Stage2.9 指标 10 项 · Assistant 156 轮 · 工具调用 150 次 · 回溯/审计 5 条 · 控制器尝试 6 次
第三阶段需要短 Rewrite：当前中文 743 个可见字符，目标 180–360；英文 290 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
摘要把医学图像分割称为医学图像分析的重要步骤，并指出连续池化和步幅卷积操作会损失空间信息；作者单位中有 Inception Institute of Artificial Intelligence。沿着这项研究的引文，在引言讨论基于图像块的脑病灶分割时，该引言把滑动窗口造成的冗余计算和无法学习全局特征列为这种方案的两个主要缺点。
继续阅读所引用的多尺度三维脑病灶分割研究。其“Discussion and Conclusion”先说明被限制在[0,1]的输出从贝叶斯意义上不是真正概率；该讨论明确指向一项2015年的较早工作，并把带随机省略单元的训练解释为深度高斯过程中的近似贝叶斯推断，同时说可由随机掩码采样估计不确定性。这个时间顺序和两条具体联系应指向被引用的较早贡献，而不是后来的改写。请据此确定终点论文，给出准确标题、稳定标识符和当前规范 GitHub 地址。
选择作者维护的实验仓库：该仓库的 README 说明它带有 UCI 数据、沿用相同的20个随机训练/测试划分，并把 BayesOpt（论文）结果与网格搜索结果并列；在 `experiment.py` 第110–217行，`best_network` 循环加载每个划分、用验证集选择超参数、在完整训练集上重训、调用预测并记录 RMSE 与 log likelihood。确认论文与仓库身份后，请在 CPU 上运行并复现该论文第5.3节表1的主回归评估，报告 Table 1 data rows 2 through 6 inclusive, both result columns（数据行第2至第6行，含首尾；两列结果）共十个单元格；采用原始的一层50单元配置、论文列出的两个候选掩码概率、验证集选出的精度、PBP 训练轮数的十倍、原始20个随机划分和仓库评估流程，并报告实测结果而非目标值。
查看英文题面
The abstract of an unnamed medical-imaging study calls segmentation an important step in medical image analysis and notes that consecutive pooling and strided convolutions lose spatial information; one affiliation is the Inception Institute of Artificial Intelligence. Following its citation, the introduction discusses patch-based brain-lesion segmentation, and the introduction lists redundant computation from sliding windows and inability to learn global features as two main drawbacks.
Read the cited multi-scale three-dimensional brain-lesion segmentation study. Its “Discussion and Conclusion” first explains that outputs bounded in [0,1] are not true Bayesian probabilities; the discussion explicitly points to an earlier 2015 work, interprets training with randomly omitted units as approximate Bayesian inference in deep Gaussian processes, and says uncertainty can be estimated from stochastic-mask samples. This chronology and the two concrete links identify the cited earlier contribution rather than a later reformulation. Give the terminal paper's exact title, a stable identifier, and its current canonical GitHub address.
Choose the author-maintained experiment repository: the repository README says it carries the UCI data, uses the same 20 random train/test splits, and places BayesOpt (paper) results beside grid-search results; in `experiment.py`, lines 110–217, the `best_network` loop loads each split, selects hyperparameters on validation data, retrains on the full training split, calls prediction, and records RMSE and log likelihood. After confirming the paper and repository identity, run the paper's main Section 5.3, Table 1 regression evaluation on CPU and report Table 1 data rows 2 through 6 inclusive, both result columns (ten cells total); use the original one-layer, 50-unit configuration, the two candidate mask probabilities listed in the paper, validation-selected precision, ten times the PBP epoch count, the original 20 random splits, and the repository evaluation workflow, and report measurements rather than target values.
终点答案与规范仓库
Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning
论文 · GitHub
稳定标识符：arXiv:1506.02142
HEAD：6eb4497628d12b0f300f4b4f6bdc386bebad565c
代码锚点
experiment_code experiment.py:110–217
Section 5.3, Table 1 and the surrounding regression-experiment description
该仓库的 README 说明它带有 UCI 数据、沿用相同的20个随机训练/测试划分，并把 BayesOpt（论文）结果与网格搜索结果并列；在 `experiment.py` 第110–217行，`best_network` 循环加载每个划分、用验证集选择超参数、在完整训练集上重训、调用预测并记录 RMSE 与 log likelihood。
查看两跳论文链与逐边 PDF 证据
- 2019 · CE-Net: Context Encoder Network for 2D Medical Image Segmentation
https://arxiv.org/pdf/1903.02740 - 2016 · Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation
https://arxiv.org/pdf/1603.05959 - 2015 · Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning
https://arxiv.org/pdf/1506.02142v6.pdf
关键讨论 Section I, INTRODUCTION
该引言把滑动窗口造成的冗余计算和无法学习全局特征列为这种方案的两个主要缺点。
随后，Kamnitsas 等人[40]采用带全连接条件随机场（CRF）的多尺度三维 CNN 来提升基于图像块的脑病灶分割。显然，该方案带来两个主要缺点：滑动窗口造成的计算冗余，以及无法学习全局特征。
查看英文证据原文
Then, Kamnitsas et al. [40] employed a multi-scale 3D CNN architecture with fully connected conditional random field (CRF) for boosting patch based brain lesion segmentation. Obviously, this solution introduces two main drawbacks: redundant computation caused from sliding window and the inability to learn global features.
关键讨论 Section 5, Discussion and Conclusion
该讨论明确指向一项2015年的较早工作，并把带随机省略单元的训练解释为深度高斯过程中的近似贝叶斯推断，同时说可由随机掩码采样估计不确定性。
虽然输出被限制在[0,1]范围内并为方便而通常称作概率，但从贝叶斯意义上说它并不是真正的概率。贝叶斯网络研究旨在缓解这一限制。一个例子是 Gal 和 Ghahramani（2015）的近期工作，他们表明可以通过采样随机掩码来估计模型置信度。
查看英文证据原文
Although the output is bounded in the [0, 1] range and commonly referred to as probability for convenience, it is not a true probability in a Bayesian sense. Research towards Bayesian networks aims to alleviate this limitation. An example is the recent work of Gal and Ghahramani (2015) who show that model confidence can be estimated via sampling the dropout mask.
