块共享骨干较简单，flatten预测头仍随长度增长。", "evidence_pages": [ 5 ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "E2", "dimension_name": "资源效率与成本收益", "score": 50, "status": "tentative", "confidence": "中", "judgment_range": [ 35, 65 ], "rationale": "仅两数据对一个模型训练时间，未说明计时口径或总调参成本。", "evidence_pages": [ 7 ], "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": 70, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "策略易移植，但sLSTM CUDA环境与参数搜索有部署成本。", "evidence_pages": [ 5, "附录3" ], "missing_information": "", "source_ids": [ "S01" ] }, { "dimension_id": "E4", "dimension_name": "后续研究启发性", "score": 75, "status": "rated", "confidence": "高", "judgment_range": null, "rationale": "区分有限样本优化与统计记忆，比较普通分块RNN很有启发。", "evidence_pages": [ 5, "附录5" ], "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": "CL02", "issue": "先行SegRNN v1明确CI、GRU与片段迭代。", "pages": [ 6 ], "action": "撤回宽泛首创，不否定特定组合。" }, { "severity": "critical", "scope": "CL03", "issue": "漂移界输入门常数不成立，噪声/紧集假设缺口，且没有分块长记忆定理。", "pages": [ 4, 5 ], "action": "暂停现有理论保证。" }, { "severity": "moderate", "scope": "reproduction", "issue": "独立附录较主文旧；表4数据标签和搜索头数不一致。", "pages": [ 7, "附录4", "附录5" ], "action": "明确版本与配置未知。" } ], "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-0095/ATT-0095_paper.pdf", "version": "2408.10006v2" }, { "source_id": "S02", "source_type": "待核历史记录", "location": "/workspace/scratch/ac01e91af073/candidate_recovery/priority_packet_20261108/papers/ATT-0095/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/Eleanorkong/P-sLSTM/blob/main/README.md", "blob_sha": "60381efe934c6cc5658d14c838381d0155fe1c31", "notes": "仅README和附录阅读，不运行模型。" }, "supplement_read": { "url": "https://raw.githubusercontent.com/Eleanorkong/P-sLSTM/main/Appendices.pdf", "local_path": "/workspace/shared/candidate_review_batch001/reviewer_b/resources/ATT-0095_Appendices_20241219.pdf", "sha256": "2beb391d55796a12ef76dcf81cad63f6293220e3c0924205a78592b60cba7bcf", "pdf_creation_date": "2024-12-19", "read_physical_pages": [ 1, 2, 3, 4, 5 ], "visual_checks_physical_pages": [ 1, 2, 4, 5 ], "not_asserted_identical_version_to_main": true }, "targeted_prior_check": { "url": "https://arxiv.org/pdf/2308.11200v1", "local_path": "/workspace/shared/candidate_review_batch001/reviewer_b/resources/SegRNN_2308.11200v1.pdf", "sha256": "dfe3eada48caa830af7356ebc74c0f053d24ae3fa916e07eff934e9453f0e662", "pages_read": [ 1, 3, 8 ], "scope": "定点核查CI、GRU、片段迭代及先行版本，不声称全文综述。" } }