@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.066_tropicalattention.zh.yaml

@@ -0,0 +1,229 @@
+id: aisb.t3.066_tropicalattention
+name: 'Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms'
+version: 0.1.0
+one_line: 基于热带几何的注意力机制，用于组合任务的神经算法推理，在分布外泛化方面表现优异，可超越PTIME扩展至NP难问题。
+task_description: |
+  本基准测试包涵盖Tropical Attention，这是一种基于热带几何的注意力机制，将注意力核提升到热带投影空间中进行推理，其推理过程是分段线性的且满足1-Lipschitz条件。基准测试评估用热带几何结构替代或增强传统softmax注意力是否能改善11个组合任务上的算法泛化能力，涵盖多项式时间算法（Floyd-Warshall、Quickselect、凸包、SCC）、NP完全问题（3SUM、均衡划分、子集和）以及NP难问题（0/1背包装箱、装箱问题、最小硬币找零）。任务在分布外（OOD）条件下进行评估，包括长度外推和数值范围泛化。主要指标是quickselect_length_ood_f1，即Quickselect任务的长度OOD F1分数。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- quickselect_length_ood_f1
+- algorithmic_reasoning_report
+- trained_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: |
+    运行一条封装的神经算法推理评估路线。加载预训练检查点，在所有11个组合任务的OOD测试集上评估。使用--job_file参数指定jobs_to_do_evaluate.csv文件，通过--job_id选择目标任务。输出每个任务的长度OOD和数值OOD F1分数。
+- id: full_train_eval
+  label: 完整训练+评估
+  description: |
+    运行完整的热带注意力训练和OOD推理基准测试工作流。从零开始在每个任务上训练100k样本，使用schedulefree优化器，然后在OOD测试集上评估。训练日志保存到outputs/<timestamp>/目录。使用--job_file参数指定jobs_to_do_train.csv文件，通过--job_id选择任务行。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://deepscientist.cc/AISB/066_tropicalattention
+    type: benchmark_archive
+    archive_type: zip
+    local_dir_name: paper-66-TropicalAttention
+  - url: https://github.com/Baran-phys/Tropical-Attention
+    type: upstream_repository
+  notes:
+  - 压缩包包含完整代码、数据加载器、模型和评估脚本
+  - 数据集在运行时通过dataloaders.py程序化生成
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: 可运行单任务训练；仅评估模式需求更低
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: |
+      推荐用于完整多任务训练运行。单GPU即可，因为参数量比Universal Transformer基线少约20%。
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torch
+  - schedulefree
+  - pandas
+  - matplotlib
+  - scikit-learn
+  notes:
+  - 详见打包的requirements.txt获取完整依赖列表
+  - TropicalAttention.py是标准Transformer块的即插即用模块
+  - 训练需要CUDA支持；推理支持CPU回退
+risk_flags:
+- ood_metric_variance
+- multi_seed_required
+risk_notes:
+- OOD F1分数在不同随机种子间存在方差（例如子集和问题为87.50±6.45）；建议进行多种子评估
+- Floyd-Warshall和分数背包的OOD分数低于分类任务
+- 每个任务训练使用100k样本；需要磁盘空间存储outputs/<timestamp>/下的产物
+recommended_when: |
+  在需要评估神经网络对图和组合问题的推理能力，并关注明确的OOD行为时使用此基准测试。特别适用于：（1）在长度外推任务上评估注意力机制；（2）在PTIME之外对NP难/完全问题求解进行基准测试；（3）将热带几何归纳偏置与softmax变换器进行比较；（4）研究神经算法推理中的多面体决策边界。
+not_recommended_when: |
+  如果需要一个无需专属推理数据集的即插即用应用基准测试，请勿使用。当仅关注分布内准确性即可，或PTIME算法推理已足够时，不适合使用。如果研究问题不涉及组合优化、注意力机制设计或OOD泛化，请避免使用。
+paper:
+  title: 'Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.17190
+  authors:
+  - Hashemi, Baran (Origins Data Science Lab, Technical University of Munich)
+  - Pasque, Kurt (Naval Postgraduate School)
+  - Teska, Chris (Naval Postgraduate School)
+  - Yoshida, Ruriko (Naval Postgraduate School)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.066_tropicalattention.zip
+  archive_type: zip
+  local_dir_name: paper-66-TropicalAttention
+  upstream_url: https://github.com/Baran-phys/Tropical-Attention
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.066_tropicalattention.zip
+  sha256: 59c566c1c49e1649bad3e39f2100133d7cb97cea8df9665ae1dea27bcd4a80db
+  size_bytes: 1618773
+display:
+  palette_seed: lime-violet-tropical
+  art_style: geometric-reasoning
+  accent_priority: high
+image_path: ../image/066_aisb.t3.066_tropicalattention.jpg
+capability_tags:
+- research_code_optimization
+- neural_algorithmic_reasoning
+- combinatorial_optimization
+- transformers
+- mathematical_reasoning
+- tropical_geometry
+- out_of_distribution_generalization
+- attention_mechanisms
+- np_complete_problems
+- np_hard_problems
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null
+experiment_anchors:
+  train_script: experiment.py
+  train_job_csv: jobs_to_do_train.csv
+  eval_job_csv: jobs_to_do_evaluate.csv
+  attention_module: TropicalAttention.py
+  eval_script: eval.py
+  metric_contract: json/metric_contract.json
+task_specifications:
+- task_name: Quickselect
+  type: classification
+  ood_metric: length_ood_f1
+  training_length_range: (10,30)
+  evaluation_length_range: (10,1024)
+  epochs: 100
+  notes: 主要OOD基准；尖锐的注意力图保留第k个元素选择能力
+- task_name: SubsetSumDecision
+  type: classification
+  complexity: NP-complete
+  training_length_range: (8,8)
+  value_range: (-5,5)
+  epochs: 100
+- task_name: Knapsack
+  type: classification
+  complexity: NP-hard
+  training_length_range: (10,20)
+  epochs: 100
+- task_name: BalancedPartition
+  type: classification
+  complexity: NP-complete
+  training_length_range: (10,30)
+  epochs: 1000
+- task_name: BinPacking
+  type: classification
+  complexity: NP-hard
+  training_length_range: (10,30)
+  epochs: 1000
+- task_name: MinCoinChange
+  type: classification
+  training_length_range: (10,20)
+  epochs: 100
+- task_name: ConvexHull
+  type: classification
+  training_length_range: (10,30)
+  epochs: 100
+- task_name: FractionalKnapsack
+  type: regression
+  training_length_range: (10,20)
+  epochs: 100
+  notes: 回归任务预期OOD性能较低
+- task_name: FloydWarshall
+  type: regression
+  training_length_range: (10,30)
+  epochs: 100
+  notes: 全源最短路径；OOD分数低于分类任务
+- task_name: SCC
+  type: classification
+  training_length_range: (10,30)
+  epochs: 100
+- task_name: 3SUM
+  type: classification
+  complexity: NP-complete
+  training_length_range: (10,30)
+  epochs: 100
+execution_notes:
+  training_invocation: 'python experiment.py --job_file jobs_to_do_train.csv --job_id
+    <row_index>
+
+    '
+  evaluation_invocation: 'python experiment.py --job_file jobs_to_do_evaluate.csv
+    --job_id <row_index>
+
+    '
+  job_csv_fields:
+  - task
+  - model_type
+  - dataset_name
+  - input_dim
+  - d_model
+  - n_heads
+  - num_layers
+  - num_epochs
+  - dropout
+  - lr
+  - seed
+  - n_samples
+  - batch_size
+  - length_range
+  - value_range
+  - noise_prob
+  output_directory: outputs/<timestamp>/
+  default_hyperparameters:
+    d_model: 64
+    n_heads: 2
+    num_layers: 1
+    dropout: 0
+    learning_rate: 0.0001
+    batch_size: 500
+    n_samples: 100000
+    seed: 999

package/AISB/catalog/aisb.t3.067_kanad.yaml

@@ -0,0 +1,193 @@
+id: aisb.t3.067_kanad
+name: 'KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks'
+version: 0.1.0
+one_line: Lightweight KAN-based time-series anomaly detection using Fourier basis
+  functions with 15% F1 improvement and 50% faster inference than vanilla KAN.
