causal-internalization-physical-law
收藏资源简介:
该数据集是论文《Causal internalization of physical law in neural networks: a formal measurement framework for representation, intervention and extrapolation across scale》的可重复性包,包含代码、实验数据和结果。数据集的目的是评估神经网络是否通过数据缩放内化隐藏物理参数(如重力、恢复系数)的负载支撑编码,还是仅学习流畅行为。测量框架包括四个指标:探针保真度(ρ)、互换/负载支撑性(ι)、外推(ξ)和变形(σ),并配有校准的正负控制。主要实验包括:物理网格(36个单元,涉及球和摆,2种尺寸×3种预算×3种种子,复制GPU/CPU),算术套件(加、减、乘、除,在模数31下内化),以及四单元域内设计空间(plain→aux-g→decode-g→appended-g)。数据形式为JSON结果文件,存储于results/目录,包含每个单元/操作/控制的测量结果。论文文件(main_filled.pdf/tex)和脚本(planted_control.py, arith_control.py等)也包含在内。该数据集适用于研究神经网络表示学习、因果结构、物理定律内化、可解释性以及缩放定律等领域。
This dataset is a reproducibility package for the paper Causal internalization of physical law in neural networks: a formal measurement framework for representation, intervention and extrapolation across scale, containing code, experimental data, and results. The dataset aims to evaluate whether neural networks internalize load-bearing encodings of hidden physical parameters (such as gravity, coefficient of restitution) through data scaling, or merely learn fluent behaviors. The measurement framework includes four metrics: probe fidelity (ρ), interchange/load-bearing (ι), extrapolation (ξ), and deformation (σ), with calibrated positive and negative controls. Main experiments include: Physical Grid (36 units, involving balls and pendulums, 2 sizes × 3 budgets × 3 seeds, replicated on GPU/CPU), Arithmetic Suite (addition, subtraction, multiplication, division, internalized under modulus 31), and a four-unit in-domain design space (plain→aux-g→decode-g→appended-g). Data is in JSON result files stored in the results/ directory, containing measurements for each unit/operation/control. Paper files (main_filled.pdf/tex) and scripts (planted_control.py, arith_control.py, etc.) are also included. This dataset is suitable for research in neural network representation learning, causal structure, internalization of physical laws, interpretability, and scaling laws.
数据集概述:因果内化物理定律的可复现性包
数据集描述
这是一个用于复现学术论文《Causal internalization of physical law in neural networks: a formal measurement framework for representation, intervention and extrapolation across scale》的完整捆绑数据集。该数据集的核心目标是衡量神经网络是否通过数据缩放构建了针对隐藏物理参数(如重力、恢复系数)的承重内部编码,还是仅表现出流畅行为。
主要衡量指标
该数据集定义了四项正式度量标准,合称“证书”:
- ρ(探针保真度)
- ι(互换性/承重性)
- ξ(外推能力)
- σ(变形能力)
所有度量均配有校准的正负对照组。
主要实验结果
1. 物理网格实验
- 实验规模:36个单元(摆球+钟摆,2种尺寸 × 3个预算 × 3个随机种子,GPU/CPU双平台复现)
- 发现:准确率随算力提升而提高,但重力探针读数处于底层水平(0.060 / 0.001),ι = 0.000,36/36个单元均无涌现现象——呈现出“随机鹦鹉”特征
2. 算术套件实验
- 任务:+、−、×、÷运算,全部在模数p=31下达到grokking状态
- 发现:grokked模型携带结果编码(ρ_res 0.67–1.00),在分布外输入上计算准确率达1.000,但在第二个模数上仅得0.517 = 435/841分——说明规律是真实存在且承重的(激活补丁翻转率达1.000),但被限制在ℤ₃₁域内
3. 四单元域内设计空间
- 结果轨迹:ι值在普通→辅助梯度→解码梯度→追加梯度间呈现 0.000 → 0.000 → 0.298 → 1.000 的变化
- 关键洞见:当目标函数选择特定代码时证书才偏离底线(潜在监督提升ρ,路由监督提升ι)——内化分解为编码准确率×控制器路由,两者皆为目标属性而非规模属性
文件构成
- 论文文件:
main_filled.pdf与.tex - 控制实验代码:
planted_control.py(可运行) - 算术套件:
arith_control.py、arith_patch_sweep.py、analyze_arith.py - 论文生成工具:
fill_paper_internalization.py - 结果与图表:
results/目录存放所有单元级JSON结果,figures/存放论文图表
Kaggle复现说明
使用数据集 sehajrsingh/causal-internalization-code(v3版本)搭配内核 sehajrsingh/causal-internalization-planted-controls 可端到端运行四单元控制套件。





