The geometry is domain-specific. Transfer across domains destroys it.
When a linear projection matrix maps one language model's internal representations into another's coordinate system, the projection works within a reasoning domain: R² = 0.50 for verbal reasoning, R² = 0.40 for mathematical reasoning (arXiv:2603.20406). Behavioral corrections reach 14-50% on TruthfulQA and 8.5-43.3% on arithmetic. The independently trained models have developed aligned geometric structures — similar enough that one model's “thinking” can be translated into another's and produce meaningful behavioral change.
Note (2026-11-12; read against the paper's full text): the cross-domain collapse is not the same in both directions: the paper reports a mean R-squared of minus 3.83 for the projection trained on TruthfulQA and tested on arithmetic, and minus 1.73 for the reverse. These are R-squared values for predicting the target model's activations (worse than predicting the mean), which the paper calls active geometric interference; the essay's statement that the projection destroys mathematical reasoning when applied is not a behavioural result reported there. In its double-dissociation tables the mean in-domain held-out R-squared across the 20 pairings is 0.343 for TruthfulQA and 0.256 for GSM8K, whereas the 0.50 and 0.40 in the essay are the abstract's headline values. The behavioral correction rates the essay quotes are ranges across the pairings (means 25.2 and 25.5 percent).
But the alignment is domain-bound. Transfer a verbal-reasoning projection matrix to mathematical reasoning, or vice versa, and the R² collapses to -3.83. Not zero — negative. Worse than random. The projection that meaningfully translates one model's verbal reasoning into another's actively destroys mathematical reasoning when applied to it.
This means language models don't develop a single unified representational geometry. They develop multiple, domain-specific geometries that share a feature space but organize it differently depending on what kind of reasoning they're doing. The same neurons participate in both verbal and mathematical reasoning, but the relationships between neurons — the directions that matter, the subspaces that encode meaning — are different for each domain.
A second counterintuitive finding: geometric alignment quality has near-zero correlation with behavioral correction rate (r = -0.07). A projection that faithfully reproduces the geometry doesn't necessarily fix the behavior. A projection that distorts the geometry sometimes does. The geometry and the behavior are connected but not by a simple mapping. Correcting the representation isn't the same as correcting the output.
Two independently trained models, converging on similar within-domain geometry, diverging across domains, and decoupled from behavior at the projection boundary. The representations are more structured and less unified than either the universality thesis or the domain-independence assumption would predict.