查看回溯与仓库迁移记录
- 论文链 · 通过
CE-Net → Efficient multi-scale 3D CNN → endpoint
已用引用方 PDF 的章节、原文和讨论主题核验。 - 规范仓库 · 否决
https://github.com/yaringal/DropoutUncertaintyCaffeModels
该仓库不能直接支持目标论文实验，未作为规范仓库。 - 规范仓库 · 通过
https://github.com/yaringal/DropoutUncertaintyExps
已克隆并核对 origin、HEAD 与论文实验内容。 - 代码特征 · 通过
experiment.py best_network evaluation loop
已核验论文内容与文件、符号和代码行的对应关系。 - Mask 与改写 · 通过
Repository-specific experiment.py and README clues followed by an explicit five-row table locator
题面保留可搜索线索，并隐藏终点身份、数据集和目标值。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：0/10；严格命中：0/10。 共审计 4 次独立尝试，当前展示最强有效尝试：corrected_dropout。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 10 项 · Assistant 38 轮 · 工具调用 37 次 · 回溯/审计 2 条
第三阶段需要短 Rewrite：当前中文 426 个可见字符，目标 180–360；英文 210 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
摘要把顺序推荐描述为理解用户兴趣演化并预测后续行为的任务；其作者单位线索包含中国科学技术大学。沿着这条线索，在第2节找到一段关于均匀与变化时间间隔的讨论：其第2节明确讨论了均匀时间间隔相较变化间隔能改善顺序推荐，并提出把原始间隔转换为均匀间隔的算子。再查该被引论文的 Experimentation 部分，其实验部分把终点方法列入非序列推荐对比模型。请据两份 PDF 确认更新→中间→终点的引用链，并识别终点论文的准确标题、稳定标识符和作者维护的 GitHub 仓库。在仓库主 Python 文件中，检查第 230–251 行的嵌入构造方法：它对应论文第 3.1.1 节的公式 (3)，逐层对称归一化邻居传播后对各层取均值。请把仓库线索直接用于 CPU 复现：运行全部三个数据集，覆盖表3三层行的六个 recall@20/ndcg@20 单元、四层行的三个指定单元和两层行的一个指定 recall@20 单元，使用嵌入维度64、K=20及均匀层组合系数，报告每个观测值且不要依据预期值调参。
查看英文题面
The abstract frames sequential recommendation as understanding users' evolving interests and predicting future behavior, and the author affiliations include the University of Science and Technology of China. Follow that clue into Section 2, where the paper discusses uniform versus varying time intervals: Section 2 explicitly discusses that uniform time intervals improve sequential recommendation over varying intervals and motivates operators that convert original intervals to uniform ones. In the cited paper's Experimentation section, its Experimentation section includes the endpoint method among the non-sequential recommendation comparison models. Using the two primary PDFs, verify the newer-to-middle-to-terminal citation chain, then identify the terminal paper's exact title, stable identifier, and author-maintained GitHub repository. In the repository's main Python file, inspect the embedding-construction method at lines 230–251: it maps to Section 3.1.1 Equation (3), propagating through symmetrically normalized neighbors layer by layer and then averaging the layers. Use the repository clue directly for a CPU reproduction: run all three bundled datasets and report every masked Table 3 cell—the six recall@20/ndcg@20 cells in the three-layer row, the three specified cells in the four-layer row, and the specified recall@20 cell in the two-layer row—with embedding size 64, K=20, and uniform layer-combination coefficients; do not tune against expected values.
终点答案与规范仓库
LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation
论文 · GitHub
稳定标识符：arXiv:2002.02126
HEAD：b067ec05f16a24e0e1efe63cd82385f0c469af20
代码锚点
formula_code LightGCN.py:230–251
Section 3.1.1, Equation (3), page 4
在仓库主 Python 文件中，检查第 230–251 行的嵌入构造方法：它对应论文第 3.1.1 节的公式 (3)，逐层对称归一化邻居传播后对各层取均值。
查看两跳论文链与逐边 PDF 证据
- 2024 · Temporal Graph Contrastive Learning for Sequential Recommendation
https://ojs.aaai.org/index.php/AAAI/article/download/28789/29511 - 2023 · Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation
https://ojs.aaai.org/index.php/AAAI/article/download/25540/25312 - 2020 · LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation
https://arxiv.org/pdf/2002.02126v4.pdf
关键讨论 Section 2.1, Sequential Recommendation
其第2节明确讨论了均匀时间间隔相较变化间隔能改善顺序推荐，并提出把原始间隔转换为均匀间隔的算子。
最近，TiCoSeRec（Dang 等，2023）发现，均匀分布的时间间隔比变化的时间间隔更能改善顺序推荐模型，因此提出了把原始时间间隔转换为均匀间隔的具体算子。
查看英文证据原文
Recently, TiCoSeRec (Dang et al. 2023) found out that uniformly distributed time intervals improve SR models better than varying time intervals, so it suggested specific operators to convert origin time intervals into uniform ones.
关键讨论 Experimentation, Comparison methods
其实验部分把终点方法列入非序列推荐对比模型。
随后，我们与以下三组推荐模型进行比较：非序列模型包括 BPR（Rendle 等，2012）和 LightGCN（He 等，2020）。
查看英文证据原文
Then, we compare with the following three groups of recommendation models. • Non-sequential models include BPR (Rendle et al. 2012) and LightGCN (He et al. 2020).
查看回溯与仓库迁移记录
- 论文链 · 通过
AAAI 2024 temporal graph contrastive learning -> AAAI 2023 interval augmentation -> arXiv 2002.02126
已用引用方 PDF 的章节、原文和讨论主题核验。 - repository · replaced
https://github.com/Wuyxin/LightGCN-parallelized-version
已由同一造题会话完成证据核验。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：0/10；严格命中：0/10。 共审计 4 次独立尝试，当前展示最强有效尝试：initial。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 7 项 · Assistant 87 轮 · 工具调用 84 次 · 回溯/审计 5 条 · 控制器尝试 3 次
第三阶段需要短 Rewrite：当前中文 493 个可见字符，目标 180–360；英文 198 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
图卷积多组学分类研究在引言第 2 页的未编号正文中提到：较早的乳腺癌预后工作把 mRNA 与 miRNA 表达特征和临床信息在隐藏层整合，以改善预后预测；该段还对被引工作给出具体评价。沿引用回溯，被引的生存分析研究开头指出，虽然许多模型能从数据预测差异生存，但仍需要能聚合并筛选相关预测因子的复杂算法。生存研究在题为“Integrating Multi-Omics Features Increased the Performances”的结果小节用一致性指数比较了 Cox-nnet、另一种神经生存模型、GLMNET 和 RSF。继续沿这条较新到较旧再到终点的引用链，终点论文第 3.1 节公式 (3) 是负对数部分似然；规范仓库中的 `deep_surv.py` 文件、`_negative_log_likelihood` 符号第 135–168 行实现公式 (3)，其中对风险取指数、形成累积风险对数分母，并返回按事件加权的负似然。请依据上述仓库线索，在 CPU 上复现第 4.1 节表 1 提议方法列的七个 C-index 结果单元，并逐项报告实际观测值。请识别终点论文的准确标题、稳定标识符和规范 GitHub URL，并说明你的复现实测结果。
查看英文题面
A graph-convolutional multi-omics classification study states in unnumbered opening body text on PDF page 2 that an earlier breast-cancer prognosis study integrated mRNA and miRNA expression with clinical information at hidden layers for breast-cancer prognosis; the passage also reports a concrete assessment of that cited work. The cited survival-analysis study opens by noting that although many models can predict differential survival from data, sophisticated algorithms are needed to aggregate and filter relevant predictors. The survival study compared its model with Cox-nnet, another neural survival model, GLMNET, and RSF using concordance index in its Results section titled “Integrating Multi-Omics Features Increased the Performances.” Follow this newer-to-older-to-terminal citation chain. The terminal paper’s Section 3.1 Equation (3) is a negative log partial likelihood. In the canonical repository, file `deep_surv.py`, symbol `_negative_log_likelihood`, lines 135–168 implements Equation (3): it exponentiates risk scores, forms the cumulative log-risk denominator, and returns the event-weighted negative likelihood. Using the repository clue above, run the seven C-index result cells in the proposed-method column of Section 4.1, Table 1 on CPU, and report each observed value. Identify the terminal paper’s exact title, stable identifier, and canonical GitHub URL, then give the measured results.