+task_description: 'This packaged benchmark covers time-series anomaly detection (TSAD)
+  using Kolmogorov-Arnold Networks (KAN) with redesigned univariate functions. The
+  core innovation replaces B-spline basis functions (used in vanilla KAN) with truncated
+  Fourier expansions, addressing KAN''s susceptibility to local disturbances in time
+  series data. The approach reformulates TSAD as approximation of time series with
+  smooth univariate functions, leveraging the insight that normal sequences exhibit
+  greater local smoothness than anomalies. The method incorporates differencing to
+  isolate trend effects and an index-based univariate function to capture fine-scale
+  periodic patterns. The local snapshot currently preserves metadata and audit material
+  rather than runnable code; executable source files require upstream restoration
+  before benchmark execution.
+
+  '
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- f1
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Review the preserved benchmark metadata, paper materials, and TSAD
+    task framing.
+- id: restore_first
+  label: Restore First
+  description: Restore the missing source tree from the public GitHub repository (github.com/CSTCloudOps/KAN-AD)
+    before treating this as a runnable benchmark.
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: github
+    url: https://github.com/CSTCloudOps/KAN-AD
+    description: Public code repository for KAN-AD implementation
+  - type: benchmark
+    description: Four publicly available TSAD benchmark datasets referenced in the
+      paper
+  notes:
+  - The current local package does not preserve an executable benchmark route; dataset
+    setup is part of upstream restoration
+  - Four popular TSAD benchmarks are used in the paper evaluation
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: CPU-only execution is plausible for the minimum analysis route
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: GPU recommended for full evaluation with restored source
+environment:
+  python: null
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages: []
+  notes:
+  - The current local snapshot does not preserve executable setup files or runnable
+    source code
+  - Environment hints here are provisional pending upstream restoration from the public
+    repository
+  - CPU-only execution is plausible for the minimum analysis route
+  - The bundled local folder does not include a README or requirements file
+  - Restore environment details from github.com/CSTCloudOps/KAN-AD
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+risk_notes:
+- The current snapshot is blocked because no executable code files are preserved
+- No benchmark execution was performed in this packaging pass
+- Treat as provisional until source restoration or richer local snapshot is recovered
+recommended_when: 'Use this entry when conducting paper-backed planning, upstream
+  restoration, or analysis of the KAN-based anomaly detection methodology. Suitable
+  for reviewing the TSAD task framing, examining the paper methodology (Fourier basis
+  functions replacing B-splines), or planning restoration from the public GitHub repository.
+
+  '
+not_recommended_when: 'Do not treat this local package as a ready-to-run anomaly benchmark.
+  No executable code is preserved in the staged snapshot. Do not attempt benchmark
+  execution without first restoring the source tree from the public repository.
