@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.102_stochasticff.zh.yaml

@@ -0,0 +1,130 @@
+id: aisb.t3.102_stochasticff
+name: 通过表征维度压缩进行随机前向学习
+version: 0.1.0
+one_line: 通过有效的维度优度函数进行非 BP 训练，该函数压缩类内表示并利用随机性作为计算资源。
+task_description: 该基准测试通过从神经响应的有效维度 (ED) 派生的新颖的“维度压缩”优度函数来实现随机前向学习 (FF)。与需要手工制作负样本并使用平方和激活作为优点的标准 FF 不同，该方法通过非中心协方差矩阵的特征值谱利用二阶统计结构。训练目标最小化单个输入的噪声同系物的 ED，同时最大化整个样本分布的 ED，从而消除显式负样本生成的需要。该方法在 CIFAR-10、CIFAR-100 和 MNIST 上实现了与其他非反向传播方法相比的竞争精度，并且固有地支持均方输出的基于能量的推理。包括用于估计单个神经元和协同神经元组合中的互信息的信息分解分析工具。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- cifar10_val_accuracy
+- cifar100_val_accuracy
+- mnist_val_accuracy
+- best_model_checkpoint
+- training_log
+- representation_analysis
+launch_profiles:
+- id: quick_cifar10
+  label: 快速 CIFAR-10
+  description: 使用默认超参数执行 CIFAR-10 训练路线。通过torchvision下载数据集，将ZCA统计数据缓存在./data/下，将检查点保存到./checkpoint/，并报告最终验证准确性。单 GPU 上的典型运行时间为 2-4 小时。
+  hardware_notes: 默认配置使用 3 个 epoch，批量大小 128，学习率 0.001，并进行余弦退火。
+- id: full_suite_analysis
+  label: 全套+分析
+  description: 跨 CIFAR-10、CIFAR-100 和 MNIST 运行多数据集训练工作流程，然后调用捆绑分析路径来生成信息分解图（图 4c/5c 等效项）。需要根据 training.py 注释手动调用特定的实验函数。
+  hardware_notes: 全套分析功能可按比例延长运行时间。第 2 阶段训练使用 60 个 epoch。
+dataset_download:
+  primary_method: torchvision
+  sources:
+  - name: CIFAR-10
+    path: torchvision.datasets.CIFAR10
+  - name: CIFAR-100
+    path: torchvision.datasets.CIFAR100
+  - name: MNIST
+    path: torchvision.datasets.MNIST
+  notes:
+  - 数据集在首次运行时自动下载并缓存在本地 ./data/ 下
+  - ZCA 白化统计数据是预先计算的，并缓存在 ./data/ 下的每个数据集
+  - 引用的 fast_mnist 片段通知帮助程序的实现；这不是数据集获取途径
+  - torchvision 提供的数据集不需要外部 API 凭据
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 在捆绑源中未检测到显式数据集或 API 凭据
+  - GitHub 存储库位于 https://github.com/ZhichaoZhu/StochasticForwardForward
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 30
+    gpu_count: 1
+    gpu_vram_gb: 8
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torch>=2.0
+  - torchvision
+  - numpy
+  - scipy
+  - scikit-learn
+  - pandas
+  - tqdm
+  notes:
+  - 请参阅捆绑的 requirements.txt 了解固定版本
+  - 请参阅捆绑的自述文件/要求以获取完整的依赖项集
+risk_flags:
+- experimental_algorithm
+- nonstandard_training_paradigm
+risk_notes:
+- FF学习不如反向传播成熟；结果可能因随机种子而异
+- 该算法系列不存在标准的微调方法
+- Hyperparameter sensitivity: ED损耗平衡系数、采样长度、投影尺寸需要调整
+recommended_when: '- 探索反向传播的替代训练范式 - 研究生物学上合理的学习算法 - 研究以随机性为资源的神经形态计算应用 - 研究二阶统计结构的表示学习 - 神经网络表示的信息论分析'
+not_recommended_when: '- 需要传统的大型模型微调工作流程 - 需要具有广泛工具的生产级训练管道 - 需要快速迭代而无需算法特定专业知识的任务'
+paper:
+  title: Stochastic Forward-Forward Learning through Representational Dimensionality Compression
+  authors:
+  - Zhu, Zhichao
+  - Qi, Yang
+  - Ma, Hengyuan
+  - Lu, Wenlian
+  - Feng, Jianfeng
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2505.16649
+  github: https://github.com/ZhichaoZhu/StochasticForwardForward
+  citation: "@article{zhu2025stochastic,\n  title={Stochastic Forward-Forward Learning through Representational Dimensionality Compression},\n  author={Zhu, Zhichao and Qi, Yang and Ma, Hengyuan and Lu, Wenlian and Feng, Jianfeng},\n  journal={arXiv preprint arXiv:2505.16649},\n  year={2025}\n}\n"
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.102_stochasticff.zip
+  archive_type: zip
+  local_dir_name: paper-102-StochasticFF
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.102_stochasticff.zip
+  sha256: 4eab3f39d4cc098ee47e9d4a31c8485f3b09752123b134a05292f5bfababe41e
+  size_bytes: 65412
+display:
+  palette_seed: amber-indigo-forward
+  art_style: learning-dynamics
+  accent_priority: medium
+image_path: ../image/102_aisb.t3.102_stochasticff.jpg
+capability_tags:
+- forward_forward_learning
+- representation_learning
+- non_backpropagation
+- effective_dimensionality
+- image_classification
+- efficient_training
+- biologically_plausible_learning
+- information_breakdown_analysis
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3

package/AISB/catalog/aisb.t3.103_qdcp.yaml

@@ -0,0 +1,150 @@
+id: aisb.t3.103_qdcp
+name: Distributed Conformal Prediction via Message Passing
+version: 0.1.0
+one_line: Distributed conformal prediction with message passing on arbitrary graph
+  topologies, providing communication-aware coverage guarantees for decentralized
+  uncertainty quantification.
+task_description: 'This packaged benchmark covers quantile-based distributed conformal
+  prediction (Q-DCP) and histogram-based distributed conformal prediction (H-DCP)
+  from ICML 2025. Each device in a decentralized network holds limited local calibration
+  data and can only communicate with neighboring nodes via message passing. Q-DCP
+  employs distributed quantile regression enhanced with ADMM-based smoothing and regularization
+  to accelerate convergence, while H-DCP uses consensus-based histogram estimation
+  to achieve hyperparameter-free coverage. The benchmark evaluates trade-offs between
+  coverage guarantees, prediction set sizes, communication overhead, and hyperparameter
+  sensitivity across different network topologies including star, ring, and path graphs.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: partial
+support_level: advanced
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- mean_set_size
+- mean_coverage
+- topology_comparison_plots
+- experiment_score_tensors
+launch_profiles:
+- id: build_scores
+  label: Build Scores
+  description: 'Run central or federated training to generate experiment score folders
+    under `experiments/`. Uses `src/run_TCT.py` or bundled shell scripts (e.g., `scripts/cifar100.sh`).
+    Score folders must match naming conventions like `{dataset}_{topology}_{model}`
+    (e.g., `cifar100_central_small_resnet14`).
+
+    '
+- id: qdcp_eval
+  label: Q-DCP Eval
+  description: 'Run the staged Q-DCP or H-DCP analysis on prepared score folders using
+    `run_qdcp_mod.py` or the notebooks. Reports coverage and set-size curves across
+    alpha levels and topology comparisons. Execute from `/repo` with pre-staged `experiments/`
+    tree.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - torchvision (CIFAR-10, CIFAR-100, FashionMNIST, MNIST, SVHN)
+  - MedMNIST (PathMNIST, BloodMNIST, TissueMNIST, DermMNIST)
+  - Fitzpatrick-SM (requires separate staging)
+  notes:
+  - Public datasets are fetched automatically by `src/run_TCT.py` or notebook loaders.
+  - Fitzpatrick and similar experiments expect separately staged local files or CSV
+    paths.
+  - The Q-DCP staged route consumes pre-computed score folders rather than raw data
+    directly.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit dataset or API credential variables detected in the snapshot.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - numpy
+  - torch
+  - matplotlib
+  - pandas
+  - tdigest
+  - ddsketch
+  - scipy
+  - notebook
+  notes:
+  - CPU-only execution is sufficient for the full benchmark route.