终点答案与规范仓库
DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network
论文 · GitHub
稳定标识符：1606.00931
HEAD：41eed003e5b892c81e7855e400861fa7a2d9da4f
代码锚点
formula_code deepsurv/deep_surv.py:135–168
Section 3.1, Equation (3), PDF page 4
规范仓库中的 `deep_surv.py` 文件、`_negative_log_likelihood` 符号第 135–168 行实现公式 (3)
查看两跳论文链与逐边 PDF 证据
- 2021 · MOGONET integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification
https://www.nature.com/articles/s41467-021-23774-w.pdf - 2019 · SALMON: Survival Analysis Learning With Multi-Omics Neural Networks on Breast Cancer
https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2019.00166/pdf - 2018 · DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network
https://arxiv.org/pdf/1606.00931v3.pdf
关键讨论 Unnumbered opening body text in the Introduction, PDF page 2
较早的乳腺癌预后工作把 mRNA 与 miRNA 表达特征和临床信息在隐藏层整合，以改善预后预测
例如，Huang 等人将 mRNA 表达和 miRNA 表达数据的特征与额外的临床信息在隐藏层整合，以更好地预测乳腺癌预后。
查看英文证据原文
For example, Huang et al.2 integrated the features of mRNA expression and miRNA expression data with additional clinical information at hidden layers for better prognosis prediction in breast cancer.
关键讨论 Results, section titled “Integrating Multi-Omics Features Increased the Performances,” PDF page 5
生存研究在题为“Integrating Multi-Omics Features Increased the Performances”的结果小节用一致性指数比较了 Cox-nnet、另一种神经生存模型、GLMNET 和 RSF
接下来，我们将 SALMON 与最先进的基于深度学习的癌症生存预后模型 Cox-nnet、另一种近期提出的 DeepSurv，以及两个传统模型——带 Cox 回归的广义线性模型 GLMNET 和 RSF——进行了比较。
查看英文证据原文
Next, we compared SALMON to the state-of-the-art Deep Learning-based cancer survival prognosis model Cox-nnet (Ching et al., 2018), as well as another recently proposed DeepSurv (Katzman et al., 2018), and two traditional models generalized linear model with Cox regression (GLMNET) (Friedman et al., 2010) and RSF (Ishwaran et al., 2008).
查看回溯与仓库迁移记录
- canonical_repo · 否决
https://github.com/czifan/DeepSurv.pytorch
已由同一造题会话完成证据核验。 - canonical_repo · 通过
https://github.com/jaredleekatzman/DeepSurv
已由同一造题会话完成证据核验。 - 论文链 · 通过
MOGONET PDF -> SALMON PDF -> endpoint PDF
已用引用方 PDF 的章节、原文和讨论主题核验。 - Mask 与改写 · 否决
prior public wording
已由同一造题会话完成证据核验。 - Mask 与改写 · 通过
revised bilingual wording
题面保留可搜索线索，并隐藏终点身份、数据集和目标值。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：5/7；严格命中：0/7。 共审计 4 次独立尝试，当前展示最强有效尝试：initial。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 10 项 · Assistant 154 轮 · 工具调用 149 次 · 回溯/审计 7 条 · 控制器尝试 5 次
第三阶段需要短 Rewrite：当前中文 490 个可见字符，目标 180–360；英文 203 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
2017年的一项研究在 Section 2 对一个由536篇维基百科文章、约23K段落和超过十万问题构成的机器阅读理解基准进行实证分析，并比较固定与动态推理轮数。该研究所引用的论文在 Section 2（Existing Datasets）中，把一个更早的纯合成阅读理解基准描述为按每项任务所需的推理类型分层，并将这类合成任务作为统一框架下的能力测试。末端论文仓库中明确的相对路径后缀是 `tasks/Negation.lua`（位于 Lua 任务包中，不是另一任务文件）；其中的 `Negation:generate_story` 位于第22–87行，在生成 yes/no 问题前采样 teleport 或 set 动作，set 分支显式选择肯定状态或不同地点来制造否定断言，随后把真值和支持事实引用写入问题。请据此给出末端论文的准确标题、稳定标识符和官方 GitHub 仓库，并紧接着沿这条明确的 `tasks/Negation.lua` 代码线索按论文设定在 CPU 上完成复现：在主文 Section 5 的 Table 3 中，从采用支持事实强监督的组合方法列读取任务行1–10，逐行报告十个实测测试准确率，不要直接抄录参考数字。
查看英文题面
A 2017 study’s Section 2 empirically analyzes a machine-comprehension benchmark built from 536 Wikipedia articles and 23K paragraphs, with over 100,000 questions, while comparing fixed and dynamically chosen reasoning turns. The study cited there, in its Section 2 (Existing Datasets), characterizes an older fully synthetic reading-comprehension benchmark as stratified by the reasoning types each task requires and presents such synthetic tasks as capability tests under one framework. The distinctive repository-relative suffix is `tasks/Negation.lua` inside the Lua task package (not another task file); its `Negation:generate_story` at lines 22–87 samples teleport or set actions before each yes/no question, with the set branch explicitly choosing an affirmative state or a different location to create a false assertion, then emits the question with its truth value and supporting-fact references. Use these clues to give the older paper’s exact title, stable identifier, and official GitHub repository, then follow the explicit `tasks/Negation.lua` code trail to carry out the paper-aligned reproduction on CPU: in the main paper’s Section 5, Table 3, take task rows 1–10 from the combined-method column under strong supporting-fact supervision and report the ten observed test-accuracy values, one per row, rather than transcribing reference numbers.
终点答案与规范仓库
Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks
论文 · GitHub
稳定标识符：1502.05698
HEAD：ccd8fd6af76d35346109e7bb5f7de9138d055e01
代码锚点
section_code lua/babi/tasks/Negation.lua:22–87
Section 3, The Tasks — Principles (PDF p.6), with the Section 4 simulation procedure
末端论文仓库中明确的相对路径后缀是 `tasks/Negation.lua`（位于 Lua 任务包中，不是另一任务文件）；其中的 `Negation:generate_story` 位于第22–87行，在生成 yes/no 问题前采样 teleport 或 set 动作，set 分支显式选择肯定状态或不同地点来制造否定断言，随后把真值和支持事实引用写入问题。
查看两跳论文链与逐边 PDF 证据
- 2017 · An Empirical Analysis of Multiple-Turn Reasoning Strategies in Reading Comprehension Tasks
https://arxiv.org/pdf/1711.03230.pdf - 2016 · SQuAD: 100,000+ Questions for Machine Comprehension of Text
https://aclanthology.org/D16-1264.pdf - 2015 · Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks
https://arxiv.org/pdf/1502.05698v10.pdf
关键讨论 Section 2, Reading Comprehension Tasks
2017年的一项研究在 Section 2 对一个由536篇维基百科文章、约23K段落和超过十万问题构成的机器阅读理解基准进行实证分析，并比较固定与动态推理轮数。
SQuAD 是一个机器阅读理解数据集，由 536 篇维基百科文章（23K 个段落）构成，包含超过 100,000 个问题。
查看英文证据原文
SQuAD is a machine comprehension dataset constructed on 536 Wikipedia articles (23K paragraphs), with more than 100,000 questions.
关键讨论 Section 2, Existing Datasets
该研究所引用的论文在 Section 2（Existing Datasets）中，把一个更早的纯合成阅读理解基准描述为按每项任务所需的推理类型分层，并将这类合成任务作为统一框架下的能力测试。
BAbI（Weston 等，2015）是一个完全合成的阅读理解数据集，按解决每项任务所需的不同推理类型分层。
查看英文证据原文
BAbI (Weston et al., 2015), a fully synthetic RC dataset, is stratified by different types of reasoning required to solve each task.
