@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.090_rstib.yaml

@@ -0,0 +1,144 @@
+id: aisb.t3.090_rstib
+name: Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting
+version: 0.1.0
+one_line: Teacher-student training of information-bottleneck-guided MLPs for robust
+  and efficient spatial-temporal forecasting.
+task_description: 'This packaged benchmark covers robust spatial-temporal forecasting
+  using the Robust Spatial-Temporal Information Bottleneck (RSTIB) principle and a
+  knowledge distillation training regime. The approach addresses the dual noise effect
+  in spatial-temporal data by first training a teacher model (STGCN), then distilling
+  its knowledge into an efficient RSTIB-MLP student. The goal is to achieve robustness
+  against noise perturbations while maintaining the computational efficiency of simple
+  MLP architectures compared to expensive graph neural networks.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- mae
+- rmse
+- mape
+- student_checkpoint
+launch_profiles:
+- id: teacher_student_train
+  label: Teacher-Student Train
+  description: 'Execute the full two-stage pipeline. First run run_teacher.sh to train
+    the STGCN teacher model, then run run_student.sh to train the RSTIB-MLP student
+    with knowledge distillation from the teacher. Training scripts use config files
+    in config/teacher/ and config/student/ respectively.
+
+    '
+- id: student_eval
+  label: Student Evaluation
+  description: 'Run evaluation of the trained RSTIB-MLP student model on held-out
+    test data. Uses train_student.py with saved checkpoint from the training profile.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: bundled
+    description: Preprocessed datasets including PEMS04 in data/ directory, loaded
+      via DataHandler_st.py
+  - type: external
+    description: PyG precompiled wheels for torch_scatter and torch_sparse (cp39-cp39-linux_x86_64)
+      required for graph operations
+  notes:
+  - PEMS04 traffic dataset is bundled in data/PEMS04/
+  - Precompiled PyG wheel files (torch_scatter-2.0.9, torch_sparse-0.6.16+pt113cu116)
+    must be installed separately
+  - See requirements.txt for additional Python dependencies
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: Minimum for running teacher or student training on PEMS04 dataset
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: Recommended for full teacher-student pipeline with larger batch sizes
+environment:
+  python: '3.9'
+  cuda: '11.7'
+  pytorch: 1.13.1
+  flash_attn: null
+  key_packages:
+  - torch==1.13.1
+  - torchvision==0.14.1
+  - torchaudio==0.13.1
+  - torch-scatter
+  - torch-sparse
+  - numpy
+  notes:
+  - Install PyTorch with CUDA 11.7 support: conda install pytorch==1.13.1 pytorch-cuda=11.7
+      -c pytorch -c nvidia
+  - Install precompiled PyG wheels: torch_scatter-2.0.9-cp39-cp39-linux_x86_64.whl
+      and torch_sparse-0.6.16+pt113cu116-cp39-cp39-linux_x86_64.whl
+  - See bundled requirements.txt for full dependency list
+risk_flags:
+- two_stage_training
+- gpu_memory_constrained
+- no_runtime_verification
+risk_notes:
+- Two-stage pipeline requires running teacher training before student training can
+  execute
+- Student model uses graph-based adjacency matrix processed via PyG; GPU memory requirements
+  may vary with spatial adjacency size
+- Static code audit confirmed executable anchors but no benchmark execution was performed
+  in packaging pass; runtime verification required before trusting metric values
+- MAPE metric is code-backed in new_metrics.py but not listed in current primary_outputs
+recommended_when: 'Use this benchmark when you want to study MLP efficiency for spatial-temporal
+  forecasting with explicit robustness to noise perturbations. The information bottleneck
+  regularization and teacher-student framework are suitable for researchers investigating
+  knowledge distillation from graph networks to simpler architectures while preserving
+  noise resistance.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need a purely graph-based
+  approach (no student MLP distillation), a probabilistic forecasting benchmark, or
+  a lightweight benchmark without GPU requirements.
+
+  '
+paper:
+  title: Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting
+  authors:
+  - Min Chen
+  - Guansong Pang
+  - Wenjun Wang
+  - Cheng Yan
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/chen25ar.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.090_rstib.zip
+  archive_type: zip
+  local_dir_name: paper-90-RSTIB
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.090_rstib.zip
+  sha256: 501c6d17e5bccd9cc64d8490c371ad914938d355be22814134f054566f302bea
+  size_bytes: 37796454
+display:
+  palette_seed: plum-cyan-bottleneck
+  art_style: robust-forecast
+  accent_priority: high
+image_path: ../image/090_aisb.t3.090_rstib.jpg

