Learned Modular Arithmetic — Optimized Schedule
Performance ablation of the uniform-transition SAIR Modular Arithmetic Challenge submission.
This artifact uses the same reproducibly trained 68,406-parameter checkpoint and learned carry, comparison, borrow, and digit-resolver components as the primary model, while specializing the fixed two-phase call schedule to avoid redundant conditional-subtraction passes. Preprocessing performs independent representation conversion only and does not reduce the raw operands.
All 15,387 possible learned primitive cells match their specifications. Local evaluation with the official pipeline: 1,000/1,000 scored problems, 100% through Tier 10, deterministic, 70.4 seconds inference on Apple MPS.
The included train.py reproduces an exact checkpoint from random
initialization using complete local primitive domains and no end-to-end modular
answers.
Source and primary research artifact: github.com/alerad/modarith-model