Machine Psychology: Code, Genomic, and Scale Extensions

This paper extends Machine Psychology beyond general language models into code, genomics, and scale. Across DeepSeek-Coder, Nucleotide Transformer, cross-domain transfer, and within-family scaling, the results show a consistent pattern: family structure remains probe-readable, native probe calibration often improves alignment, but cross-domain transport still requires richer full-space geometry than small invariant subspaces can provide本研究将机器心理学（Machine Psychology）的研究范畴从通用大语言模型拓展至代码、基因组学与规模化建模领域。针对DeepSeek-Coder、核苷酸Transformer（Nucleotide Transformer）、跨域迁移以及同家族缩放四类实验情境，研究结果呈现出一致的规律：模型家族的内部结构仍可被探测探针解析，原生探针校准通常可提升模型对齐效果，但跨域传递仍需要相较于小型不变子空间更为丰富的全空间几何特性。