package/AISB/catalog/aisb.t3.090_rstib.zh.yaml

@@ -0,0 +1,109 @@
+id: aisb.t3.090_rstib
+name: 用于鲁棒时空预测的信息瓶颈引导 MLP
+version: 0.1.0
+one_line: 信息瓶颈引导 MLP 的师生培训，以实现稳健、高效的时空预测。
+task_description: 该打包基准涵盖使用鲁棒时空信息瓶颈 (RSTIB) 原理和知识蒸馏训练机制进行鲁棒时空预测。该方法首先训练教师模型 (STGCN)，然后将其知识提炼成高效的 RSTIB-MLP 学生，从而解决时空数据中的双重噪声效应。目标是实现对噪声扰动的鲁棒性，同时保持简单 MLP 架构与昂贵的图神经网络相比的计算效率。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- mae
+- rmse
+- mape
+- student_checkpoint
+launch_profiles:
+- id: teacher_student_train
+  label: 师生列车
+  description: 执行完整的两阶段管道。首先运行 run_teacher.sh 来训练 STGCN 教师模型，然后运行 ​​run_student.sh 通过教师的知识蒸馏来训练 RSTIB-MLP 学生。训练脚本分别使用 config/teacher/ 和 config/student/ 中的配置文件。
+- id: student_eval
+  label: 学生评价
+  description: 根据保留的测试数据对经过训练的 RSTIB-MLP 学生模型进行评估。使用 train_student.py 和训练配置文件中保存的检查点。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: bundled
+    description: Preprocessed datasets including PEMS04 in data/ directory, loaded via DataHandler_st.py
+  - type: external
+    description: PyG precompiled wheels for torch_scatter and torch_sparse (cp39-cp39-linux_x86_64) required for graph operations
+  notes:
+  - PEMS04流量数据集捆绑在data/PEMS04/中
+  - 预编译的PyG轮文件（torch_scatter-2.0.9、torch_sparse-0.6.16+pt113cu116）必须单独安装
+  - 请参阅 requirements.txt 了解其他 Python 依赖项
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: 在 PEMS04 数据集上运行教师或学生培训的最低要求
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 推荐用于具有较大批量的完整师生管道
+environment:
+  python: '3.9'
+  cuda: '11.7'
+  pytorch: 1.13.1
+  flash_attn: null
+  key_packages:
+  - torch==1.13.1
+  - torchvision==0.14.1
+  - torchaudio==0.13.1
+  - torch-scatter
+  - torch-sparse
+  - numpy
+  notes:
+  - Install PyTorch with CUDA 11.7 support: conda 安装 pytorch==1.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
+  - Install precompiled PyG wheels: torch_scatter-2.0.9-cp39-cp39-linux_x86_64.whl 和 torch_sparse-0.6.16+pt113cu116-cp39-cp39-linux_x86_64.whl
+  - 请参阅捆绑的 requirements.txt 以获取完整的依赖项列表
+risk_flags:
+- two_stage_training
+- gpu_memory_constrained
+- no_runtime_verification
+risk_notes:
+- 两阶段管道需要先运行教师培训，然后才能执行学生培训
+- 学生模型使用通过 PyG 处理的基于图的邻接矩阵； GPU 内存要求可能会因空间邻接大小而异
+- 静态代码审计确认了可执行锚点，但在打包过程中未执行基准测试；在信任指标值之前需要运行时验证
+- MAPE 指标在 new_metrics.py 中由代码支持，但未在当前 primary_outputs 中列出
+recommended_when: 当您想要研究时空预测的 MLP 效率以及对噪声扰动的显式鲁棒性时，请使用此基准。信息瓶颈正则化和师生框架适合研究人员研究从图网络到更简单的架构的知识蒸馏，同时保持抗噪性。
+not_recommended_when: 如果您需要纯粹基于图形的方法（无学生 MLP 蒸馏）、概率预测基准或没有 GPU 要求的轻量级基准，请不要使用此基准。
+paper:
+  title: Information Bottleneck-guided MLPs for Robust Spatial-temporal Forecasting
+  authors:
+  - Min Chen
+  - Guansong Pang
+  - Wenjun Wang
+  - Cheng Yan
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/chen25ar.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.090_rstib.zip
+  archive_type: zip
+  local_dir_name: paper-90-RSTIB
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.090_rstib.zip
+  sha256: 501c6d17e5bccd9cc64d8490c371ad914938d355be22814134f054566f302bea
+  size_bytes: 37796454
+display:
+  palette_seed: plum-cyan-bottleneck
+  art_style: robust-forecast
+  accent_priority: high
+image_path: ../image/090_aisb.t3.090_rstib.jpg