+  - Bundled requirements file is a conda export targeting a Windows py311 environment.
+  - Core calibration logic is in `src/DCP_pinball_smooth.py` and `src/class_CP_QQ.py`.
+  - See bundled README/requirements for the full dependency set.
+risk_flags:
+- route_caveat
+- hardcoded_path_dependency
+risk_notes:
+- The snapshot contains hard-coded `/repo` paths; execution requires mounting to this
+  path.
+- The staged Q-DCP runner expects a prepared `experiments/` tree with score folders.
+- Score folders must follow naming conventions or helpers.load_scores will fail.
+- Some routes depend on separately staged local files (e.g., Fitzpatrick experiments).
+- No benchmark execution was performed in this packaging pass; static audit only.
+recommended_when: 'Use this benchmark for communication-aware uncertainty quantification
+  over decentralized devices on arbitrary graph topologies, or when evaluating trade-offs
+  between coverage guarantees, set sizes, and communication overhead in distributed
+  settings.
+
+  '
+not_recommended_when: 'Do not use if you need a centralized prediction benchmark with
+  no network topology, or if you require execution without pre-staging an experiments
+  directory.
+
+  '
+paper:
+  title: Distributed Conformal Prediction via Message Passing
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2501.14544
+  github: https://github.com/HaifengWen/Distributed-Conformal-Prediction
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.103_qdcp.zip
+  archive_type: zip
+  local_dir_name: paper-103-QDCP
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.103_qdcp.zip
+  sha256: f4c79a3d5e5a28a6b2d53f67023f6a15efd49b4ab2204e0f90812cc9f9f43e1c
+  size_bytes: 881452
+display:
+  palette_seed: teal-ivory-message
+  art_style: network-calibration
+  accent_priority: medium
+image_path: ../image/103_aisb.t3.103_qdcp.jpg
+capability_tags:
+- conformal_prediction
+- distributed_learning
+- uncertainty_quantification
+- message_passing
+- research_code_optimization
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3

package/AISB/catalog/aisb.t3.103_qdcp.zh.yaml

@@ -0,0 +1,116 @@
+id: aisb.t3.103_qdcp
+name: 通过消息传递的分布式共形预测
+version: 0.1.0
+one_line: 分布式共形预测，消息在任意图拓扑上传递，为分散的不确定性量化提供通信感知的覆盖保证。
+task_description: 该打包基准涵盖 ICML 2025 中基于分位数的分布式共形预测 (Q-DCP) 和基于直方图的分布式共形预测 (H-DCP)。去中心化网络中的每个设备都保存有限的本地校准数据，并且只能通过消息传递与相邻节点进行通信。 Q-DCP 采用基于 ADMM 的平滑和正则化增强的分布式分位数回归来加速收敛，而 H-DCP 使用基于共识的直方图估计来实现无超参数覆盖。该基准评估不同网络拓扑（包括星形图、环形图和路径图）之间的覆盖保证、预测集大小、通信开销和超参数敏感性之间的权衡。
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: partial
+support_level: advanced
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- mean_set_size
+- mean_coverage
+- topology_comparison_plots
+- experiment_score_tensors
+launch_profiles:
+- id: build_scores
+  label: 建立分数
+  description: 运行中央或联合训练以在 `experiments/` 下生成实验分数文件夹。使用 `src/run_TCT.py` 或捆绑的 shell 脚本（e.g.、`scripts/cifar100.sh`）。乐谱文件夹必须符合命名约定，例如 `{dataset}_{topology}_{model}` (e.g., `cifar100_central_small_resnet14`)。
+- id: qdcp_eval
+  label: Q-DCP评估
+  description: 使用 `run_qdcp_mod.py` 或笔记本在准备好的乐谱文件夹上运行分阶段 Q-DCP 或 H-DCP 分析。报告 alpha 级别的覆盖范围和设置大小曲线以及拓扑比较。使用预先准备的 `experiments/` 树从 `/repo` 执行。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - torchvision (CIFAR-10, CIFAR-100, FashionMNIST, MNIST, SVHN)
+  - MedMNIST (PathMNIST, BloodMNIST, TissueMNIST, DermMNIST)