+
+  '
+paper:
+  title: 'KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/zhou25u.html
+  github_url: https://github.com/CSTCloudOps/KAN-AD
+  authors:
+  - Quan Zhou
+  - Changhua Pei
+  - Fei Sun
+  - Jing Han
+  - Zhengwei Gao
+  - Haiming Zhang
+  - Gaogang Xie
+  - Dan Pei
+  - Jianhui Li
+  affiliations:
+  - Computer Network Information Center, Chinese Academy of Sciences
+  - University of Chinese Academy of Sciences
+  - Hangzhou Institute for Advanced Study
+  - Institute of Computing Technology, Chinese Academy of Sciences
+  - ZTE
+  - Department of Computer Science, Tsinghua University
+  - School of Frontier Sciences, Nanjing University
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.067_kanad.zip
+  archive_type: zip
+  local_dir_name: paper-67-KAN-AD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.067_kanad.zip
+  sha256: 39c46b0ec921cd9342fdeda7a58471f5b5bff4e1f0747f9629424822710e8a96
+  size_bytes: 27708
+display:
+  palette_seed: peach-slate-anomaly
+  art_style: lightweight-analytics
+  accent_priority: medium
+image_path: ../image/067_aisb.t3.067_kanad.jpg
+capability_tags:
+- research_code_optimization
+- anomaly_detection
+- time_series
+- kolmogorov_arnold_networks
+- scientific_ml
+- univariate_tsad
+- fourier_basis
+- lightweight_models
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null
+metric_evidence:
+  primary_metric: f1
+  f1_status: blocked-no-code
+  f1_claim: 'Paper reports 15% average F1 improvement over state-of-the-art baselines
+    on four TSAD benchmarks
+
+    '
+  performance_notes:
+  - Peak improvement exceeds 27% over baselines
+  - 50% faster inference compared to original KAN
+  - Fewer than 1,000 trainable parameters
+  - Note: These claims require verification through restored execution
+technical_summary:
+  problem_domain: Univariate time series anomaly detection (UTS-TSAD)
+  input_format: Univariate time series x = [x_0, x_1, ..., x_t] with anomaly labels
+  core_methodology:
+  - Kolmogorov-Arnold Network backbone with redesigned univariate functions
+  - B-splines replaced with truncated Fourier expansions
+  - Index-based univariate function for fine-scale periodicity
+  - Differencing incorporated to isolate trend effects
+  key_insight: Normal sequences exhibit greater local smoothness than anomalies; reformulating
+    TSAD as smooth function approximation mitigates overfitting to local disturbances
+  three_stage_pipeline:
+  - mapping: Decompose input time window into multiple univariate functions
+  - reducing: Combine functions through learned coefficients to reconstruct normal
+      patterns
+  - projection: Leverage for anomaly scoring
+  evaluation_datasets:
+  - Four publicly available TSAD benchmark datasets (specific names not preserved
+    in snapshot)
+  baselines_compared:
+  - TimesNet
+  - Vanilla KAN
+  - Other forecasting-based TSAD methods
+  - VAE-based approaches (FCVAE)

package/AISB/catalog/aisb.t3.067_kanad.zh.yaml

@@ -0,0 +1,167 @@
+id: aisb.t3.067_kanad
+name: 'KAN-AD：基于Kolmogorov-Arnold网络的时间序列异常检测'
+version: 0.1.0
+one_line: 基于KAN的轻量级时间序列异常检测，采用傅里叶基函数，F1提升15%，推理速度比标准KAN快50%。
+task_description: '该基准测试包涵盖使用Kolmogorov-Arnold网络（KAN）和重新设计的单变量函数进行时间序列异常检测（TSAD）。核心创新是用截断傅里叶展开替换B样条基函数（标准KAN使用），解决了KAN对时间序列数据局部扰动的敏感性。该方法将TSAD重新表述为用平滑单变量函数逼近时间序列，利用正常序列比异常具有更强局部平滑性这一洞察。该方法结合了差分以隔离趋势效应，以及基于索引的单变量函数以捕捉细粒度周期性模式。目前本地快照仅保留元数据和审计材料，而非可运行代码；基准测试执行前需要上游恢复可执行源文件。
+
+  '
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- f1
+launch_profiles:
+- id: analysis_only
+  label: 仅分析
+  description: 审查保存的基准测试元数据、论文资料和TSAD任务框架。
+- id: restore_first
+  label: 先恢复
+  description: 在将此作为可运行基准测试之前，从公共GitHub仓库（github.com/CSTCloudOps/KAN-AD）恢复缺失的源代码树。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: github
+    url: https://github.com/CSTCloudOps/KAN-AD
+    description: KAN-AD实现的公共代码仓库
+  - type: benchmark
+    description: 论文中引用的四个公开可用的TSAD基准测试数据集
+  notes:
+  - 当前本地包未保留可执行的基准测试路径；数据集设置是上游恢复的一部分
+  - 论文评估使用了四个流行的TSAD基准测试
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: CPU-only execution is plausible for the minimum analysis route
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 8
+    notes: GPU recommended for full evaluation with restored source
+environment:
+  python: null
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages: []
+  notes:
+  - 当前本地快照未保留可执行的设置文件或可运行源代码
+  - 此处的环境提示为临时性质，等待从公共仓库进行上游恢复
+  - 最小分析路径支持纯CPU执行
+  - 打包的本地文件夹不包含README或requirements文件
+  - 从github.com/CSTCloudOps/KAN-AD恢复环境详情
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- blocked_no_code
+risk_notes:
+- 由于未保存可执行代码文件，当前快照被阻止
+- 本次打包过程中未执行基准测试
+- 在源代码恢复或获得更完整的本地快照之前，视为临时版本
+recommended_when: '在开展基于论文的规划、上游恢复或分析KAN-based异常检测方法时使用此条目。适用于审查TSAD任务框架、检查论文方法论（用傅里叶基函数替换B样条），或规划从公共GitHub仓库恢复代码。
+
+  '
+not_recommended_when: '不要将本地包视为可立即运行的异常检测基准测试。暂存快照中未保存可执行代码。在未首先从公共仓库恢复源代码树的情况下，不要尝试执行基准测试。
+
+  '
+paper:
+  title: 'KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/zhou25u.html
+  github_url: https://github.com/CSTCloudOps/KAN-AD
+  authors:
+  - Quan Zhou
+  - Changhua Pei
+  - Fei Sun
+  - Jing Han
+  - Zhengwei Gao
+  - Haiming Zhang
+  - Gaogang Xie
+  - Dan Pei
+  - Jianhui Li
+  affiliations:
+  - Computer Network Information Center, Chinese Academy of Sciences
+  - University of Chinese Academy of Sciences
+  - Hangzhou Institute for Advanced Study
+  - Institute of Computing Technology, Chinese Academy of Sciences
+  - ZTE
+  - Department of Computer Science, Tsinghua University
+  - School of Frontier Sciences, Nanjing University
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.067_kanad.zip
+  archive_type: zip
+  local_dir_name: paper-67-KAN-AD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.067_kanad.zip
+  sha256: 39c46b0ec921cd9342fdeda7a58471f5b5bff4e1f0747f9629424822710e8a96
+  size_bytes: 27708
+display:
+  palette_seed: peach-slate-anomaly
+  art_style: lightweight-analytics
+  accent_priority: medium
+image_path: ../image/067_aisb.t3.067_kanad.jpg
+capability_tags:
+- research_code_optimization
+- anomaly_detection
+- time_series
+- kolmogorov_arnold_networks
+- scientific_ml
+- univariate_tsad
+- fourier_basis
+- lightweight_models
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null
+metric_evidence:
+  primary_metric: f1
+  f1_status: blocked-no-code
+  f1_claim: '论文报告在四个TSAD基准测试上相比最先进方法平均F1提升15%
+
+    '
+  performance_notes:
+  - 相比基线峰值提升超过27%
+  - 相比原始KAN推理速度快50%
+  - 可训练参数少于1000个
+  - 注意：这些声明需要通过恢复执行进行验证
+technical_summary:
+  problem_domain: 单变量时间序列异常检测（UTS-TSAD）
+  input_format: 带异常标签的单变量时间序列 x = [x_0, x_1, ..., x_t]
+  core_methodology:
+  - 基于Kolmogorov-Arnold网络骨干和重新设计的单变量函数
+  - 用截断傅里叶展开替换B样条
+  - 基于索引的单变量函数用于细粒度周期性
+  - 结合差分以隔离趋势效应
+  key_insight: 正常序列比异常具有更强的局部平滑性；将TSAD重新表述为平滑函数逼近可减轻对局部扰动的过拟合
+  three_stage_pipeline:
+  - mapping: 将输入时间窗口分解为多个单变量函数
+  - reducing: 通过学习系数组合函数以重构正常模式
+  - projection: 用于异常评分
+  evaluation_datasets:
+  - 四个公开可用的TSAD基准测试数据集（快照中未保留具体名称）
+  baselines_compared:
+  - TimesNet
+  - Vanilla KAN
+  - 其他基于预测的TSAD方法
+  - 基于VAE的方法（FCVAE）