查看回溯与仓库迁移记录
- 论文链 · 否决
HISTORIAE -> SQuAD -> endpoint (prior metadata branch)
PDF 证据不足或只是普通引用，已回溯其他分支。 - 论文链 · 通过
An Empirical Analysis of Multiple-Turn Reasoning Strategies -> SQuAD -> endpoint
已用引用方 PDF 的章节、原文和讨论主题核验。 - repository · 通过
facebook/bAbI-tasks versus facebookarchive/bAbI-tasks
已由同一造题会话完成证据核验。 - 代码特征 · 通过
formula -> section-code -> experiment workflow
已核验论文内容与文件、符号和代码行的对应关系。 - black_box_repair · 否决
generic generate_story / WhereIsActor ambiguity
已由同一造题会话完成证据核验。 - rewrite · 通过
exact Negation.lua path plus affirmative-versus-different-location behavior
已由同一造题会话完成证据核验。 - evidence · 通过
middle-paper section label
已由同一造题会话完成证据核验。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：10/10；严格命中：3/10。 共审计 4 次独立尝试，当前展示最强有效尝试：corrected_babi_harp_kerr。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 9 项 · Assistant 112 轮 · 工具调用 108 次 · 回溯/审计 5 条 · 控制器尝试 4 次
第三阶段需要短 Rewrite：当前中文 522 个可见字符，目标 180–360；英文 245 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
一篇 AAAI-23 的图嵌入论文开篇提出两个明确问题：能否从更小的摘要图学习节点嵌入，以及原图与摘要图上的嵌入有何关系？它在配置模型重构图下，为 DeepWalk、LINE 和 GCN 给出闭式分析，并用线性恢复映射连接两者。在相关工作中，它还具体描述了一个较晚出现的谱多层框架：先融合拓扑与节点属性，反复按高谱相似度合并节点，再用图滤波器细化嵌入。中间研究的摘要把该方法概括为四个内核——图融合、谱图粗化、图嵌入和嵌入细化；其传导式基线模型小节将该框架与两个更早的多层方法比较，并指出随机游走基线在嵌入质量或可扩展性上的提升。请按较新—中间—终点顺序找出三篇论文，并依据引用 PDF 的原文核对两条关系。解析终点论文的作者维护仓库，确认克隆内容与论文实现一致。在克隆的作者维护仓库中，检查 src/scoring.py 的 scoring 函数（第 40–107 行）：它将特征和标签打乱十次，训练 one-vs-rest 逻辑回归，并平均 Macro F1；沿这条评测路径直接进行下面的 CPU 复现。请在 CPU 上完成 Table 2 的复现实验：按照论文列出的设置和十次分类打乱，在该表的三个数据集列上运行三个骨干变体，并报告全部九个 Macro F1 单元格；不得用期望值替代实测结果。
查看英文题面
An AAAI-23 graph-embedding paper opens with two precise questions: can node embeddings be learned from a smaller summary graph, and what is the relationship between embeddings on the original and summary graphs? It answers them with closed-form analyses for DeepWalk, LINE, and GCN under a configuration-based reconstructed graph and a linear restoration map. In its related-work discussion, it describes a later spectral multi-level framework that first fuses topology with node attributes, repeatedly merges nodes with high spectral similarity, and refines embeddings with a graph filter. The middle study’s abstract defines a four-kernel framework—graph fusion, spectral graph coarsening, graph embedding, and embedding refinement—and its transductive-baseline section compares that framework with two earlier multi-level methods, noting gains over random-walk baselines in embedding quality or scalability. Identify the three papers in newer-to-middle-to-terminal order and verify both citation relations from the exact citing-PDF wording. Resolve the terminal paper’s author-maintained repository and confirm that the clone matches the paper implementation. In the cloned author-maintained repository, inspect the scoring function in src/scoring.py (lines 40–107): it shuffles features and labels ten times, fits one-vs-rest logistic regression, and averages Macro F1; follow this evaluation path directly for the CPU reproduction below. On CPU, carry out the Table 2 reproduction: run the three backbone variants over the table’s three dataset columns with the paper’s listed settings and ten classification shuffles, and report all nine Macro F1 cells; do not substitute expected values for measured results.
终点答案与规范仓库
HARP: Hierarchical Representation Learning for Networks
论文 · GitHub
稳定标识符：1706.07845
HEAD：72a13bf05ca58095f72574bb7adfc2718ae4d029
代码锚点
experiment_code src/scoring.py:40–107
Multi-label Classification paragraph in the Experiment section
在克隆的作者维护仓库中，检查 src/scoring.py 的 scoring 函数（第 40–107 行）：它将特征和标签打乱十次，训练 one-vs-rest 逻辑回归，并平均 Macro F1；沿这条评测路径直接进行下面的 CPU 复现。
查看两跳论文链与逐边 PDF 证据
- 2023 · A Provable Framework of Learning Graph Embeddings via Summarization
https://shenghua-liu.github.io/assets/pdf/papers/aaai2023-gelsum.pdf - 2020 · GraphZoom: A multi-level spectral approach for accurate and scalable graph embedding
https://arxiv.org/pdf/1910.02370 - 2018 · HARP: Hierarchical Representation Learning for Networks
https://arxiv.org/pdf/1706.07845v2.pdf
关键讨论 Related Work, Embedding learning by summarization
一篇 AAAI-23 的图嵌入论文开篇提出两个明确问题：能否从更小的摘要图学习节点嵌入，以及原图与摘要图上的嵌入有何关系？它在配置模型重构图下，为 DeepWalk、LINE 和 GCN 给出闭式分析，并用线性恢复映射连接两者。在相关工作中，它还具体描述了一个较晚出现的谱多层框架：先融合拓扑与节点属性，反复按高谱相似度合并节点，再用图滤波器细化嵌入。
GraphZoom（Deng 等，2020）额外采用图融合步骤来结合结构信息和特征信息，随后使用基于谱相似度的谱粗化方法合并节点；然后通过图滤波器优化嵌入，以确保特征平滑。
查看英文证据原文
GraphZoom (Deng et al. 2020) employs an extra graph fusion step to combine the structural information and feature information, and then uses a spectral coarsening method to merge nodes based on their spectral similarities. Embeddings are then refined by a graph filter to ensure feature smoothness.
关键讨论 Section 4.1, Experimental Setup (Transductive baseline models)
中间研究的摘要把该方法概括为四个内核——图融合、谱图粗化、图嵌入和嵌入细化；其传导式基线模型小节将该框架与两个更早的多层方法比较，并指出随机游走基线在嵌入质量或可扩展性上的提升。
此外，我们将 GraphZoom 与两个多层图嵌入框架进行比较：HARP（Chen 等，2018）和 MILE（Liang 等，2018）；这些框架据报道在嵌入质量或可扩展性方面改进了 DeepWalk 和 node2vec。
查看英文证据原文
In addition, we compare GraphZoom with two multi-level graph embedding frameworks: HARP (Chen et al., 2018) and MILE (Liang et al., 2018), which have reported improvement over DeepWalk and node2vec in either embedding quality or scalability.