+  - Fitzpatrick-SM (requires separate staging)
+  notes:
+  - 公共数据集由 `src/run_TCT.py` 或笔记本加载程序自动获取。
+  - Fitzpatrick 和类似的实验期望单独暂存的本地文件或 CSV 路径。
+  - Q-DCP 分阶段路线使用预先计算的分数文件夹，而不是直接使用原始数据。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 快照中未检测到显式数据集或 API 凭证变量。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - numpy
+  - torch
+  - matplotlib
+  - pandas
+  - tdigest
+  - ddsketch
+  - scipy
+  - notebook
+  notes:
+  - 仅 CPU 执行足以满足完整的基准测试路线。
+  - 捆绑的需求文件是针对 Windows py311 环境的 conda 导出。
+  - 核心校准逻辑位于`src/DCP_pinball_smooth.py`和`src/class_CP_QQ.py`中。
+  - 请参阅捆绑的自述文件/要求以获取完整的依赖项集。
+risk_flags:
+- route_caveat
+- hardcoded_path_dependency
+risk_notes:
+- 快照包含硬编码的 `/repo` 路径；执行需要挂载到该路径。
+- 分阶段的 Q-DCP 运行程序需要一个准备好的 `experiments/` 树，其中包含分数文件夹。
+- 乐谱文件夹必须遵循命名约定，否则 helpers.__MASK0__ 将失败。
+- 一些路由依赖于单独暂存的本地文件（e.g.，Fitzpatrick 实验）。
+- 在此打包过程中未执行任何基准测试；仅静态审核。
+recommended_when: 使用此基准对任意图拓扑上的分散设备进行通信感知的不确定性量化，或者在评估分布式设置中的覆盖保证、集大小和通信开销之间的权衡时。
+not_recommended_when: 如果您需要没有网络拓扑的集中式预测基准，或者需要在不预先暂存实验目录的情况下执行，请勿使用。
+paper:
+  title: Distributed Conformal Prediction via Message Passing
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2501.14544
+  github: https://github.com/HaifengWen/Distributed-Conformal-Prediction
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.103_qdcp.zip
+  archive_type: zip
+  local_dir_name: paper-103-QDCP
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.103_qdcp.zip
+  sha256: f4c79a3d5e5a28a6b2d53f67023f6a15efd49b4ab2204e0f90812cc9f9f43e1c
+  size_bytes: 881452
+display:
+  palette_seed: teal-ivory-message
+  art_style: network-calibration
+  accent_priority: medium
+image_path: ../image/103_aisb.t3.103_qdcp.jpg
+capability_tags:
+- conformal_prediction
+- distributed_learning
+- uncertainty_quantification
+- message_passing
+- research_code_optimization
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3

package/AISB/catalog/aisb.t3.104_balancedactiveinf.yaml

@@ -0,0 +1,137 @@
+id: aisb.t3.104_balancedactiveinf
+name: Balanced Active Inference
+version: 0.1.0
+one_line: Active inference for regression with cube-method balanced sampling, calibrated
+  uncertainty, and asymptotically efficient population mean estimation.
+task_description: 'This packaged benchmark implements balanced active inference combining
+  the cube method for balanced sampling with active inference to address variance
+  inflation in label-scarce regression settings. The framework trains two XGBoost
+  models: a label predictor and an uncertainty estimator. Sampling probabilities use
+  a τ-mixed rule (τ=0.5 default), followed by cube-method balancing via constrained
+  optimization. The population mean is estimated using a Generalized Difference estimator.
+  Evaluated against classical simple random sampling, uniform active inference, and
+  traditional Poisson-based active inference baselines. Supports synthetic (Linear,
+  Nonlinear, Friedman) and real-world regression datasets (Bike, Communities, Concrete,
+  Energy, Life, Superconductor).
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- ci_width_cube_active
+- coverage_cube_active
+- prediction_rmse
+- comparison_figure
+launch_profiles:
+- id: demo_notebook
+  label: Demo Notebook
+  description: 'Run the bundled demo.ipynb to generate Friedman synthetic data, train
+    dual XGBoost models, perform sampling budget sweeps, and produce publication-quality
+    comparison figures.