package/AISB/catalog/aisb.t3.091_timeawarecausal.yaml

@@ -0,0 +1,132 @@
+id: aisb.t3.091_timeawarecausal
+name: Learning Time-Aware Causal Representation for Model Generalization in Evolving
+  Domains
+version: 0.1.0
+one_line: Evolving-domain generalization via time-aware causal representation learning
+  using a sequential VAE framework that mitigates spurious correlations.
+task_description: 'This packaged benchmark covers evolving-domain generalization (EDG)
+  through SYNC (Static-DYNamic Causal Representation Learning), a time-aware causal
+  approach that learns representations robust to temporal distribution shift. The
+  method introduces a time-aware structural causal model (SCM) that explicitly models
+  dynamic causal factors and causal mechanism drifts. A sequential VAE framework captures
+  evolving patterns, while information-theoretic objectives preserve intra-class compactness
+  of causal factors across and within temporal domains. The approach theoretically
+  guarantees optimal causal predictors per time domain.
+
+  '
+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:
+- average_accuracy
+- causal_representation_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: Run one packaged SYNC evaluation route on a single evolving-domain
+    dataset from the bundled suite.
+- id: full_train_eval
+  label: Full Train + Eval
+  description: Run the full time-aware causal representation learning workflow end
+    to end, covering all bundled datasets.
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://github.com/WonderSeven/LSSAE
+    note: Primary upstream source for bundled datasets
+  - url: https://deepscientist.cc/AISB/091_timeawarecausal
+    note: Local archive with prepackaged data splits
+  notes:
+  - 'Bundled datasets: Circle, Sine (synthetic), RMNIST, Portraits, Caltran, PowerSupply,
+    ONP'
+  - 'Caltran: vehicle presence detection with temporal lighting confounds (day/night
+    spurious correlations)'
+  - 'RMNIST: rotated MNIST with concept drift across temporal domains'
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+  notes:
+  - Tested on NVIDIA GeForce RTX 4090
+  - Single GPU is sufficient; multi-GPU training is not required
+environment:
+  python: 3.8.18
+  cuda: '11.6'
+  pytorch: 1.12.1
+  flash_attn: null
+  key_packages:
+  - torch==1.12.1
+  - torchvision==0.13.1
+  - torchaudio==0.12.1
+  - DomainBed
+  - LSSAE
+  notes:
+  - Full dependency set in bundled requirements.txt
+  - 'Install via: conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1
+    cudatoolkit=11.6 -c pytorch -c conda-forge'
+risk_flags:
+- spurious_correlation_risk
+- temporal_drift_risk
+risk_notes:
+- Caltran contains lighting-based spurious correlations that must be explicitly mitigated
+- Model relies on learning time-aware causal features rather than shortcut features
+recommended_when: 'Use this benchmark when you need a causal-representation approach
+  for evolving domains, spurious-correlation mitigation in temporal distribution shift,
+  or benchmark EDG methods with causal guarantees.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need a small static-domain
+  generalization task, a ready-to-use image classification benchmark without temporal
+  structure, or methods that do not require causal disentanglement.
+
+  '
+paper:
+  title: Learning Time-Aware Causal Representation for Model Generalization in Evolving
+    Domains
+  authors:
+  - Zhuo He
+  - Shuang Li
+  - Wenze Song
+  - Longhui Yuan
+  - Jian Liang
+  - Han Li
+  - Kun Gai
+  venue: ICML 2025
+  year: 2025
+  url: https://openreview.net/forum?id=KhCKypSaqx
+  note: Kuaishou Technology; PMLR 267
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.091_timeawarecausal.zip
+  archive_type: zip
+  local_dir_name: paper-91-TimeAwareCausal
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.091_timeawarecausal.zip
+  sha256: f135895ff69ff80efe2ba4f8752a8677949d0f078a2bff940425f43f6ab228d9
+  size_bytes: 187191114
+display:
+  palette_seed: navy-rose-causal
+  art_style: evolving-domain
+  accent_priority: high
+image_path: ../image/091_aisb.t3.091_timeawarecausal.jpg