查看回溯与仓库迁移记录
- 论文链 · 通过
2023 summary-graph theory -> 2020 spectral four-kernel framework -> 2018 hierarchical embedding
已用引用方 PDF 的章节、原文和讨论主题核验。 - chain_rewrite · 否决
GraphZoom -> MILE -> HARP interpretation from black-box feedback
已由同一造题会话完成证据核验。 - repository · 通过
https://github.com/GTmac/HARP
已由同一造题会话完成证据核验。 - 代码特征 · 通过
Multi-label Classification evaluation workflow
已核验论文内容与文件、符号和代码行的对应关系。 - mask_retry · 通过
rewrite both prompts with titles, endpoint name, owner/slug, datasets, expected values, and internal labels masked
已由同一造题会话完成证据核验。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：9/9；严格命中：0/9。 共审计 4 次独立尝试，当前展示最强有效尝试：corrected_babi_harp_kerr。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 8 项 · Assistant 38 轮 · 工具调用 36 次 · 回溯/审计 5 条 · 控制器尝试 2 次
第三阶段需要短 Rewrite：当前中文 499 个可见字符，目标 180–360；英文 199 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
某 IJCAI 论文研究错位红外/可见光融合，将跨模态生成与多级配准结合起来：估计畸变图像和伪红外图像之间的位移场，并重建配准图像；作者单位之一是 Dalian University of Technology。在第3.3节的定量与定性比较中，RFN被列为与该融合框架对照的现有方法之一。随后，一篇发表于 Information Fusion（2021）的论文在引言中讨论更早的方法：在引言第1节，作者说明预训练 ResNet-50 直接提取深层特征，并用 ZCA 与 l1 范数得到决策图。沿这两条有实质内容的引用边追踪代码端点。论文第3.2节公式（5）对应仓库文件 `whitening_norm.m` 中的符号 `l1_norm`，第11–27行；该实现用带填充的局部窗口对各通道绝对特征取平均。根据这些线索确定端点论文的完整标题、稳定标识符和作者维护的规范 GitHub 仓库。请直接使用上述仓库线索完成指定的 CPU 复现：在 CPU 上运行 Fig.9 的(1)–(8)全部八个图像对，在 Section 4.5 Objective Evaluation 的 Table 7 最末 Conv5 子列读取 EPIa，并逐行报告八个实测值，不使用隐藏的参考数值。
查看英文题面
An IJCAI paper studies misaligned infrared/visible fusion by combining a cross-modality generation step with multi-level registration: it predicts a displacement field between distorted and pseudo-infrared images and reconstructs a registered image; one affiliation is Dalian University of Technology. In Section 3.3, the quantitative and qualitative comparison lists RFN among the existing methods evaluated against the proposed fusion framework. A 2021 Information Fusion paper then discusses the older method in its Introduction: In Section 1, the authors say that pre-trained ResNet-50 extracts deep features directly and that ZCA with l1-norm obtains the decision map. Follow these two substantive citation edges to the code-bearing endpoint. Section 3.2, Equation (5), maps to the repository file `whitening_norm.m`, symbol `l1_norm`, lines 11–27; it applies a padded local window and averages channelwise absolute features. Identify the endpoint's exact title, stable identifier, and canonical author-maintained GitHub repository. Use that repository clue directly for the specified CPU reproduction: on CPU, reproduce the selected Table 7 result by running all eight Fig.9 image-pair cases (1)–(8), reading EPIa from the final Conv5 subcolumn in Section 4.5 Objective Evaluation, and reporting every measured value row by row without using hidden reference values.
终点答案与规范仓库
Infrared and Visible Image Fusion with ResNet and zero-phase component analysis
论文 · GitHub
稳定标识符：arXiv:1806.07119
HEAD：5a60e487bb4ec345827c6d8905486d20431a134a
代码锚点
formula_code whitening_norm.m:11–27
Section 3.2, Equation (5), page 8
论文第3.2节公式（5）对应仓库文件 `whitening_norm.m` 中的符号 `l1_norm`，第11–27行；该实现用带填充的局部窗口对各通道绝对特征取平均。
查看两跳论文链与逐边 PDF 证据
- 2022 · Unsupervised Misaligned Infrared and Visible Image Fusion via Cross-Modality Image Generation and Registration
https://www.ijcai.org/proceedings/2022/0487.pdf - 2021 · RFN-Nest: An end-to-end residual fusion network for infrared and visible images
https://arxiv.org/pdf/2103.04286 - 2019 · Infrared and Visible Image Fusion with ResNet and zero-phase component analysis
https://arxiv.org/pdf/1806.07119v7.pdf
关键讨论 Section 3.3, Comparing with SOTA in IVIF
在第3.3节的定量与定性比较中，RFN被列为与该融合框架对照的现有方法之一。
我们对所提出的方法与若干现有先进红外可见光图像融合方法进行定量和定性评价，其中包括 DENSE、DIDFuse、FGAN、GAN-FM、MFEIF、RFN 和 U2F；这些方法配合 FlowNet+STN 与 VoxelMorph，这两者是本文用于跨模态图像配准的典型算法。
查看英文证据原文
We perform quantitative and qualitative evaluations for the proposed method against several state-of-the-art IVIF methods, including DENSE [Li and Wu, 2019], DIDFuse [Zhao et al., 2020], FGAN [Ma et al., 2019], GAN-FM [Zhang et al., 2021a], MFEIF [Liu et al., 2021a], RFN [Li et al., 2021], and U2F [Xu et al., 2022], equipped with FlowNet+STN and VoxelMorph, two typical image alignment algorithms used to cross-modality image registration in our work.
关键讨论 Section 1, Introduction
在引言第1节，作者说明预训练 ResNet-50 直接提取深层特征，并用 ZCA 与 l1 范数得到决策图。
文献[23]直接使用预训练的 ResNet-50 从源图像提取深层特征，并通过零相位分量分析（ZCA）和 l1 范数得到决策图。
查看英文证据原文
In [23], the pre-trained ResNet-50 [36] is utilized to extract deep features from the source images directly. A decision map is obtained by zero-phase component analysis(ZCA) and l1 -norm.
查看回溯与仓库迁移记录
- 规范仓库 · rejected_alias
exceptionLi/imagefusion_resnet50
已由同一造题会话完成证据核验。 - 规范仓库 · 通过
hli1221/imagefusion_resnet50
已克隆并核对 origin、HEAD 与论文实验内容。 - 论文链 · 通过
IJCAI 2022 -> Information Fusion 2021 -> endpoint
已用引用方 PDF 的章节、原文和讨论主题核验。 - Mask 与改写 · rejected_leak
prior prompt containing an internal pipeline label
已由同一造题会话完成证据核验。 - Mask 与改写 · accepted_mask
endpoint title, owner, slug, datasets, method label, expected values
已由同一造题会话完成证据核验。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：8/8；严格命中：0/8。 共审计 4 次独立尝试，当前展示最强有效尝试：cpu_recovery。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 10 项 · Assistant 69 轮 · 工具调用 66 次 · 回溯/审计 5 条 · 控制器尝试 3 次
第三阶段需要短 Rewrite：当前中文 470 个可见字符，目标 180–360；英文 211 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
在线投资组合更新把约束步写成沿居中价格相对量移动后投影回单纯形；请沿以下两条有实质内容的引文线索定位终点。其引言报告：深度强化学习在一年内预测12种加密货币时击败了均匀买入并持有策略。被引的投资组合管理研究把任务定义为在金融产品间分配资金，并说明网络不限于加密货币；其结果部分将CNN智能体与三种投资组合算法比较，并指出累积收益落后于被引的在线策略。请据此给出终点论文的准确标题、稳定标识符（DOI或arXiv）和作者维护的官方GitHub仓库。论文第4.2节的第一个约束更新（公式(2)）由仓库 `Strategy/` 目录中的 MATLAB expert-update 例程第23–31行实现，其中调用 `simplex_projection`；说明损失、居中的价格相对量步长与单纯形投影如何对应，并在同一仓库基础上直接进行 CPU 复现：以 variant 0、epsilon=0.5、初始财富 S0=1 运行完整序列，报告表4（第5.4.1节）所选策略行的六个原始数据列及反转区块前四列的十个 cumulative-wealth 单元格、各自表格坐标和观测值。
查看英文题面
A portfolio update writes its constrained step as motion along centered price relatives followed by projection onto the simplex. Trace these two substantive citation clues. Its Introduction reports that deep reinforcement learning beat uniform buy-and-hold when predicting 12 cryptocurrencies over one year. The cited portfolio-management study defines portfolio management as allocating funds among financial products and notes that its network is not limited to cryptocurrency. In its Results section, the CNN agent is compared with three portfolio algorithms and loses to the cited online strategy in cumulative return. Identify the terminal paper by its exact title, a stable identifier (DOI or arXiv), and the author-maintained official GitHub repository. The paper's Section 4.2 first constrained update (Equation (2)) is implemented in the repository's MATLAB expert-update routine under `Strategy/`, lines 23–31, including the `simplex_projection` call; explain how the loss, centered price-relative step, and simplex projection map to code, then continue directly with the CPU task. On CPU, use this repository to reproduce the specified run: execute variant 0 with epsilon=0.5 and initial wealth S0=1 over the full sequences, then report the ten cumulative-wealth cells in Table 4 (Section 5.4.1)—the six original-data columns and the first four reversed-data columns—with each table coordinate and observed value.
