Language, Not Code Abstraction: Natural-language Training Shapes Brain-aligned Geometry

A study on arXiv (2611.09027v1) trained 25 GPT-2-small models on TinyStories, Python, C functions, x86-64 assembly and machine code, and found natural-language training produced closer alignment with brain data than code training. On the Pereira-243 and Tuckute2024 benchmarks, TinyStories exceeded the mean of the four code models by Δr = 0.053 and 0.059 (Holm-corrected p<0.001), while code training left alignment near random-weight levels; the advantage was benchmark-specific, with no corrected language-code difference on Blank2014 and no language benefit on Fedorenko2016. TinyStories seeds converged on the most similar representations, with fit-free correspondence to neural dissimilarity peaking at ρ = 0.14, twice the code and random-weight levels, and updating only the embedding matrix on English improved mean-over-depth alignment for every code model more than an equal update on its own corpus.

arXiv:2611.09027v1 Announce Type: new Abstract: Code-trained language models produce positive brain scores, which raises the question of whether abstract structure learned from programming languages can substitute for natural-language training. We trained 25 GPT-2-small models on TinyStories, Python, C functions, x86-64 assembly, and the corresponding machine code, and scored them using a shared tokenizer and encoding procedure against matched random-weight models. On Pereira-243 and Tuckute2024, TinyStories exceeds the mean of the four code models by $\Delta r = 0.053$ and $0.059$ Holm-corrected $p<0.001$ , while code training leaves alignment near the random-weight level. The language advantage is benchmark-specific: Blank2014 shows no corrected language-code difference, and Fedorenko2016 shows no language benefit. Geometry tracks the encoding result on Pereira-243: TinyStories seeds converge on the most similar representations, and its fit-free correspondence with neural dissimilarity peaks at $\rho = 0.14$, twice the code and random-weight levels. Binary-trained models have about three effective dimensions, so seed agreement alone does not imply neural correspondence. Updating only the embedding matrix on English improves mean-over-depth alignment for every code model more than an equal update on its own corpus. Within this architecture and training setup, natural-language training produces a distinct representational geometry with closer neural correspondence on Pereira-243