+
+    '
+- id: full_budget_sweep
+  label: Full Budget Sweep
+  description: 'Execute the complete simulation experiment with T=10,000 Monte Carlo
+    repetitions across budgets 0.03–0.45, exporting CSV metrics and PDF comparison
+    plots.
+
+    '
+dataset_download:
+  primary_method: procedural
+  sources: []
+  notes:
+  - Benchmark generates Friedman-style synthetic regression data procedurally inside
+    src/data_generation.py.
+  - Formula: Y = 10*sin(π*X₁*X₂) + 20*(X₃-0.5)² + 10*X₄ + 5*X₅ + ε, with ε~N(0, 0.09).
+  - Features X₁–X₅ are informative; X₆–X₁₀ are noise variables.
+  - No external dataset download required; real-world datasets are optional extensions.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No API credentials required; purely CPU-bound procedural computation.
+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: '3.8'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - xgboost==1.5.0
+  - rpy2==3.4.0
+  - numpy>=1.21.0
+  - pandas>=1.3.0
+  - scikit-learn>=1.0.0
+  - scipy>=1.7.0
+  - matplotlib>=3.4.0
+  - seaborn>=0.11.0
+  - joblib>=1.1.0
+  notes:
+  - CPU-only execution; no GPU required.
+  - R runtime required for BalancedSampling package: install via R console with install.packages("BalancedSampling").
+  - Python+R interoperability via rpy2; verify R installation before executing.
+risk_flags:
+- route_caveat
+risk_notes:
+- Execution depends on Python + R interoperability through rpy2 and the BalancedSampling
+  R package.
+- Ensure R is installed and rpy2 can load the BalancedSampling library before running.
+- Synthetic-data-driven route; no external dataset download or network access needed.
+- Parallel execution via joblib (n_jobs parameter) requires sufficient RAM.
+- Full budget sweep (T=10,000 trials) is computationally intensive; consider reducing
+  n_trials for quick validation.
+recommended_when: 'Use this benchmark for CPU-friendly active learning research with
+  variance estimation, uncertainty calibration, or statistical efficiency comparisons.
+  Ideal for regression tasks with limited labeling budgets where reliable confidence
+  intervals and RMSE reduction are primary metrics.
+
+  '
+not_recommended_when: 'Do not use this if you require large-scale multimodal benchmarks,
+  deep learning pipelines, or GPU-accelerated training. Not suited for tasks needing
+  external dataset downloads or real-time API access.
+
+  '
+paper:
+  title: Balanced Active Inference
+  authors: Boyu Chen, Zhixiang Zhou, Liuhua Peng, Zhonglei Wang
+  institution: Xiamen University, Shanghai Innovation Institute, The University of
+    Melbourne
+  venue: NeurIPS
+  year: 2025
+  url: https://neurips.cc/virtual/2025/poster/117263
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.104_balancedactiveinf.zip
+  archive_type: zip
+  local_dir_name: paper-104-BalancedActiveInf
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.104_balancedactiveinf.zip
+  sha256: d1c5e0c466d3f1bdd4fdfe96fa9161ea425e88dbf0d0b2ab9f6c2a3992d8c338
+  size_bytes: 35848
+display:
+  palette_seed: lemon-slate-balance