package/AISB/catalog/aisb.t3.091_timeawarecausal.zh.yaml

@@ -0,0 +1,107 @@
+id: aisb.t3.091_timeawarecausal
+name: 学习时间感知因果表示以实现不断发展的领域中的模型泛化
+version: 0.1.0
+one_line: 使用顺序 VAE 框架通过时间感知因果表示学习来进行演化域泛化，从而减少虚假相关性。
+task_description: 该打包基准测试涵盖通过 SYNC（静态-动态因果表示学习）的演化域泛化 (EDG)，SYNC 是一种时间感知因果方法，可学习对时间分布变化具有鲁棒性的表示。该方法引入了时间感知的结构因果模型（SCM），该模型明确地模拟动态因果因素和因果机制漂移。顺序 VAE 框架捕获不断变化的模式，而信息论目标则保持跨时域和时域内因果因素的类内紧凑性。该方法理论上保证了每个时域的最佳因果预测器。
+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:
+- average_accuracy
+- causal_representation_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 在捆绑套件中的单个演进域数据集上运行一个打包的 SYNC 评估路线。
+- id: full_train_eval
+  label: 完整训练+评估
+  description: 端到端运行完整的时间感知因果表示学习工作流程，涵盖所有捆绑数据集。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://github.com/WonderSeven/LSSAE
+    note: Primary upstream source for bundled datasets
+  - url: https://deepscientist.cc/AISB/091_timeawarecausal
+    note: Local archive with prepackaged data splits
+  notes:
+  - 捆绑数据集：Circle、Sine（合成）、RMNIST、Portraits、Caltran、PowerSupply、ONP
+  - Caltran：车辆存在检测与时间照明混淆（白天/夜间杂散相关性）
+  - RMNIST：跨时域概念漂移的旋转 MNIST
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+  notes:
+  - 在 NVIDIA GeForce RTX 4090 上测试
+  - 单GPU就足够了；不需要多GPU训练
+environment:
+  python: 3.8.18
+  cuda: '11.6'
+  pytorch: 1.12.1
+  flash_attn: null
+  key_packages:
+  - torch==1.12.1
+  - torchvision==0.13.1
+  - torchaudio==0.12.1
+  - DomainBed
+  - LSSAE
+  notes:
+  - 捆绑的 requirements.txt 中设置了完整的依赖项
+  - 通过安装： conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
+risk_flags:
+- spurious_correlation_risk
+- temporal_drift_risk
+risk_notes:
+- Caltran 包含基于照明的杂散相关性，必须明确缓解
+- 模型依赖于学习时间感知的因果特征而不是快捷特征
+recommended_when: 当您需要用于不断发展的领域的因果表示方法、时间分布变化中的杂散相关缓解或具有因果保证的基准 EDG 方法时，请使用此基准。
+not_recommended_when: 如果您需要小型静态域泛化任务、没有时间结构的即用型图像分类基准或不需要因果解开的方法，请不要使用此基准。
+paper:
+  title: Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains
+  authors:
+  - Zhuo He
+  - Shuang Li
+  - Wenze Song
+  - Longhui Yuan
+  - Jian Liang
+  - Han Li
+  - Kun Gai
+  venue: ICML 2025
+  year: 2025
+  url: https://openreview.net/forum?id=KhCKypSaqx
+  note: 快手科技； PMLR 267
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.091_timeawarecausal.zip
+  archive_type: zip
+  local_dir_name: paper-91-TimeAwareCausal
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.091_timeawarecausal.zip
+  sha256: f135895ff69ff80efe2ba4f8752a8677949d0f078a2bff940425f43f6ab228d9
+  size_bytes: 187191114
+display:
+  palette_seed: navy-rose-causal
+  art_style: evolving-domain
+  accent_priority: high
+image_path: ../image/091_aisb.t3.091_timeawarecausal.jpg

