Language → Action

A guided tour through the internals of a tiny transformer learning to map synonymous natural-language commands onto a small action vocabulary.

1. The task
2. Baseline: a 1-layer model, no synonyms
3. Attention patterns
4. Direct logit attribution
5. Turning the data dial: synonyms
6. Turning the model dial: depth
7. Activation patching: the causal story
8. Synthesis: across both axes