ADR-036: learning-theory widens to cover training dynamics

Context

The topics vocabulary in luria.yaml is closed, and a word it lacks is added by decision, with a label and a blurb. learning-theory’s blurb read: “why learning generalizes — compression and description length, Kolmogorov complexity and sufficient statistics, rate–distortion and PAC-style bounds, model selection”.

On 2026-10-09 the record took in a run of papers whose subject is how networks learn rather than why they generalize: exact and closed-form training dynamics (LIT-855, LIT-859, LIT-862), grokking (LIT-858), the statistical physics of attention (LIT-857) and topological change during training (LIT-861). Five readers reported the same gap. Each tagged the work learning-theory as the nearest fit and said the blurb did not cover it. Older entries had done the same (LIT-370, LIT-242).

Decision

learning-theory keeps its label and widens its blurb to:

why learning generalizes and how it proceeds — compression and description length, Kolmogorov complexity and sufficient statistics, rate–distortion and PAC-style bounds, model selection; and training dynamics: solvable models of learning trajectories, stages, plateaus and phase transitions in training, and the implicit bias of optimisation. A work whose subject is how to train a model better is ML practice and goes to the anthology

The entries already tagged this way stand as tagged. No retagging is needed.

Alternatives

• A separate training-dynamics word. Offered to the owner, who chose widening instead. The works sit on both sides at once: implicit bias explains generalization through the dynamics, and grokking is a generalization result reached through a trajectory. Two words would split one literature.
• Leave the blurb. Rejected: the guard catching the same gap five times in a day is a bug report about the vocabulary, not about the readers.

Consequences

Readers filing training-dynamics work now have a word that says what they mean. The boundary with the anthology is unchanged: theory of how learning proceeds is held here; advice on training is the anthology’s (training-optimization).