package/AISB/catalog/aisb.t3.092_kmeanslocalopt.yaml

@@ -0,0 +1,138 @@
+id: aisb.t3.092_kmeanslocalopt
+name: Modified K-means Algorithm with Local Optimality Guarantees
+version: 0.1.0
+one_line: Optimize k-means with guaranteed local optimality while preserving classic
+  algorithmic simplicity and same computational complexity.
+task_description: 'This packaged benchmark covers modified K-means algorithms (LO-K-means)
+  that guarantee convergence to locally optimal solutions, addressing a fundamental
+  misconception in the field. The original K-means algorithm (Lloyd''s algorithm)
+  does not always converge to a locally optimal solution despite widespread belief
+  since the 1980s. The benchmark implements two variants: C-LO_K_means (continuous
+  local optimality) and D-LO_K_means (discrete local optimality), along with Min_D_LO_K_means
+  which balances improved accuracy with computational efficiency. All variants maintain
+  the same per-iteration time complexity as the original K-means algorithm. The implementation
+  supports weighted K-means clustering using general Bregman divergences, extending
+  beyond squared Euclidean distance.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- mean_clustering_loss
+- minimum_clustering_loss
+- average_computation_time_s
+- average_number_of_iterations
+launch_profiles:
+- id: python_eval
+  label: Python Eval
+  description: Run the packaged Python implementation comparing Standard K-means,
+    C-LO_K_means, D-LO_K_means, and Min_D_LO_K_means variants. Evaluates clustering
+    loss and iteration count using Iris dataset. Primary code anchor is py/run.py
+    with algorithm implementation in py/LO_K_means.py.
+- id: cpp_eval
+  label: C++ Eval
+  description: Use the primary C++ implementation for comprehensive local-optimality
+    experiment family. Includes test_multiple_runs.cpp, test_real_world.cpp, and test_news20.cpp
+    variants. Header implementation in cpp/LO_K_means.hpp with settings in cpp/Setting.hpp.
+dataset_download:
+  primary_method: bundled
+  sources:
+  - name: Iris dataset
+    url: embedded
+    path: data/Iris.txt
+    format: custom
+    description: Tab-separated text format with N, D header followed by data points
+  notes:
+  - Iris.txt bundled with N rows and D-dimensional points
+  - C++ implementation includes additional real-world and synthetic test scenarios
+  - Dataset loading handled by make_unique() function for weighted point processing
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+environment:
+  python: '3.9'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - numpy
+  notes:
+  - CPU-only benchmark; no GPU acceleration required or beneficial
+  - Pure numerical computation; torch/flash-attention not applicable
+  - Python route requires numpy for matrix operations and random generation
+  - C++ route compiles with standard g++/clang++ toolchain
+  - See bundled README/requirements.txt for full dependency set
+risk_flags:
+- no_runtime_trust
+- static_audit_only
+risk_notes:
+- Static code audit confirms executable anchors for all staged metrics
+- No benchmark execution performed during this packaging pass
+- Metric values marked as provisional until runtime verification completes
+- Primary metric mean_clustering_loss code-backed via py/LO_K_means.py
+recommended_when: 'Use this benchmark when you need a classic clustering task with
+  strong algorithmic clarity, modest hardware requirements, and CPU-only execution.
+  Appropriate for evaluating local optimality guarantees, comparing K-means variants,
+  or studying convergence behavior on low-dimensional tabular data. Excellent for
+  algorithm-focused research with minimal infrastructure overhead.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you require GPU acceleration,
+  deep learning workflows, high-dimensional embedding clustering, or large-scale distributed
+  training. Not suitable for benchmarks requiring neural network backbones or vision-language
+  components.
+
+  '
+paper:
+  title: Modified K-means Algorithm with Local Optimality Guarantees
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/li25bt.html
+  github_url: https://github.com/lmingyi/LO-K-means
+  authors:
+  - Li, Mingyi (University of Tokyo)
+  - Metel, Michael R. (Huawei Noah's Ark Lab)
+  - Takeda, Akiko (University of Tokyo / RIKEN AIP)
+display:
+  palette_seed: sand-slate-centroid
+  art_style: algorithmic-minimal
+  accent_priority: medium
+  tags:
+  - clustering
+  - optimization
+  - local-optimality
+  - bregman-divergence
+  - cpu-only
+  - icml2025
+image_path: ../image/092_aisb.t3.092_kmeanslocalopt.jpg
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.092_kmeanslocalopt.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.092_kmeanslocalopt.zip
+  archive_type: zip
+  sha256: 86ab62322776504678fb2616e825a32d89f9598e1f0a5900bfeb1231ba8a4ed6
+  size_bytes: 133834