Results; Extended Data 2a and 9; SI: meaning
Semantic alignment and transfer
Consensus geometry aligns with independent semantic representations and supports category transfer. Residual eigenvector sharing beyond a displacement-magnitude control is smaller and concentrated in leading modes.
Loading recorded comparison…
What the shared geometry represents
Loading recorded comparison…
Choosing a meaningful comparison for shared directions
Loading recorded comparison…
Cost, sensitivity and sharing
Loading recorded comparison…
Methods and interpretation
Semantic batteries, transfer probes and subspace-overlap controls use distinct endpoints. Most raw leading overlap is explained by shared displacement magnitudes. The residual measures alignment remaining after this control.
Procedure in the paper: Prepare the natural, templated and semantic batteries, acquire the fixed independent semantic encoder, compute consensus and residual geometries, and evaluate the frozen probes and null constructions.
Code and data
v0.1.0 · 74 KB · View source on GitHub ↗
Fresh six-model measurements on the fixed natural, templated and factual batteries; semantic/control comparisons, complete cross-model logistic probes and native spectral/null measurements with current context-bootstrap analyses. Historical model snapshot equivalence is assumed from unchanged upstream identities.
v0.1.0 · 69 KB · View source on GitHub ↗
Fixed-prompt and natural-text axis sharing, salience-matched nulls, the corrected natural-context bootstrap and all six rank bands.
v0.1.0 · 91 KB · View source on GitHub ↗
Representation metrics, coordinate changes, natural-text and n-gram baselines, corpus mediation, reference scale, output distances and permutation calibration.
v0.1.0 · 66 KB · View source on GitHub ↗
Four-model, six-objective matched-effect steering and its model/domain robustness analysis.
v0.1.0 · 66 KB · View source on GitHub ↗
Vocabulary centering, word-frequency controls and the decision-axis anisotropy comparison.
Each standalone package includes code, shared helpers, required small inputs and reference results, with setup and commands in its README. Model weights and public datasets are obtained separately where needed.
Download complete source (v0.1.0) for all experiments and the companion website.