Different Language Models Learn Similar Number Representations
6 hours ago
- #number-representation
- #feature-learning
- #language-models
- Language models learn periodic features representing numbers with dominant periods at 2, 5, and 10.
- A two-tiered hierarchy exists: all models learn features with period-T spikes, but only some achieve geometric separability for mod-T classification.
- Fourier domain sparsity is necessary but insufficient for mod-T geometric separability.
- Data, architecture, optimizer, and tokenizer influence whether models develop geometrically separable features.
- Two main pathways to acquiring these features: complementary co-occurrence signals in general language data and multi-token addition problems.
- The study demonstrates convergent evolution in feature learning across diverse models and training signals.