+  art_style: active-estimation
+  accent_priority: medium
+image_path: ../image/104_aisb.t3.104_balancedactiveinf.jpg

package/AISB/catalog/aisb.t3.104_balancedactiveinf.zh.yaml

@@ -0,0 +1,104 @@
+id: aisb.t3.104_balancedactiveinf
+name: 平衡主动推理
+version: 0.1.0
+one_line: 使用立方体方法平衡抽样、校准不确定性和渐近有效总体均值估计进行回归的主动推理。
+task_description: '该打包基准测试实现了平衡主动推理，将平衡采样的立方体方法与主动推理相结合，以解决标签稀缺回归设置中的方差膨胀问题。 The framework trains two XGBoost models: a label predictor and an uncertainty estimator.采样概率使用 τ 混合规则（默认值 τ=0.5），然后通过约束优化进行立方体方法平衡。总体平均值是使用广义差值估计器估计的。根据经典的简单随机采样、均匀主动推理和传统的基于泊松的主动推理基线进行评估。支持合成（线性、非线性、弗里德曼）和现实世界回归数据集（自行车、社区、混凝土、能源、生命、超导）。'
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- ci_width_cube_active
+- coverage_cube_active
+- prediction_rmse
+- comparison_figure
+launch_profiles:
+- id: demo_notebook
+  label: 演示笔记本
+  description: 运行捆绑的 demo.ipynb 以生成 Friedman 合成数据、训练双 XGBoost 模型、执行抽样预算扫描并生成出版质量的比较数据。
+- id: full_budget_sweep
+  label: 全面预算扫荡
+  description: 在预算 0.03–0.45 范围内执行 T=10,000 次蒙特卡罗重复的完整模拟实验，导出 CSV 指标和 PDF 比较图。
+dataset_download:
+  primary_method: procedural
+  sources: []
+  notes:
+  - Benchmark 在 src/data_generation.py 内按程序生成弗里德曼式综合回归数据。
+  - Formula: Y = 10*sin(π*X₁*X2) + 20*(X₃-0.5)² + 10*X₄ + 5*X₅ + ε，其中 ε~N(0, 0.09)。
+  - 特征 X₁–X₅ 具有丰富的信息； X₆–X₁₀ 是噪声变量。
+  - 无需下载外部数据集；现实世界的数据集是可选的扩展。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 无需 API 凭证；纯粹受 CPU 限制的过程计算。
+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: '3.8'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - xgboost==1.5.0
+  - rpy2==3.4.0
+  - numpy>=1.21.0
+  - pandas>=1.3.0
+  - scikit-learn>=1.0.0
+  - scipy>=1.7.0
+  - matplotlib>=3.4.0
+  - seaborn>=0.11.0
+  - joblib>=1.1.0
+  notes:
+  - 仅 CPU 执行；无需 GPU。
+  - R runtime required for BalancedSampling package: 通过 R 控制台使用 install.packages("BalancedSampling") 安装。
+  - 通过 rpy2 实现 Python+R 互操作；执行前验证 R 安装。
+risk_flags:
+- route_caveat
+risk_notes:
+- 执行取决于通过 rpy2 和 BalancedSampling R 包实现的 Python + R 互操作性。
+- 确保 R 已安装并且 rpy2 在运行前可以加载 BalancedSampling 库。
+- 综合数据驱动路线；无需外部数据集下载或网络访问。
+- 通过 joblib（n_jobs 参数）并行执行需要足够的 RAM。
+- 全面预算扫描（T=10,000 次试验）需要大量计算；考虑减少 n_trials 以进行快速验证。
+recommended_when: 使用此基准进行 CPU 友好的主动学习研究，包括方差估计、不确定性校准或统计效率比较。非常适合标签预算有限的回归任务，其中可靠的置信区间和 RMSE 降低是主要指标。
+not_recommended_when: 如果您需要大规模多模式基准测试、深度学习管道或 GPU 加速训练，请勿使用此选项。不适合需要外部数据集下载或实时 API 访问的任务。
+paper:
+  title: Balanced Active Inference
+  authors: Boyu Chen, Zhixiang Zhou, Liuhua Peng, Zhonglei Wang
+  institution: Xiamen University, Shanghai Innovation Institute, The University of Melbourne
+  venue: NeurIPS
+  year: 2025
+  url: https://neurips.cc/virtual/2025/poster/117263
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.104_balancedactiveinf.zip
+  archive_type: zip
+  local_dir_name: paper-104-BalancedActiveInf
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.104_balancedactiveinf.zip
+  sha256: d1c5e0c466d3f1bdd4fdfe96fa9161ea425e88dbf0d0b2ab9f6c2a3992d8c338
+  size_bytes: 35848
+display:
+  palette_seed: lemon-slate-balance
+  art_style: active-estimation
+  accent_priority: medium
+image_path: ../image/104_aisb.t3.104_balancedactiveinf.jpg