终点答案与规范仓库
PAMR: Passive aggressive mean reversion strategy for portfolio selection
论文 · GitHub
稳定标识符：doi:10.1007/s10994-012-5281-z
HEAD：9120783cd59a7966b0f78e2b5668030a4378b8af
代码锚点
formula_code Strategy/pamr_expert.m:23–31
Section 4.2, Equation (2), printed p. 231
论文第4.2节的第一个约束更新（公式(2)）由仓库 `Strategy/` 目录中的 MATLAB expert-update 例程第23–31行实现，其中调用 `simplex_projection`
查看两跳论文链与逐边 PDF 证据
- 2018 · Anticipating Cryptocurrency Prices Using Machine Learning
https://arxiv.org/pdf/1805.08550 - 2017 · Cryptocurrency portfolio management with deep reinforcement learning
https://arxiv.org/pdf/1612.01277 - 2012 · PAMR: Passive aggressive mean reversion strategy for portfolio selection
https://link.springer.com/content/pdf/10.1007/s10994-012-5281-z.pdf
关键讨论 Introduction
其引言报告：深度强化学习在一年内预测12种加密货币时击败了均匀买入并持有策略。
深度强化学习被指出在一年期间预测12种加密货币价格时击败了均匀买入并持有策略。
查看英文证据原文
Deep reinforcement learning was showed to beat the uniform buy and hold strategy [47] in predicting the prices of 12 cryptocurrencies over one year period [48].
关键讨论 Section VII, subsection A (Results)
其结果部分将CNN智能体与三种投资组合算法比较，并指出累积收益落后于被引的在线策略。
图2、图4和表III的结果表明，我们的CNN智能体表现优于大多数基准及其他比较算法，但在累积收益方面仅输给被引的被动攻击性均值反转方法。
查看英文证据原文
The result in the Fig. 2, 4 and Table III shows that the performance of our CNN agent outperforms most of the benchmarks and other compared algorithms, only losing to Passive Aggressive Mean Reversion, in term of accumulative return.
查看回溯与仓库迁移记录
- 论文链 · 通过
2018 cryptocurrency-price study -> 2017 deep-reinforcement portfolio study -> 2012 terminal study
已用引用方 PDF 的章节、原文和讨论主题核验。 - repository · replaced
Marigold/universal-portfolios
已由同一造题会话完成证据核验。 - repository · 通过
OLPS/OLPS
已由同一造题会话完成证据核验。 - 代码特征 · 通过
Strategy/pamr_expert.m lines 23-31
已核验论文内容与文件、符号和代码行的对应关系。 - mask · replaced
prior questions containing the endpoint acronym and internal stage label
已由同一造题会话完成证据核验。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：10/10；严格命中：10/10。 共审计 4 次独立尝试，当前展示最强有效尝试：initial。
造题强校验通过 严格 Mask 通过 验题通过
Stage2.9 指标 7 项 · Assistant 54 轮 · 工具调用 52 次 · 回溯/审计 5 条 · 控制器尝试 2 次
第三阶段需要短 Rewrite：当前中文 498 个可见字符，目标 180–360；英文 214 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
候选补丁会先被编译并接受测试：端点论文的 Section IV-D 描述了验证流程，它优先完全成功的补丁，并保留能减少失败而不引入新失败的最佳部分成功补丁。请从一条反向引用路径识别这篇论文。较新的 API 误用自动修复实证研究在 Section 4.3 明确采用补丁候选数量作为修复工具的效率指标；沿该引用到 2020 年 ICSE 的中间研究，其中 Section 3.3 讨论用语义相似性判断生成补丁正确性，并根据作为研究资料提供的人工标注补丁建立正确性分类。沿这两个实质语义关系到达端点后，请给出其完整标题、公开论文 URL、稳定标识符和规范 GitHub URL。请在克隆的规范仓库中定位名为 testGeneratedPatches 的 Java 修复例程会编译候选补丁、重新运行失败测试，并区分完全成功与部分成功修复，给出仓库相对路径、符号和行范围。识别端点后，直接使用该仓库执行 CPU 复现：在 Table VIII 的六个项目列采用正常故障定位 APR 配置和完整的可疑语句排序列表，报告 Fully Fixed 行每个 x/y 单元格的分子及 P*(%) 汇总列的 plausible-patch correctness 百分比，并返回全部七个结果记录。
查看英文题面
A patch-validation workflow first compiles candidates and tests them; Section IV-D of the endpoint prioritizes the first fully fixing patch and retains the best partially fixing one when it reduces failures without adding new failures. Identify that endpoint through two substantive citation links. In Section 4.3, a newer empirical study of automated repair for API misuses explicitly adopts the number of patch candidates as an efficiency measure; follow that citation to the 2020 ICSE middle study. Its Section 3.3 discusses semantic-similarity rules for judging generated-patch correctness and builds its correctness taxonomy from manually labeled patches supplied as research artifacts. Follow this semantic path to the endpoint and give its exact title, public paper URL, stable identifier, and canonical GitHub URL. In the cloned canonical repository, locate the Java fixer routine named testGeneratedPatches that compiles candidates, reruns failing tests, and distinguishes fully from partially successful fixes; provide the repository-relative path, symbol, and line range. After identifying the endpoint, use that repository directly for the CPU reproduction: run the normal fault-localization APR configuration with the ranked full suspicious-statement list across all six project columns in Table VIII, report the numerator in each Fully Fixed x/y cell and the aggregate plausible-patch correctness percentage in the P*(%) column, and return all seven requested result records.
终点答案与规范仓库
AVATAR : Fixing Semantic Bugs with Fix Patterns of Static Analysis Violations
论文 · GitHub
稳定标识符：1812.07270
HEAD：49389fd819cb0a10f33cacc257c9de5a137b8155
代码锚点
section_code src/main/java/edu/lu/uni/serval/avatar/AbstractFixer.java:227–336
Section IV-D, Patch Validation
名为 testGeneratedPatches 的 Java 修复例程会编译候选补丁、重新运行失败测试，并区分完全成功与部分成功修复
查看两跳论文链与逐边 PDF 证据
- 2021 · Evaluating Automatic Program Repair Capabilities to Repair API Misuses
https://discovery.ucl.ac.uk/10124292/1/api_repair_tools_study_camera_ready_submitted.pdf - 2020 · On the efficiency of test suite based program repair
https://arxiv.org/pdf/2008.00914 - 2019 · AVATAR : Fixing Semantic Bugs with Fix Patterns of Static Analysis Violations
https://arxiv.org/pdf/1812.07270v3.pdf
关键讨论 Section 4.3, RQ2: How efficient is a repair tool when it automatically repairs an API misuse?
