Crystalline representations in learned models
We find that internal features organise into lattice-like structures, and show how to use them for control.
Large models learn internal representations that are often described as tangled. In new work, we find that a meaningful fraction of features organise into regular, lattice-like geometry — and that this regularity can be measured and used.
Summary of findings
- Features associated with compositional concepts arrange along a small number of shared axes.
- These axes are stable across training runs and model scales within a family.
- Interventions along the axes produce predictable, compositional changes in behaviour.
Regular structure is what makes a system legible. Where we find it, we can build on it.
We release our measurement tools alongside the paper so that others can reproduce and extend the results.