采用补丁候选数量作为修复工具的效率指标
其次，我们采用补丁候选数量（NPC）作为效率指标，该指标最近由 Liu 等人[10]提出。
查看英文证据原文
Second, we use as an efficiency measure the number of patch candidates (NPC) as recently introduced by Liu et al. [10].
关键讨论 Section 3.3.3, Patch Validation (within Experiment Settings)
根据作为研究资料提供的人工标注补丁建立正确性分类
这一分类体系基于 ACS、SimFix、AVATAR、FixMiner、kPAR 和 TBar 生成的补丁；这些工具的作者研究了正确性，并将人工标注的补丁作为研究资料提供。
查看英文证据原文
This taxonomy is based on the patches generated by ACS, SimFix, AVATAR, FixMiner, kPAR, and TBar whose authors investigated correctness and provided their manually labeled patches as research artifacts.
查看回溯与仓库迁移记录
- admission · 通过
Stage2.9 Table VIII contract
已由同一造题会话完成证据核验。 - 论文链 · 通过
2021 API-misuse study -> 2020 efficiency study -> endpoint paper
已用引用方 PDF 的章节、原文和讨论主题核验。 - repository · replaced
SerVal-Repair/AVATAR
已由同一造题会话完成证据核验。 - feature · 通过
Section IV-D to AbstractFixer.testGeneratedPatches
已由同一造题会话完成证据核验。 - mask_and_rewrite · 通过
Bilingual public prompts after retry leak audit
已由同一造题会话完成证据核验。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：0/7；严格命中：0/7。 共审计 4 次独立尝试，当前展示最强有效尝试：cpu_recovery。
造题强校验通过 严格 Mask 通过 验题有结果
Stage2.9 指标 10 项 · Assistant 64 轮 · 工具调用 62 次 · 回溯/审计 4 条 · 控制器尝试 2 次
第三阶段需要短 Rewrite：当前中文 513 个可见字符，目标 180–360；英文 247 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
该研究用 Mino 时间把径向和极向运动解耦，并把长时段的极端质量比波形拼成多个声部；文中明确说，某个表示“由 van de Meent 提供得特别简洁”，并在附录总结其公式。随后，一项研究 Kerr 时空中的自旋平行输运在引言写道，Fujita 和 Hikida 于 2009 年推导了 Kerr 时空有界测地线的闭式解析解，并将其方法扩展到 Marck 方程；该文第二节还先重述这些测地线，第三节才应用 Marck 形式。在克隆的规范仓库中，检查 stable.py 的 StableOrbit.fundamental_frequencies（第165–188行）：它取出 Υr、Υθ、Υφ、Γ，并返回各 Υ/Γ，这正对应式(9)。请使用该克隆仓库和随附论文执行 CPU 复现，严格计算 Section 3.4/Table 1 的这十个单元：e=0.01、θ=20° 的 Ωr；e=0.01、θ=45° 的 Ωθ；e=0.01、θ=70° 的 Ωφ；e=0.05、θ=20° 的 Ωr；e=0.05、θ=45° 的 Ωθ 和 Ωφ；e=0.05、θ=70° 的 Ωφ；e=0.09、θ=20° 的 Ωr；e=0.09、θ=45° 的 Ωθ；以及 e=0.09、θ=70° 的 Ωφ；报告每个实际观测指标，不要根据参考值调参。
查看英文题面
The study uses Mino time to decouple radial and polar motion and builds long extreme-mass-ratio-inspiral waveforms by stitching snapshots into multiple voices, including a generic inclined and eccentric example. The waveform study says that a particularly clean representation is provided by van de Meent and that it summarizes his formulas in Appendix A. A later study of parallel transport in Kerr spacetime states that closed form analytic solutions for bound geodesics in Kerr spacetime were derived by Fujita and Hikida in 2009, and that it extends their method to Marck’s equation; its Section 2 first reviews those geodesics, while Section 3 applies Marck’s formalism. In the cloned canonical repository, inspect stable.py's StableOrbit.fundamental_frequencies method at lines 165-188: it unpacks Upsilon_r, Upsilon_theta, Upsilon_phi, and Gamma and returns each Upsilon divided by Gamma, implementing Equation (9). Using that cloned repository and the bundled paper, run the CPU reproduction for exactly these ten Section 3.4/Table 1 cells: Omega_r at (e=0.01, theta=20 deg), Omega_theta at (e=0.01, theta=45 deg), Omega_phi at (e=0.01, theta=70 deg), Omega_r at (e=0.05, theta=20 deg), both Omega_theta and Omega_phi at (e=0.05, theta=45 deg), Omega_phi at (e=0.05, theta=70 deg), Omega_r at (e=0.09, theta=20 deg), Omega_theta at (e=0.09, theta=45 deg), and Omega_phi at (e=0.09, theta=70 deg); report every observed metric and do not tune inputs to reference values.
终点答案与规范仓库
Analytical solutions of bound timelike geodesic orbits in Kerr spacetime
论文 · GitHub
稳定标识符：0906.1420
HEAD：58dac1d06edaeb6774d8e0775ead2b34bae573e1
代码锚点
formula_code kerrgeopy/stable.py:165–188
Section 3.1, page 6, Equation (9)
在克隆的规范仓库中，检查 stable.py 的 StableOrbit.fundamental_frequencies（第165–188行）：它取出 Υr、Υθ、Υφ、Γ，并返回各 Υ/Γ，这正对应式(9)。
查看两跳论文链与逐边 PDF 证据
- 2021 · Adiabatic waveforms for extreme mass-ratio inspirals via multivoice decomposition in time and frequency
https://arxiv.org/pdf/2102.02713 - 2020 · Analytic solutions for parallel transport along generic bound geodesics in Kerr spacetime
https://arxiv.org/pdf/1906.05090 - 2009 · Analytical solutions of bound timelike geodesic orbits in Kerr spacetime
https://arxiv.org/pdf/0906.1420v2.pdf
关键讨论 Section II, Kerr geodesics (paragraph introducing the inclination parameter)
文中明确说，某个表示“由 van de Meent 提供得特别简洁”，并在附录总结其公式。
Schmidt [29] 首先展示了如何计算一般 Kerr 轨道的 (E, Lz, Q)；van de Meent [32] 提供了一个特别简洁的表示。我们在附录 A 总结他的公式。
查看英文证据原文
Schmidt [29] first showed how to compute (E, Lz, Q) for generic Kerr orbits; a particularly clean representation is provided by van de Meent [32]. We summarize his formulas in Appendix A
关键讨论 Section 1, Introduction
随后，一项研究 Kerr 时空中的自旋平行输运在引言写道，Fujita 和 Hikida 于 2009 年推导了 Kerr 时空有界测地线的闭式解析解，并将其方法扩展到 Marck 方程；该文第二节还先重述这些测地线，第三节才应用 Marck 形式。
Fujita 和 Hikida 于 2009 年推导了 Kerr 时空有界测地线的闭式解析解。本文采用他们的方法，并将其扩展为 Marck 平行输运方程的解。
查看英文证据原文
Closed form analytic solutions for bound geodesics in Kerr spacetime were derived by Fujita and Hikida [19] in 2009. This paper takes their method an extends it to a solution for Marck’s equation for parallel transport.
查看回溯与仓库迁移记录
- 论文链 · 否决
2021 multivoice waveforms -> 2018 gravitational self-force -> 2009 endpoint
PDF 证据不足或只是普通引用，已回溯其他分支。 - 论文链 · 通过
2021 multivoice waveforms -> 2020 parallel transport -> 2009 endpoint
已用引用方 PDF 的章节、原文和讨论主题核验。 - 规范仓库 · 通过
https://github.com/blackholeperturbationtoolkit/kerrgeopy
已克隆并核对 origin、HEAD 与论文实验内容。 - 代码特征 · 通过
Equation (9) -> StableOrbit.fundamental_frequencies
已核验论文内容与文件、符号和代码行的对应关系。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：10/10；严格命中：10/10。 共审计 4 次独立尝试，当前展示最强有效尝试：corrected_babi_harp_kerr。
造题强校验通过 严格 Mask 通过 验题通过
Stage2.9 指标 10 项 · Assistant 46 轮 · 工具调用 45 次 · 回溯/审计 3 条
第三阶段需要短 Rewrite：当前中文 604 个可见字符，目标 180–360；英文 260 词，目标 80–160。链路和 Repo Feature 可直接复用。
旧单会话题面（仅校准，不发布）
一项关于公平性测试的综合综述按技术路线组织测试输入生成研究。它把被引方法概括为一种面向深度神经网络的梯度引导生成方法：全局阶段把种子迭代推向决策边界，局部阶段继续用梯度搜索已发现歧视实例的邻域。请沿着这条引用找到该方法论文。
在实验设置中，这项工作又给出了一条非常具体的早期工作线索：对于表格型输入，它从 GitHub 获取被引实现，并将该工作三个局部搜索变体中定向程度最高的一个作为实验基线，因为被引论文报告该变体表现最佳。请继续沿这条基线引用，给出三篇论文的准确标题和稳定标识符，并给出最早论文当前由作者维护的规范 GitHub 仓库 URL。
在该规范仓库的当前克隆中，定位第 97–136 行的 `Local_Perturbation.__call__` 实现，并说明其中的参数选择、带符号扰动、歧视性检查、概率更新与归一化如何落实论文第 4.2 节的局部搜索过程及其概率更新算法；随后仅使用 CPU、以性别敏感参数、阈值 0、δ∈{−1,+1} 且 Δv=Δpr=0.001 的配置复现第 5 节表 3 中定向程度最高的生成器，并报告原样保留的全部十个结果单元格：六个分类器行（Fair SVM、SVM、MLPC、Random Forest、Decision Tree、Ensemble）的“生成输入中歧视性输入所占百分比”，以及按表中顺序前四行的“生成的歧视性输入数量”。请报告实测值而不是验证参考值，保持论文配置不变，并如实说明任何偏差或失败。
查看英文题面
A comprehensive survey of fairness testing organizes test-input generation by technique. It describes the cited method as gradient-guided generation for deep neural networks: the global phase iteratively moves seeds toward the decision boundary, and the local phase continues using gradients to search neighborhoods of discovered discriminatory instances. Follow that citation to the method paper.
In its experimental setup, that paper makes a second, unusually concrete link to older work: it obtains the cited implementation from GitHub and uses that work's most-directed one of three local-search variants as an experimental baseline because the cited paper reports the variant performs best. Follow this baseline citation and identify the exact titles and stable identifiers of all three papers, then give the older paper's current author-maintained canonical GitHub repository URL.
In the canonical repository, locate the `Local_Perturbation.__call__` implementation at lines 97–136 in the current clone and explain how its parameter selection, signed perturbation, discrimination check, probability updates, and normalization implement the local-search procedure in Section 4.2 and its probability-update algorithm; using CPU only, reproduce Section 5, Table 3 for the most-directed generator with the gender-sensitive parameter, threshold 0, δ∈{−1,+1}, and Δv=Δpr=0.001, and report all ten preserved result cells: the percentage of generated inputs that are discriminatory for all six classifier rows (Fair SVM, SVM, MLPC, Random Forest, Decision Tree, and Ensemble), plus the number of discriminatory inputs generated for the first four of those rows in table order. Report observed values rather than verification references, retain the paper's configurations, and state any mismatch or failure honestly.
终点答案与规范仓库
Automated Directed Fairness Testing
论文 · GitHub
稳定标识符：arXiv:1807.00468
HEAD：54909641c178f8fc45a3cba0e1140611cf6a016b
代码锚点
section_code Aequitas_Fully_Directed.py:97–136
Section 4.2 Local Search and Algorithm 4 Aeqitas fully-directed update probability
定位第 97–136 行的 `Local_Perturbation.__call__` 实现
查看两跳论文链与逐边 PDF 证据
- 2024 · Fairness Testing: A Comprehensive Survey and Analysis of Trends
https://arxiv.org/pdf/2207.10223 - 2021 · Automatic Fairness Testing of Neural Classifiers through Adversarial Sampling
https://arxiv.org/pdf/2107.08176 - 2018 · Automated Directed Fairness Testing
https://arxiv.org/pdf/1807.00468v2.pdf
关键讨论 Section 4.1.2 Search-based Test Input Generation, subsection Two-phase search-based techniques
它把被引方法概括为一种面向深度神经网络的梯度引导生成方法：全局阶段把种子迭代推向决策边界，局部阶段继续用梯度搜索已发现歧视实例的邻域。
Zhang 等人提出了 ADF，一种面向深度神经网络的梯度引导生成方法。在全局搜索阶段，ADF 在梯度引导下通过反复将种子输入扰动向决策边界来定位决策边界附近的歧视性实例；在局部搜索阶段，ADF 进一步利用梯度引导搜索已找到的个体歧视性实例的邻域，以发现更多歧视性实例。
查看英文证据原文
Zhang et al. [313, 315] proposed ADF, a gradient-guided generation approach for DNNs. In the global search phase, ADF locates the discriminatory instances near the decision boundary by iteratively perturbing a seed input towards the decision boundary with the guidance of gradient. In the local search phase, ADF further uses gradients as the guidance to search the neighborhood of the found individual discriminatory instances to discover more discriminatory instances.
关键讨论 Section 5.1 Experimental Setup
对于表格型输入，它从 GitHub 获取被引实现，并将该工作三个局部搜索变体中定向程度最高的一个作为实验基线，因为被引论文报告该变体表现最佳。
我们从 GitHub 获取了 AEQUITAS 的实现，并根据文献 [14] 的描述重新实现了 SG，因为其实现并未公开。还要注意，AEQUITAS 提出了三种不同的局部搜索算法；由于文献 [13] 报告其性能最佳，我们在评估中采用了 fully-directed 算法。
查看英文证据原文
We obtained the implementation of AEQUITAS from GitHub4 and re-implemented SG according to the description in [14] since their implementation is not publicly available. Further notice that AEQUITAS proposed 3 different local search algorithms, we adopted the fully-directed algorithm in our evaluation since it has the best performance according to [13].
查看回溯与仓库迁移记录
- 论文链 · 否决
2021 survey -> 2020 empirical benchmark -> 2018 endpoint
PDF 证据不足或只是普通引用，已回溯其他分支。 - 论文链 · 通过
2024 fairness-testing survey -> 2021 adversarial-sampling method -> 2018 endpoint
已用引用方 PDF 的章节、原文和讨论主题核验。 - 代码特征 · 通过
Section 4.2 local-search prose -> Aequitas_Fully_Directed.py Local_Perturbation.__call__
已核验论文内容与文件、符号和代码行的对应关系。
查看历史 K8s 校准结果
论文/仓库身份：正确；两跳链/代码锚点：正确；CPU 指标：0/10；严格命中：0/10。 共审计 4 次独立尝试，当前展示最强有效尝试：initial。
造题强校验通过 严格 Mask 通过 验题有结果